add rest of the lora loader mixins to the docs.

* first add a script for DC-AE;

* DC-AE init

* replace triton with custom implementation

* 1. rename file and remove un-used codes;

* no longer rely on omegaconf and dataclass

* replace custom activation with diffuers activation

* remove dc_ae attention in attention_processor.py

* iinherit from ModelMixin

* inherit from ConfigMixin

* dc-ae reduce to one file

* update downsample and upsample

* clean code

* support DecoderOutput

* remove get_same_padding and val2tuple

* remove autocast and some assert

* update ResBlock

* remove contents within super().__init__

* Update src/diffusers/models/autoencoders/dc_ae.py

Co-authored-by: YiYi Xu <yixu310@gmail.com>

* remove opsequential

* update other blocks to support the removal of build_norm

* remove build encoder/decoder project in/out

* remove inheritance of RMSNorm2d from LayerNorm

* remove reset_parameters for RMSNorm2d

Co-authored-by: YiYi Xu <yixu310@gmail.com>

* remove device and dtype in RMSNorm2d __init__

Co-authored-by: YiYi Xu <yixu310@gmail.com>

* Update src/diffusers/models/autoencoders/dc_ae.py

Co-authored-by: YiYi Xu <yixu310@gmail.com>

* Update src/diffusers/models/autoencoders/dc_ae.py

Co-authored-by: YiYi Xu <yixu310@gmail.com>

* Update src/diffusers/models/autoencoders/dc_ae.py

Co-authored-by: YiYi Xu <yixu310@gmail.com>

* remove op_list & build_block

* remove build_stage_main

* change file name to autoencoder_dc

* move LiteMLA to attention.py

* align with other vae decode output;

* add DC-AE into init files;

* update

* make quality && make style;

* quick push before dgx disappears again

* update

* make style

* update

* update

* fix

* refactor

* refactor

* refactor

* update

* possibly change to nn.Linear

* refactor

* make fix-copies

* replace vae with ae

* replace get_block_from_block_type to get_block

* replace downsample_block_type from Conv to conv for consistency

* add scaling factors

* incorporate changes for all checkpoints

* make style

* move mla to attention processor file; split qkv conv to linears

* refactor

* add tests

* from original file loader

* add docs

* add standard autoencoder methods

* combine attention processor

* fix tests

* update

* minor fix

* minor fix

* minor fix & in/out shortcut rename

* minor fix

* make style

* fix paper link

* update docs

* update single file loading

* make style

* remove single file loading support; todo for DN6

* Apply suggestions from code review

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* add abstract

* 1. add DCAE into diffusers;
2. make style and make quality;

* add DCAE_HF into diffusers;

* bug fixed;

* add SanaPipeline, SanaTransformer2D into diffusers;

* add sanaLinearAttnProcessor2_0;

* first update for SanaTransformer;

* first update for SanaPipeline;

* first success run SanaPipeline;

* model output finally match with original model with the same intput;

* code update;

* code update;

* add a flow dpm-solver scripts

* 🎉[important update]
1. Integrate flow-dpm-sovler into diffusers;
2. finally run successfully on both `FlowMatchEulerDiscreteScheduler` and `FlowDPMSolverMultistepScheduler`;

* 🎉🔧[important update & fix huge bugs!!]
1. add SanaPAGPipeline & several related Sana linear attention operators;
2. `SanaTransformer2DModel` not supports multi-resolution input;
2. fix the multi-scale HW bugs in SanaPipeline and SanaPAGPipeline;
3. fix the flow-dpm-solver set_timestep() init `model_output` and `lower_order_nums` bugs;

* remove prints;

* add convert sana official checkpoint to diffusers format Safetensor.

* Update src/diffusers/models/transformers/sana_transformer_2d.py

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* Update src/diffusers/models/transformers/sana_transformer_2d.py

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* Update src/diffusers/models/transformers/sana_transformer_2d.py

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* Update src/diffusers/pipelines/pag/pipeline_pag_sana.py

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* Update src/diffusers/models/transformers/sana_transformer_2d.py

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* Update src/diffusers/models/transformers/sana_transformer_2d.py

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* Update src/diffusers/pipelines/sana/pipeline_sana.py

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* Update src/diffusers/pipelines/sana/pipeline_sana.py

Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* update Sana for DC-AE's recent commit;

* make style && make quality

* Add StableDiffusion3PAGImg2Img Pipeline + Fix SD3 Unconditional PAG (#9932)

* fix progress bar updates in SD 1.5 PAG Img2Img pipeline

---------

Co-authored-by: Vinh H. Pham <phamvinh257@gmail.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* make the vae can be None in `__init__` of `SanaPipeline`

* Update src/diffusers/models/transformers/sana_transformer_2d.py

Co-authored-by: hlky <hlky@hlky.ac>

* change the ae related code due to the latest update of DCAE branch;

* change the ae related code due to the latest update of DCAE branch;

* 1. change code based on AutoencoderDC;
2. fix the bug of new GLUMBConv;
3. run success;

* update for solving conversation.

* 1. fix bugs and run convert script success;
2. Downloading ckpt from hub automatically;

* make style && make quality;

* 1. remove un-unsed parameters in init;
2. code update;

* remove test file

* refactor; add docs; add tests; update conversion script

* make style

* make fix-copies

* refactor

* udpate pipelines

* pag tests and refactor

* remove sana pag conversion script

* handle weight casting in conversion script

* update conversion script

* add a processor

* 1. add bf16 pth file path;
2. add complex human instruct in pipeline;

* fix fast \tests

* change gemma-2-2b-it ckpt to a non-gated repo;

* fix the pth path bug in conversion script;

* change grad ckpt to original; make style

* fix the complex_human_instruct bug and typo;

* remove dpmsolver flow scheduler

* apply review suggestions

* change the `FlowMatchEulerDiscreteScheduler` to default `DPMSolverMultistepScheduler` with flow matching scheduler.

* fix the tokenizer.padding_side='right' bug;

* update docs

* make fix-copies

* fix imports

* fix docs

* add integration test

* update docs

* update examples

* fix convert_model_output in schedulers

* fix failing tests

---------

Co-authored-by: Junyu Chen <chenjydl2003@gmail.com>
Co-authored-by: YiYi Xu <yixu310@gmail.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: chenjy2003 <70215701+chenjy2003@users.noreply.github.com>
Co-authored-by: Aryan <aryan@huggingface.co>
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>
Co-authored-by: hlky <hlky@hlky.ac>

.github/workflows/nightly_tests.yml

@@ -238,12 +238,13 @@ jobs:
 
   run_flax_tpu_tests:
     name: Nightly Flax TPU Tests
-    runs-on: docker-tpu
+    runs-on:
+      group: gcp-ct5lp-hightpu-8t
     if: github.event_name == 'schedule'
 
     container:
       image: diffusers/diffusers-flax-tpu
-      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/ --privileged
+      options: --shm-size "16gb" --ipc host --privileged ${{ vars.V5_LITEPOD_8_ENV}} -v /mnt/hf_cache:/mnt/hf_cache
     defaults:
       run:
         shell: bash
@@ -519,4 +520,4 @@ jobs:
 #        if: always()
 #        run: |
 #          pip install slack_sdk tabulate
-#          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
+#          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY

.github/workflows/pr_test_peft_backend.yml

@@ -1,134 +0,0 @@
-name: Fast tests for PRs - PEFT backend
-
-on:
-  pull_request:
-    branches:
-      - main
-    paths:
-      - "src/diffusers/**.py"
-      - "tests/**.py"
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  DIFFUSERS_IS_CI: yes
-  OMP_NUM_THREADS: 4
-  MKL_NUM_THREADS: 4
-  PYTEST_TIMEOUT: 60
-
-jobs:
-  check_code_quality:
-    runs-on: ubuntu-22.04
-    steps:
-      - uses: actions/checkout@v3
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.8"
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install .[quality]
-      - name: Check quality
-        run: make quality
-      - name: Check if failure
-        if: ${{ failure() }}
-        run: |
-          echo "Quality check failed. Please ensure the right dependency versions are installed with 'pip install -e .[quality]' and run 'make style && make quality'" >> $GITHUB_STEP_SUMMARY
-
-  check_repository_consistency:
-    needs: check_code_quality
-    runs-on: ubuntu-22.04
-    steps:
-      - uses: actions/checkout@v3
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.8"
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install .[quality]
-      - name: Check repo consistency
-        run: |
-          python utils/check_copies.py
-          python utils/check_dummies.py
-          make deps_table_check_updated
-      - name: Check if failure
-        if: ${{ failure() }}
-        run: |
-          echo "Repo consistency check failed. Please ensure the right dependency versions are installed with 'pip install -e .[quality]' and run 'make fix-copies'" >> $GITHUB_STEP_SUMMARY
-
-  run_fast_tests:
-    needs: [check_code_quality, check_repository_consistency]
-    strategy:
-      fail-fast: false
-      matrix:
-        lib-versions: ["main", "latest"]
-
-
-    name: LoRA - ${{ matrix.lib-versions }}
-
-    runs-on:
-      group: aws-general-8-plus
-
-    container:
-      image: diffusers/diffusers-pytorch-cpu
-      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/
-
-    defaults:
-      run:
-        shell: bash
-
-    steps:
-    - name: Checkout diffusers
-      uses: actions/checkout@v3
-      with:
-        fetch-depth: 2
-
-    - name: Install dependencies
-      run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m uv pip install -e [quality,test]
-        # TODO (sayakpaul, DN6): revisit `--no-deps`
-        if [ "${{ matrix.lib-versions }}" == "main" ]; then
-            python -m pip install -U peft@git+https://github.com/huggingface/peft.git --no-deps
-            python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
-            pip uninstall accelerate -y && python -m uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
-        else
-            python -m uv pip install -U peft --no-deps
-            python -m uv pip install -U transformers accelerate --no-deps
-        fi
-
-    - name: Environment
-      run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python utils/print_env.py
-
-    - name: Run fast PyTorch LoRA CPU tests with PEFT backend
-      run: |
-        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
-        python -m pytest -n 4 --max-worker-restart=0 --dist=loadfile \
-          -s -v \
-          --make-reports=tests_${{ matrix.lib-versions }} \
-          tests/lora/
-        python -m pytest -n 4 --max-worker-restart=0 --dist=loadfile \
-          -s -v \
-          --make-reports=tests_models_lora_${{ matrix.lib-versions }} \
-          tests/models/ -k "lora"
-
-
-    - name: Failure short reports
-      if: ${{ failure() }}
-      run: |
-        cat reports/tests_${{ matrix.lib-versions }}_failures_short.txt
-        cat reports/tests_models_lora_${{ matrix.lib-versions }}_failures_short.txt
-
-    - name: Test suite reports artifacts
-      if: ${{ always() }}
-      uses: actions/upload-artifact@v4
-      with:
-        name: pr_${{ matrix.lib-versions }}_test_reports
-        path: reports

.github/workflows/pr_tests.yml

@@ -234,3 +234,67 @@ jobs:
       with:
         name: pr_${{ matrix.config.report }}_test_reports
         path: reports
+
+  run_lora_tests:
+    needs: [check_code_quality, check_repository_consistency]
+    strategy:
+      fail-fast: false
+
+    name: LoRA tests with PEFT main
+
+    runs-on:
+      group: aws-general-8-plus
+
+    container:
+      image: diffusers/diffusers-pytorch-cpu
+      options: --shm-size "16gb" --ipc host -v /mnt/hf_cache:/mnt/cache/
+
+    defaults:
+      run:
+        shell: bash
+
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test]
+        # TODO (sayakpaul, DN6): revisit `--no-deps`
+        python -m pip install -U peft@git+https://github.com/huggingface/peft.git --no-deps
+        python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
+        pip uninstall accelerate -y && python -m uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
+
+    - name: Environment
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python utils/print_env.py
+
+    - name: Run fast PyTorch LoRA tests with PEFT
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m pytest -n 4 --max-worker-restart=0 --dist=loadfile \
+          -s -v \
+          --make-reports=tests_peft_main \
+          tests/lora/
+        python -m pytest -n 4 --max-worker-restart=0 --dist=loadfile \
+          -s -v \
+          --make-reports=tests_models_lora_peft_main \
+          tests/models/ -k "lora"
+
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/tests_lora_failures_short.txt
+        cat reports/tests_models_lora_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v4
+      with:
+        name: pr_main_test_reports
+        path: reports
+

.github/workflows/push_tests.yml

@@ -161,11 +161,11 @@ jobs:
 
   flax_tpu_tests:
     name: Flax TPU Tests
-    runs-on: docker-tpu
+    runs-on:
+      group: gcp-ct5lp-hightpu-8t
     container:
       image: diffusers/diffusers-flax-tpu
-      options: --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/ --privileged
-    defaults:
+      options: --shm-size "16gb" --ipc host --privileged ${{ vars.V5_LITEPOD_8_ENV}} -v /mnt/hf_cache:/mnt/hf_cache    defaults:
       run:
         shell: bash
     steps:

docs/source/en/_toctree.yml

@@ -252,6 +252,8 @@
         title: SD3ControlNetModel
       - local: api/models/controlnet_sparsectrl
         title: SparseControlNetModel
+      - local: api/models/controlnet_union
+        title: ControlNetUnionModel
       title: ControlNets
     - sections:
       - local: api/models/allegro_transformer3d
@@ -272,6 +274,8 @@
         title: LatteTransformer3DModel
       - local: api/models/lumina_nextdit2d
         title: LuminaNextDiT2DModel
+      - local: api/models/ltx_video_transformer3d
+        title: LTXVideoTransformer3DModel
       - local: api/models/mochi_transformer3d
         title: MochiTransformer3DModel
       - local: api/models/pixart_transformer2d
@@ -280,6 +284,8 @@
         title: PriorTransformer
       - local: api/models/sd3_transformer2d
         title: SD3Transformer2DModel
+      - local: api/models/sana_transformer2d
+        title: SanaTransformer2DModel
       - local: api/models/stable_audio_transformer
         title: StableAudioDiTModel
       - local: api/models/transformer2d
@@ -310,6 +316,8 @@
         title: AutoencoderKLAllegro
       - local: api/models/autoencoderkl_cogvideox
         title: AutoencoderKLCogVideoX
+      - local: api/models/autoencoderkl_ltx_video
+        title: AutoencoderKLLTXVideo
       - local: api/models/autoencoderkl_mochi
         title: AutoencoderKLMochi
       - local: api/models/asymmetricautoencoderkl
@@ -368,6 +376,8 @@
       title: ControlNet-XS
     - local: api/pipelines/controlnetxs_sdxl
       title: ControlNet-XS with Stable Diffusion XL
+    - local: api/pipelines/controlnet_union
+      title: ControlNetUnion
     - local: api/pipelines/dance_diffusion
       title: Dance Diffusion
     - local: api/pipelines/ddim
@@ -404,6 +414,8 @@
       title: Latte
     - local: api/pipelines/ledits_pp
       title: LEDITS++
+    - local: api/pipelines/ltx_video
+      title: LTX
     - local: api/pipelines/lumina
       title: Lumina-T2X
     - local: api/pipelines/marigold
@@ -424,6 +436,8 @@
       title: PixArt-α
     - local: api/pipelines/pixart_sigma
       title: PixArt-Σ
+    - local: api/pipelines/sana
+      title: Sana
     - local: api/pipelines/self_attention_guidance
       title: Self-Attention Guidance
     - local: api/pipelines/semantic_stable_diffusion

docs/source/en/api/loaders/lora.md

@@ -17,6 +17,9 @@ LoRA is a fast and lightweight training method that inserts and trains a signifi
 - [`StableDiffusionLoraLoaderMixin`] provides functions for loading and unloading, fusing and unfusing, enabling and disabling, and more functions for managing LoRA weights. This class can be used with any model.
 - [`StableDiffusionXLLoraLoaderMixin`] is a [Stable Diffusion (SDXL)](../../api/pipelines/stable_diffusion/stable_diffusion_xl) version of the [`StableDiffusionLoraLoaderMixin`] class for loading and saving LoRA weights. It can only be used with the SDXL model.
 - [`SD3LoraLoaderMixin`] provides similar functions for [Stable Diffusion 3](https://huggingface.co/blog/sd3).
+- [`FluxLoraLoaderMixin`] provides similar functions for [Flux](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux).
+- [`CogVideoXLoraLoaderMixin`] provides similar functions for [CogVideoX](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogvideox).
+- [`Mochi1LoraLoaderMixin`] provides similar functions for [Mochi](https://huggingface.co/docs/diffusers/main/en/api/pipelines/mochi).
 - [`AmusedLoraLoaderMixin`] is for the [`AmusedPipeline`].
 - [`LoraBaseMixin`] provides a base class with several utility methods to fuse, unfuse, unload, LoRAs and more.
 
@@ -38,6 +41,18 @@ To learn more about how to load LoRA weights, see the [LoRA](../../using-diffuse
 
 [[autodoc]] loaders.lora_pipeline.SD3LoraLoaderMixin
 
+## FluxLoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.FluxLoraLoaderMixin
+
+## CogVideoXLoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.CogVideoXLoraLoaderMixin
+
+## Mochi1LoraLoaderMixin
+
+[[autodoc]] loaders.lora_pipeline.Mochi1LoraLoaderMixin
+
 ## AmusedLoraLoaderMixin
 
 [[autodoc]] loaders.lora_pipeline.AmusedLoraLoaderMixin

docs/source/en/api/models/autoencoder_dc.md

@@ -37,6 +37,26 @@ from diffusers import AutoencoderDC
 ae = AutoencoderDC.from_pretrained("mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers", torch_dtype=torch.float32).to("cuda")
 ```
 
+## Load a model in Diffusers via `from_single_file`
+
+```python
+from difusers import AutoencoderDC
+
+ckpt_path = "https://huggingface.co/mit-han-lab/dc-ae-f32c32-sana-1.0/blob/main/model.safetensors"
+model = AutoencoderDC.from_single_file(ckpt_path) 
+
+```
+
+The `AutoencoderDC` model has `in` and `mix` single file checkpoint variants that have matching checkpoint keys, but use different scaling factors. It is not possible for Diffusers to automatically infer the correct config file to use with the model based on just the checkpoint and will default to configuring the model using the `mix` variant config file. To override the automatically determined config, please use the `config` argument when using single file loading with `in` variant checkpoints. 
+
+```python
+from diffusers import AutoencoderDC
+
+ckpt_path = "https://huggingface.co/mit-han-lab/dc-ae-f128c512-in-1.0/blob/main/model.safetensors"
+model = AutoencoderDC.from_single_file(ckpt_path, config="mit-han-lab/dc-ae-f128c512-in-1.0-diffusers")
+```
+
+
 ## AutoencoderDC
 
 [[autodoc]] AutoencoderDC

docs/source/en/api/models/autoencoderkl_ltx_video.md

@@ -0,0 +1,37 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# AutoencoderKLLTXVideo
+
+The 3D variational autoencoder (VAE) model with KL loss used in [LTX](https://huggingface.co/Lightricks/LTX-Video) was introduced by Lightricks.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import AutoencoderKLLTXVideo
+
+vae = AutoencoderKLLTXVideo.from_pretrained("TODO/TODO", subfolder="vae", torch_dtype=torch.float32).to("cuda")
+```
+
+## AutoencoderKLLTXVideo
+
+[[autodoc]] AutoencoderKLLTXVideo
+    - decode
+    - encode
+    - all
+
+## AutoencoderKLOutput
+
+[[autodoc]] models.autoencoders.autoencoder_kl.AutoencoderKLOutput
+
+## DecoderOutput
+
+[[autodoc]] models.autoencoders.vae.DecoderOutput

docs/source/en/api/models/controlnet_union.md

@@ -0,0 +1,35 @@
+<!--Copyright 2024 The HuggingFace Team and The InstantX Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# ControlNetUnionModel
+
+ControlNetUnionModel is an implementation of ControlNet for Stable Diffusion XL.
+
+The ControlNet model was introduced in [ControlNetPlus](https://github.com/xinsir6/ControlNetPlus) by xinsir6. It supports multiple conditioning inputs without increasing computation.
+
+*We design a new architecture that can support 10+ control types in condition text-to-image generation and can generate high resolution images visually comparable with midjourney. The network is based on the original ControlNet architecture, we propose two new modules to: 1 Extend the original ControlNet to support different image conditions using the same network parameter. 2 Support multiple conditions input without increasing computation offload, which is especially important for designers who want to edit image in detail, different conditions use the same condition encoder, without adding extra computations or parameters.*
+
+## Loading
+
+By default the [`ControlNetUnionModel`] should be loaded with [`~ModelMixin.from_pretrained`].
+
+```py
+from diffusers import StableDiffusionXLControlNetUnionPipeline, ControlNetUnionModel
+
+controlnet = ControlNetUnionModel.from_pretrained("xinsir/controlnet-union-sdxl-1.0")
+pipe = StableDiffusionXLControlNetUnionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet)
+```
+
+## ControlNetUnionModel
+
+[[autodoc]] ControlNetUnionModel
+

docs/source/en/api/models/ltx_video_transformer3d.md

@@ -0,0 +1,30 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# LTXVideoTransformer3DModel
+
+A Diffusion Transformer model for 3D data from [LTX](https://huggingface.co/Lightricks/LTX-Video) was introduced by Lightricks.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import LTXVideoTransformer3DModel
+
+transformer = LTXVideoTransformer3DModel.from_pretrained("TODO/TODO", subfolder="transformer", torch_dtype=torch.bfloat16).to("cuda")
+```
+
+## LTXVideoTransformer3DModel
+
+[[autodoc]] LTXVideoTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

docs/source/en/api/models/sana_transformer2d.md

@@ -0,0 +1,34 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# SanaTransformer2DModel
+
+A Diffusion Transformer model for 2D data from [SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers](https://huggingface.co/papers/2410.10629) was introduced from NVIDIA and MIT HAN Lab, by Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, Song Han.
+
+The abstract from the paper is:
+
+*We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096×4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: (1) Deep compression autoencoder: unlike traditional AEs, which compress images only 8×, we trained an AE that can compress images 32×, effectively reducing the number of latent tokens. (2) Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. (3) Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. (4) Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence. As a result, Sana-0.6B is very competitive with modern giant diffusion model (e.g. Flux-12B), being 20 times smaller and 100+ times faster in measured throughput. Moreover, Sana-0.6B can be deployed on a 16GB laptop GPU, taking less than 1 second to generate a 1024×1024 resolution image. Sana enables content creation at low cost. Code and model will be publicly released.*
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import SanaTransformer2DModel
+
+transformer = SanaTransformer2DModel.from_pretrained("Efficient-Large-Model/Sana_1600M_1024px_diffusers", subfolder="transformer", torch_dtype=torch.float16)
+```
+
+## SanaTransformer2DModel
+
+[[autodoc]] SanaTransformer2DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

docs/source/en/api/pipelines/controlnet_union.md

@@ -0,0 +1,35 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# ControlNetUnion
+
+ControlNetUnionModel is an implementation of ControlNet for Stable Diffusion XL.
+
+The ControlNet model was introduced in [ControlNetPlus](https://github.com/xinsir6/ControlNetPlus) by xinsir6. It supports multiple conditioning inputs without increasing computation.
+
+*We design a new architecture that can support 10+ control types in condition text-to-image generation and can generate high resolution images visually comparable with midjourney. The network is based on the original ControlNet architecture, we propose two new modules to: 1 Extend the original ControlNet to support different image conditions using the same network parameter. 2 Support multiple conditions input without increasing computation offload, which is especially important for designers who want to edit image in detail, different conditions use the same condition encoder, without adding extra computations or parameters.*
+
+
+## StableDiffusionXLControlNetUnionPipeline
+[[autodoc]] StableDiffusionXLControlNetUnionPipeline
+	- all
+	- __call__
+
+## StableDiffusionXLControlNetUnionImg2ImgPipeline
+[[autodoc]] StableDiffusionXLControlNetUnionImg2ImgPipeline
+	- all
+	- __call__
+
+## StableDiffusionXLControlNetUnionInpaintPipeline
+[[autodoc]] StableDiffusionXLControlNetUnionInpaintPipeline
+	- all
+	- __call__

docs/source/en/api/pipelines/ltx_video.md

@@ -0,0 +1,68 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License. -->
+
+# LTX
+
+[LTX Video](https://huggingface.co/Lightricks/LTX-Video) is the first DiT-based video generation model capable of generating high-quality videos in real-time. It produces 24 FPS videos at a 768x512 resolution faster than they can be watched. Trained on a large-scale dataset of diverse videos, the model generates high-resolution videos with realistic and varied content. We provide a model for both text-to-video as well as image + text-to-video usecases.
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.md) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading.md#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+## Loading Single Files
+
+Loading the original LTX Video checkpoints is also possible with [`~ModelMixin.from_single_file`].
+
+```python
+import torch
+from diffusers import AutoencoderKLLTXVideo, LTXImageToVideoPipeline, LTXVideoTransformer3DModel
+
+single_file_url = "https://huggingface.co/Lightricks/LTX-Video/ltx-video-2b-v0.9.safetensors"
+transformer = LTXVideoTransformer3DModel.from_single_file(single_file_url, torch_dtype=torch.bfloat16)
+vae = AutoencoderKLLTXVideo.from_single_file(single_file_url, torch_dtype=torch.bfloat16)
+pipe = LTXImageToVideoPipeline.from_pretrained("Lightricks/LTX-Video", transformer=transformer, vae=vae, torch_dtype=torch.bfloat16)
+
+# ... inference code ...
+```
+
+Alternatively, the pipeline can be used to load the weights with [~FromSingleFileMixin.from_single_file`].
+
+```python
+import torch
+from diffusers import LTXImageToVideoPipeline
+from transformers import T5EncoderModel, T5Tokenizer
+
+single_file_url = "https://huggingface.co/Lightricks/LTX-Video/ltx-video-2b-v0.9.safetensors"
+text_encoder = T5EncoderModel.from_pretrained("Lightricks/LTX-Video", subfolder="text_encoder", torch_dtype=torch.bfloat16)
+tokenizer = T5Tokenizer.from_pretrained("Lightricks/LTX-Video", subfolder="tokenizer", torch_dtype=torch.bfloat16)
+pipe = LTXImageToVideoPipeline.from_single_file(single_file_url, text_encoder=text_encoder, tokenizer=tokenizer, torch_dtype=torch.bfloat16)
+```
+
+## LTXPipeline
+
+[[autodoc]] LTXPipeline
+  - all
+  - __call__
+
+## LTXImageToVideoPipeline
+
+[[autodoc]] LTXImageToVideoPipeline
+  - all
+  - __call__
+
+## LTXPipelineOutput
+
+[[autodoc]] pipelines.ltx.pipeline_output.LTXPipelineOutput

docs/source/en/api/pipelines/sana.md

@@ -0,0 +1,65 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License. -->
+
+# SanaPipeline
+
+[SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers](https://huggingface.co/papers/2410.10629) from NVIDIA and MIT HAN Lab, by Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, Song Han.
+
+The abstract from the paper is:
+
+*We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096×4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: (1) Deep compression autoencoder: unlike traditional AEs, which compress images only 8×, we trained an AE that can compress images 32×, effectively reducing the number of latent tokens. (2) Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. (3) Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. (4) Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence. As a result, Sana-0.6B is very competitive with modern giant diffusion model (e.g. Flux-12B), being 20 times smaller and 100+ times faster in measured throughput. Moreover, Sana-0.6B can be deployed on a 16GB laptop GPU, taking less than 1 second to generate a 1024×1024 resolution image. Sana enables content creation at low cost. Code and model will be publicly released.*
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.md) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading.md#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+This pipeline was contributed by [lawrence-cj](https://github.com/lawrence-cj). The original codebase can be found [here](https://github.com/NVlabs/Sana). The original weights can be found under [hf.co/Efficient-Large-Model]https://huggingface.co/Efficient-Large-Model).
+
+Available models:
+
+| Model | Recommended dtype |
+|:-----:|:-----------------:|
+| [`Efficient-Large-Model/Sana_1600M_1024px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px_diffusers) | `torch.float16` |
+| [`Efficient-Large-Model/Sana_1600M_1024px_MultiLing_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px_MultiLing_diffusers) | `torch.float16` |
+| [`Efficient-Large-Model/Sana_1600M_1024px_BF16_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px_BF16_diffusers) | `torch.bfloat16` |
+| [`Efficient-Large-Model/Sana_1600M_512px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_512px_diffusers) | `torch.float16` |
+| [`Efficient-Large-Model/Sana_1600M_512px_MultiLing_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_512px_MultiLing_diffusers) | `torch.float16` |
+| [`Efficient-Large-Model/Sana_600M_1024px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_600M_1024px_diffusers) | `torch.float16` |
+| [`Efficient-Large-Model/Sana_600M_512px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_600M_512px_diffusers) | `torch.float16` |
+
+Refer to [this](https://huggingface.co/collections/Efficient-Large-Model/sana-673efba2a57ed99843f11f9e) collection for more information.
+
+<Tip>
+
+Make sure to pass the `variant` argument for downloaded checkpoints to use lower disk space. Set it to `"fp16"` for models with recommended dtype as `torch.float16`, and `"bf16"` for models with recommended dtype as `torch.bfloat16`. By default, `torch.float32` weights are downloaded, which use twice the amount of disk storage. Additionally, `torch.float32` weights can be downcasted on-the-fly by specifying the `torch_dtype` argument. Read about it in the [docs](https://huggingface.co/docs/diffusers/v0.31.0/en/api/pipelines/overview#diffusers.DiffusionPipeline.from_pretrained).
+
+</Tip>
+
+## SanaPipeline
+
+[[autodoc]] SanaPipeline
+  - all
+  - __call__
+
+## SanaPAGPipeline
+
+[[autodoc]] SanaPAGPipeline
+  - all
+  - __call__
+
+## SanaPipelineOutput
+
+[[autodoc]] pipelines.sana.pipeline_output.SanaPipelineOutput

examples/community/README_community_scripts.md

@@ -241,27 +241,15 @@ from diffusers import StableDiffusionPipeline
 from diffusers.callbacks import PipelineCallback, MultiPipelineCallbacks
 from diffusers.configuration_utils import register_to_config
 import torch
-from typing import Any, Dict, Optional
+from typing import Any, Dict, Tuple, Union
 
 
-pipeline: StableDiffusionPipeline = StableDiffusionPipeline.from_pretrained(
-    "stable-diffusion-v1-5/stable-diffusion-v1-5",
-    torch_dtype=torch.float16,
-    variant="fp16",
-    use_safetensors=True,
-).to("cuda")
-pipeline.safety_checker = None
-pipeline.requires_safety_checker = False
-
-
-class SDPromptScheduleCallback(PipelineCallback):
+class SDPromptSchedulingCallback(PipelineCallback):
     @register_to_config
     def __init__(
         self,
-        prompt: str,
-        negative_prompt: Optional[str] = None,
-        num_images_per_prompt: int = 1,
-        cutoff_step_ratio=1.0,
+        encoded_prompt: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]],
+        cutoff_step_ratio=None,
         cutoff_step_index=None,
     ):
         super().__init__(
@@ -275,6 +263,10 @@ class SDPromptScheduleCallback(PipelineCallback):
     ) -> Dict[str, Any]:
         cutoff_step_ratio = self.config.cutoff_step_ratio
         cutoff_step_index = self.config.cutoff_step_index
+        if isinstance(self.config.encoded_prompt, tuple):
+            prompt_embeds, negative_prompt_embeds = self.config.encoded_prompt
+        else:
+            prompt_embeds = self.config.encoded_prompt
 
         # Use cutoff_step_index if it's not None, otherwise use cutoff_step_ratio
         cutoff_step = (
@@ -284,34 +276,164 @@ class SDPromptScheduleCallback(PipelineCallback):
         )
 
         if step_index == cutoff_step:
-            prompt_embeds, negative_prompt_embeds = pipeline.encode_prompt(
-                prompt=self.config.prompt,
-                negative_prompt=self.config.negative_prompt,
-                device=pipeline._execution_device,
-                num_images_per_prompt=self.config.num_images_per_prompt,
-                do_classifier_free_guidance=pipeline.do_classifier_free_guidance,
-            )
             if pipeline.do_classifier_free_guidance:
                 prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
             callback_kwargs[self.tensor_inputs[0]] = prompt_embeds
         return callback_kwargs
 
+
+pipeline: StableDiffusionPipeline = StableDiffusionPipeline.from_pretrained(
+    "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    torch_dtype=torch.float16,
+    variant="fp16",
+    use_safetensors=True,
+).to("cuda")
+pipeline.safety_checker = None
+pipeline.requires_safety_checker = False
+
 callback = MultiPipelineCallbacks(
     [
-        SDPromptScheduleCallback(
-            prompt="Official portrait of a smiling world war ii general, female, cheerful, happy, detailed face, 20th century, highly detailed, cinematic lighting, digital art painting by Greg Rutkowski",
-            negative_prompt="Deformed, ugly, bad anatomy",
-            cutoff_step_ratio=0.25,
-        )
+        SDPromptSchedulingCallback(
+            encoded_prompt=pipeline.encode_prompt(
+                prompt=f"prompt {index}",
+                negative_prompt=f"negative prompt {index}",
+                device=pipeline._execution_device,
+                num_images_per_prompt=1,
+                # pipeline.do_classifier_free_guidance can't be accessed until after pipeline is ran
+                do_classifier_free_guidance=True,
+            ),
+            cutoff_step_index=index,
+        ) for index in range(1, 20)
     ]
 )
 
 image = pipeline(
-    prompt="Official portrait of a smiling world war ii general, male, cheerful, happy, detailed face, 20th century, highly detailed, cinematic lighting, digital art painting by Greg Rutkowski",
-    negative_prompt="Deformed, ugly, bad anatomy",
+    prompt="prompt"
+    negative_prompt="negative prompt",
     callback_on_step_end=callback,
     callback_on_step_end_tensor_inputs=["prompt_embeds"],
 ).images[0]
 torch.cuda.empty_cache()
 image.save('image.png')
 ```
+
+```python
+from diffusers import StableDiffusionXLPipeline
+from diffusers.callbacks import PipelineCallback, MultiPipelineCallbacks
+from diffusers.configuration_utils import register_to_config
+import torch
+from typing import Any, Dict, Tuple, Union
+
+
+class SDXLPromptSchedulingCallback(PipelineCallback):
+    @register_to_config
+    def __init__(
+        self,
+        encoded_prompt: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]],
+        add_text_embeds: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]],
+        add_time_ids: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]],
+        cutoff_step_ratio=None,
+        cutoff_step_index=None,
+    ):
+        super().__init__(
+            cutoff_step_ratio=cutoff_step_ratio, cutoff_step_index=cutoff_step_index
+        )
+
+    tensor_inputs = ["prompt_embeds", "add_text_embeds", "add_time_ids"]
+
+    def callback_fn(
+        self, pipeline, step_index, timestep, callback_kwargs
+    ) -> Dict[str, Any]:
+        cutoff_step_ratio = self.config.cutoff_step_ratio
+        cutoff_step_index = self.config.cutoff_step_index
+        if isinstance(self.config.encoded_prompt, tuple):
+            prompt_embeds, negative_prompt_embeds = self.config.encoded_prompt
+        else:
+            prompt_embeds = self.config.encoded_prompt
+        if isinstance(self.config.add_text_embeds, tuple):
+            add_text_embeds, negative_add_text_embeds = self.config.add_text_embeds
+        else:
+            add_text_embeds = self.config.add_text_embeds
+        if isinstance(self.config.add_time_ids, tuple):
+            add_time_ids, negative_add_time_ids = self.config.add_time_ids
+        else:
+            add_time_ids = self.config.add_time_ids
+
+        # Use cutoff_step_index if it's not None, otherwise use cutoff_step_ratio
+        cutoff_step = (
+            cutoff_step_index
+            if cutoff_step_index is not None
+            else int(pipeline.num_timesteps * cutoff_step_ratio)
+        )
+
+        if step_index == cutoff_step:
+            if pipeline.do_classifier_free_guidance:
+                prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+                add_text_embeds = torch.cat([negative_add_text_embeds, add_text_embeds])
+                add_time_ids = torch.cat([negative_add_time_ids, add_time_ids])
+            callback_kwargs[self.tensor_inputs[0]] = prompt_embeds
+            callback_kwargs[self.tensor_inputs[1]] = add_text_embeds
+            callback_kwargs[self.tensor_inputs[2]] = add_time_ids
+        return callback_kwargs
+
+
+pipeline: StableDiffusionXLPipeline = StableDiffusionXLPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    torch_dtype=torch.float16,
+    variant="fp16",
+    use_safetensors=True,
+).to("cuda")
+
+callbacks = []
+for index in range(1, 20):
+    (
+        prompt_embeds,
+        negative_prompt_embeds,
+        pooled_prompt_embeds,
+        negative_pooled_prompt_embeds,
+    ) = pipeline.encode_prompt(
+        prompt=f"prompt {index}",
+        negative_prompt=f"prompt {index}",
+        device=pipeline._execution_device,
+        num_images_per_prompt=1,
+        # pipeline.do_classifier_free_guidance can't be accessed until after pipeline is ran
+        do_classifier_free_guidance=True,
+    )
+    text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+    add_time_ids = pipeline._get_add_time_ids(
+        (1024, 1024),
+        (0, 0),
+        (1024, 1024),
+        dtype=prompt_embeds.dtype,
+        text_encoder_projection_dim=text_encoder_projection_dim,
+    )
+    negative_add_time_ids = pipeline._get_add_time_ids(
+        (1024, 1024),
+        (0, 0),
+        (1024, 1024),
+        dtype=prompt_embeds.dtype,
+        text_encoder_projection_dim=text_encoder_projection_dim,
+    )
+    callbacks.append(
+        SDXLPromptSchedulingCallback(
+            encoded_prompt=(prompt_embeds, negative_prompt_embeds),
+            add_text_embeds=(pooled_prompt_embeds, negative_pooled_prompt_embeds),
+            add_time_ids=(add_time_ids, negative_add_time_ids),
+            cutoff_step_index=index,
+        )
+    )
+
+
+callback = MultiPipelineCallbacks(callbacks)
+
+image = pipeline(
+    prompt="prompt",
+    negative_prompt="negative prompt",
+    callback_on_step_end=callback,
+    callback_on_step_end_tensor_inputs=[
+        "prompt_embeds",
+        "add_text_embeds",
+        "add_time_ids",
+    ],
+).images[0]
+```

examples/community/pipeline_flux_rf_inversion.py

@@ -648,6 +648,8 @@ class RFInversionFluxPipeline(
         height: Optional[int] = None,
         width: Optional[int] = None,
         eta: float = 1.0,
+        decay_eta: Optional[bool] = False,
+        eta_decay_power: Optional[float] = 1.0,
         strength: float = 1.0,
         start_timestep: float = 0,
         stop_timestep: float = 0.25,
@@ -880,12 +882,9 @@ class RFInversionFluxPipeline(
                     v_t = -noise_pred
                     v_t_cond = (y_0 - latents) / (1 - t_i)
                     eta_t = eta if start_timestep <= i < stop_timestep else 0.0
-                    if start_timestep <= i < stop_timestep:
-                        # controlled vector field
-                        v_hat_t = v_t + eta * (v_t_cond - v_t)
-
-                    else:
-                        v_hat_t = v_t
+                    if decay_eta:
+                        eta_t = eta_t * (1 - i / num_inference_steps) ** eta_decay_power  # Decay eta over the loop
+                    v_hat_t = v_t + eta_t * (v_t_cond - v_t)
 
                     # SDE Eq: 17 from https://arxiv.org/pdf/2410.10792
                     latents = latents + v_hat_t * (sigmas[i] - sigmas[i + 1])

examples/dreambooth/train_dreambooth.py

@@ -1300,16 +1300,17 @@ def main(args):
                     # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                     # This is discussed in Section 4.2 of the same paper.
                     snr = compute_snr(noise_scheduler, timesteps)
-                    base_weight = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
 
                     if noise_scheduler.config.prediction_type == "v_prediction":
                         # Velocity objective needs to be floored to an SNR weight of one.
-                        mse_loss_weights = base_weight + 1
+                        divisor = snr + 1
                     else:
-                        # Epsilon and sample both use the same loss weights.
-                        mse_loss_weights = base_weight
+                        divisor = snr
+
+                    mse_loss_weights = (
+                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / divisor
+                    )
+
                     loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                     loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                     loss = loss.mean()

examples/flux-control/README.md

@@ -0,0 +1,204 @@
+# Training Flux Control
+
+This (experimental) example shows how to train Control LoRAs with [Flux](https://huggingface.co/black-forest-labs/FLUX.1-dev) by conditioning it with additional structural controls (like depth maps, poses, etc.). We provide a script for full fine-tuning, too, refer to [this section](#full-fine-tuning). To know more about Flux Control family, refer to the following resources:
+
+* [Docs](https://github.com/black-forest-labs/flux/blob/main/docs/structural-conditioning.md) by Black Forest Labs
+* Diffusers docs ([1](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux#canny-control), [2](https://huggingface.co/docs/diffusers/main/en/api/pipelines/flux#depth-control))
+
+To incorporate additional condition latents, we expand the input features of Flux.1-Dev from 64 to 128. The first 64 channels correspond to the original input latents to be denoised, while the latter 64 channels correspond to control latents. This expansion happens on the `x_embedder` layer, where the combined latents are projected to the expected feature dimension of rest of the network. Inference is performed using the `FluxControlPipeline`.
+
+> [!NOTE]
+> **Gated model**
+>
+> As the model is gated, before using it with diffusers you first need to go to the [FLUX.1 [dev] Hugging Face page](https://huggingface.co/black-forest-labs/FLUX.1-dev), fill in the form and accept the gate. Once you are in, you need to log in so that your system knows you’ve accepted the gate. Use the command below to log in:
+
+```bash
+huggingface-cli login
+```
+
+The example command below shows how to launch fine-tuning for pose conditions. The dataset ([`raulc0399/open_pose_controlnet`](https://huggingface.co/datasets/raulc0399/open_pose_controlnet)) being used here already has the pose conditions of the original images, so we don't have to compute them.
+
+```bash
+accelerate launch train_control_lora_flux.py \
+  --pretrained_model_name_or_path="black-forest-labs/FLUX.1-dev" \
+  --dataset_name="raulc0399/open_pose_controlnet" \
+  --output_dir="pose-control-lora" \
+  --mixed_precision="bf16" \
+  --train_batch_size=1 \
+  --rank=64 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --learning_rate=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=5000 \
+  --validation_image="openpose.png" \
+  --validation_prompt="A couple, 4k photo, highly detailed" \
+  --offload \
+  --seed="0" \
+  --push_to_hub
+```
+
+`openpose.png` comes from [here](https://huggingface.co/Adapter/t2iadapter/resolve/main/openpose.png).
+
+You need to install `diffusers` from the branch of [this PR](https://github.com/huggingface/diffusers/pull/9999). When it's merged, you should install `diffusers` from the `main`.
+
+The training script exposes additional CLI args that might be useful to experiment with:
+
+* `use_lora_bias`: When set, additionally trains the biases of the `lora_B` layer. 
+* `train_norm_layers`: When set, additionally trains the normalization scales. Takes care of saving and loading.
+* `lora_layers`: Specify the layers you want to apply LoRA to. If you specify "all-linear", all the linear layers will be LoRA-attached.
+
+### Training with DeepSpeed
+
+It's possible to train with [DeepSpeed](https://github.com/microsoft/DeepSpeed), specifically leveraging the Zero2 system optimization. To use it, save the following config to an YAML file (feel free to modify as needed):
+
+```yaml
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  gradient_accumulation_steps: 1
+  gradient_clipping: 1.0
+  offload_optimizer_device: cpu
+  offload_param_device: cpu
+  zero3_init_flag: false
+  zero_stage: 2
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+enable_cpu_affinity: false
+machine_rank: 0
+main_training_function: main
+mixed_precision: bf16
+num_machines: 1
+num_processes: 1
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false
+```
+
+And then while launching training, pass the config file:
+
+```bash
+accelerate launch --config_file=CONFIG_FILE.yaml ...
+```
+
+### Inference
+
+The pose images in our dataset were computed using the [`controlnet_aux`](https://github.com/huggingface/controlnet_aux) library. Let's install it first:
+
+```bash
+pip install controlnet_aux
+```
+
+And then we are ready:
+
+```py
+from controlnet_aux import OpenposeDetector
+from diffusers import FluxControlPipeline
+from diffusers.utils import load_image
+from PIL import Image
+import numpy as np
+import torch 
+
+pipe = FluxControlPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to("cuda")
+pipe.load_lora_weights("...") # change this.
+
+open_pose = OpenposeDetector.from_pretrained("lllyasviel/Annotators")
+
+# prepare pose condition.
+url = "https://huggingface.co/Adapter/t2iadapter/resolve/main/people.jpg"
+image = load_image(url)
+image = open_pose(image, detect_resolution=512, image_resolution=1024)
+image = np.array(image)[:, :, ::-1]           
+image = Image.fromarray(np.uint8(image))
+
+prompt = "A couple, 4k photo, highly detailed"
+
+gen_images = pipe(
+  prompt=prompt,
+  condition_image=image,
+  num_inference_steps=50,
+  joint_attention_kwargs={"scale": 0.9},
+  guidance_scale=25., 
+).images[0]
+gen_images.save("output.png")
+```
+
+## Full fine-tuning
+
+We provide a non-LoRA version of the training script `train_control_flux.py`. Here is an example command:
+
+```bash
+accelerate launch --config_file=accelerate_ds2.yaml train_control_flux.py \
+  --pretrained_model_name_or_path="black-forest-labs/FLUX.1-dev" \
+  --dataset_name="raulc0399/open_pose_controlnet" \
+  --output_dir="pose-control" \
+  --mixed_precision="bf16" \
+  --train_batch_size=2 \
+  --dataloader_num_workers=4 \
+  --gradient_accumulation_steps=4 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --proportion_empty_prompts=0.2 \
+  --learning_rate=5e-5 \
+  --adam_weight_decay=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="cosine" \
+  --lr_warmup_steps=1000 \
+  --checkpointing_steps=1000 \
+  --max_train_steps=10000 \
+  --validation_steps=200 \
+  --validation_image "2_pose_1024.jpg" "3_pose_1024.jpg" \
+  --validation_prompt "two friends sitting by each other enjoying a day at the park, full hd, cinematic" "person enjoying a day at the park, full hd, cinematic" \
+  --offload \
+  --seed="0" \
+  --push_to_hub
+```
+
+Change the `validation_image` and `validation_prompt` as needed.
+
+For inference, this time, we will run:
+
+```py
+from controlnet_aux import OpenposeDetector
+from diffusers import FluxControlPipeline, FluxTransformer2DModel
+from diffusers.utils import load_image
+from PIL import Image
+import numpy as np
+import torch 
+
+transformer = FluxTransformer2DModel.from_pretrained("...") # change this.
+pipe = FluxControlPipeline.from_pretrained(
+  "black-forest-labs/FLUX.1-dev",  transformer=transformer, torch_dtype=torch.bfloat16
+).to("cuda")
+
+open_pose = OpenposeDetector.from_pretrained("lllyasviel/Annotators")
+
+# prepare pose condition.
+url = "https://huggingface.co/Adapter/t2iadapter/resolve/main/people.jpg"
+image = load_image(url)
+image = open_pose(image, detect_resolution=512, image_resolution=1024)
+image = np.array(image)[:, :, ::-1]           
+image = Image.fromarray(np.uint8(image))
+
+prompt = "A couple, 4k photo, highly detailed"
+
+gen_images = pipe(
+  prompt=prompt,
+  condition_image=image,
+  num_inference_steps=50,
+  guidance_scale=25., 
+).images[0]
+gen_images.save("output.png")
+```
+
+## Things to note
+
+* The scripts provided in this directory are experimental and educational. This means we may have to tweak things around to get good results on a given condition. We believe this is best done with the community 🤗
+* The scripts are not memory-optimized but we offload the VAE and the text encoders to CPU when they are not used. 
+* We can extract LoRAs from the fully fine-tuned model. While we currently don't provide any utilities for that, users are welcome to refer to [this script](https://github.com/Stability-AI/stability-ComfyUI-nodes/blob/master/control_lora_create.py) that provides a similar functionality. 

examples/flux-control/requirements.txt

@@ -0,0 +1,6 @@
+transformers==4.47.0
+wandb
+torch
+torchvision
+accelerate==1.2.0
+peft>=0.14.0

scripts/convert_ltx_to_diffusers.py

@@ -0,0 +1,209 @@
+import argparse
+from typing import Any, Dict
+
+import torch
+from safetensors.torch import load_file
+from transformers import T5EncoderModel, T5Tokenizer
+
+from diffusers import AutoencoderKLLTXVideo, FlowMatchEulerDiscreteScheduler, LTXPipeline, LTXVideoTransformer3DModel
+
+
+def remove_keys_(key: str, state_dict: Dict[str, Any]):
+    state_dict.pop(key)
+
+
+TOKENIZER_MAX_LENGTH = 128
+
+TRANSFORMER_KEYS_RENAME_DICT = {
+    "patchify_proj": "proj_in",
+    "adaln_single": "time_embed",
+    "q_norm": "norm_q",
+    "k_norm": "norm_k",
+}
+
+TRANSFORMER_SPECIAL_KEYS_REMAP = {}
+
+VAE_KEYS_RENAME_DICT = {
+    # decoder
+    "up_blocks.0": "mid_block",
+    "up_blocks.1": "up_blocks.0",
+    "up_blocks.2": "up_blocks.1.upsamplers.0",
+    "up_blocks.3": "up_blocks.1",
+    "up_blocks.4": "up_blocks.2.conv_in",
+    "up_blocks.5": "up_blocks.2.upsamplers.0",
+    "up_blocks.6": "up_blocks.2",
+    "up_blocks.7": "up_blocks.3.conv_in",
+    "up_blocks.8": "up_blocks.3.upsamplers.0",
+    "up_blocks.9": "up_blocks.3",
+    # encoder
+    "down_blocks.0": "down_blocks.0",
+    "down_blocks.1": "down_blocks.0.downsamplers.0",
+    "down_blocks.2": "down_blocks.0.conv_out",
+    "down_blocks.3": "down_blocks.1",
+    "down_blocks.4": "down_blocks.1.downsamplers.0",
+    "down_blocks.5": "down_blocks.1.conv_out",
+    "down_blocks.6": "down_blocks.2",
+    "down_blocks.7": "down_blocks.2.downsamplers.0",
+    "down_blocks.8": "down_blocks.3",
+    "down_blocks.9": "mid_block",
+    # common
+    "conv_shortcut": "conv_shortcut.conv",
+    "res_blocks": "resnets",
+    "norm3.norm": "norm3",
+    "per_channel_statistics.mean-of-means": "latents_mean",
+    "per_channel_statistics.std-of-means": "latents_std",
+}
+
+VAE_SPECIAL_KEYS_REMAP = {
+    "per_channel_statistics.channel": remove_keys_,
+    "per_channel_statistics.mean-of-means": remove_keys_,
+    "per_channel_statistics.mean-of-stds": remove_keys_,
+}
+
+
+def get_state_dict(saved_dict: Dict[str, Any]) -> Dict[str, Any]:
+    state_dict = saved_dict
+    if "model" in saved_dict.keys():
+        state_dict = state_dict["model"]
+    if "module" in saved_dict.keys():
+        state_dict = state_dict["module"]
+    if "state_dict" in saved_dict.keys():
+        state_dict = state_dict["state_dict"]
+    return state_dict
+
+
+def update_state_dict_inplace(state_dict: Dict[str, Any], old_key: str, new_key: str) -> Dict[str, Any]:
+    state_dict[new_key] = state_dict.pop(old_key)
+
+
+def convert_transformer(
+    ckpt_path: str,
+    dtype: torch.dtype,
+):
+    PREFIX_KEY = ""
+
+    original_state_dict = get_state_dict(load_file(ckpt_path))
+    transformer = LTXVideoTransformer3DModel().to(dtype=dtype)
+
+    for key in list(original_state_dict.keys()):
+        new_key = key[len(PREFIX_KEY) :]
+        for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict_inplace(original_state_dict, key, new_key)
+
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, original_state_dict)
+
+    transformer.load_state_dict(original_state_dict, strict=True)
+    return transformer
+
+
+def convert_vae(ckpt_path: str, dtype: torch.dtype):
+    original_state_dict = get_state_dict(load_file(ckpt_path))
+    vae = AutoencoderKLLTXVideo().to(dtype=dtype)
+
+    for key in list(original_state_dict.keys()):
+        new_key = key[:]
+        for replace_key, rename_key in VAE_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        update_state_dict_inplace(original_state_dict, key, new_key)
+
+    for key in list(original_state_dict.keys()):
+        for special_key, handler_fn_inplace in VAE_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, original_state_dict)
+
+    vae.load_state_dict(original_state_dict, strict=True)
+    return vae
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--transformer_ckpt_path", type=str, default=None, help="Path to original transformer checkpoint"
+    )
+    parser.add_argument("--vae_ckpt_path", type=str, default=None, help="Path to original vae checkpoint")
+    parser.add_argument(
+        "--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory"
+    )
+    parser.add_argument(
+        "--typecast_text_encoder",
+        action="store_true",
+        default=False,
+        help="Whether or not to apply fp16/bf16 precision to text_encoder",
+    )
+    parser.add_argument("--save_pipeline", action="store_true")
+    parser.add_argument("--output_path", type=str, required=True, help="Path where converted model should be saved")
+    parser.add_argument("--dtype", default="fp32", help="Torch dtype to save the model in.")
+    return parser.parse_args()
+
+
+DTYPE_MAPPING = {
+    "fp32": torch.float32,
+    "fp16": torch.float16,
+    "bf16": torch.bfloat16,
+}
+
+VARIANT_MAPPING = {
+    "fp32": None,
+    "fp16": "fp16",
+    "bf16": "bf16",
+}
+
+
+if __name__ == "__main__":
+    args = get_args()
+
+    transformer = None
+    dtype = DTYPE_MAPPING[args.dtype]
+    variant = VARIANT_MAPPING[args.dtype]
+
+    if args.save_pipeline:
+        assert args.transformer_ckpt_path is not None and args.vae_ckpt_path is not None
+
+    if args.transformer_ckpt_path is not None:
+        transformer: LTXVideoTransformer3DModel = convert_transformer(args.transformer_ckpt_path, dtype)
+        if not args.save_pipeline:
+            transformer.save_pretrained(
+                args.output_path, safe_serialization=True, max_shard_size="5GB", variant=variant
+            )
+
+    if args.vae_ckpt_path is not None:
+        vae: AutoencoderKLLTXVideo = convert_vae(args.vae_ckpt_path, dtype)
+        if not args.save_pipeline:
+            vae.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB", variant=variant)
+
+    if args.save_pipeline:
+        text_encoder_id = "google/t5-v1_1-xxl"
+        tokenizer = T5Tokenizer.from_pretrained(text_encoder_id, model_max_length=TOKENIZER_MAX_LENGTH)
+        text_encoder = T5EncoderModel.from_pretrained(text_encoder_id, cache_dir=args.text_encoder_cache_dir)
+
+        if args.typecast_text_encoder:
+            text_encoder = text_encoder.to(dtype=dtype)
+
+        # Apparently, the conversion does not work anymore without this :shrug:
+        for param in text_encoder.parameters():
+            param.data = param.data.contiguous()
+
+        scheduler = FlowMatchEulerDiscreteScheduler(
+            use_dynamic_shifting=True,
+            base_shift=0.95,
+            max_shift=2.05,
+            base_image_seq_len=1024,
+            max_image_seq_len=4096,
+            shift_terminal=0.1,
+        )
+
+        pipe = LTXPipeline(
+            scheduler=scheduler,
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+        )
+
+        pipe.save_pretrained(args.output_path, safe_serialization=True, variant=variant, max_shard_size="5GB")

scripts/convert_sana_to_diffusers.py

@@ -0,0 +1,307 @@
+#!/usr/bin/env python
+from __future__ import annotations
+
+import argparse
+import os
+from contextlib import nullcontext
+
+import torch
+from accelerate import init_empty_weights
+from huggingface_hub import hf_hub_download, snapshot_download
+from termcolor import colored
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from diffusers import (
+    AutoencoderDC,
+    DPMSolverMultistepScheduler,
+    FlowMatchEulerDiscreteScheduler,
+    SanaPipeline,
+    SanaTransformer2DModel,
+)
+from diffusers.models.modeling_utils import load_model_dict_into_meta
+from diffusers.utils.import_utils import is_accelerate_available
+
+
+CTX = init_empty_weights if is_accelerate_available else nullcontext
+
+ckpt_ids = [
+    "Efficient-Large-Model/Sana_1600M_1024px_MultiLing/checkpoints/Sana_1600M_1024px_MultiLing.pth",
+    "Efficient-Large-Model/Sana_1600M_1024px_BF16/checkpoints/Sana_1600M_1024px_BF16.pth",
+    "Efficient-Large-Model/Sana_1600M_512px_MultiLing/checkpoints/Sana_1600M_512px_MultiLing.pth",
+    "Efficient-Large-Model/Sana_1600M_1024px/checkpoints/Sana_1600M_1024px.pth",
+    "Efficient-Large-Model/Sana_1600M_512px/checkpoints/Sana_1600M_512px.pth",
+    "Efficient-Large-Model/Sana_600M_1024px/checkpoints/Sana_600M_1024px_MultiLing.pth",
+    "Efficient-Large-Model/Sana_600M_512px/checkpoints/Sana_600M_512px_MultiLing.pth",
+]
+# https://github.com/NVlabs/Sana/blob/main/scripts/inference.py
+
+
+def main(args):
+    cache_dir_path = os.path.expanduser("~/.cache/huggingface/hub")
+
+    if args.orig_ckpt_path is None or args.orig_ckpt_path in ckpt_ids:
+        ckpt_id = args.orig_ckpt_path or ckpt_ids[0]
+        snapshot_download(
+            repo_id=f"{'/'.join(ckpt_id.split('/')[:2])}",
+            cache_dir=cache_dir_path,
+            repo_type="model",
+        )
+        file_path = hf_hub_download(
+            repo_id=f"{'/'.join(ckpt_id.split('/')[:2])}",
+            filename=f"{'/'.join(ckpt_id.split('/')[2:])}",
+            cache_dir=cache_dir_path,
+            repo_type="model",
+        )
+    else:
+        file_path = args.orig_ckpt_path
+
+    print(colored(f"Loading checkpoint from {file_path}", "green", attrs=["bold"]))
+    all_state_dict = torch.load(file_path, weights_only=True)
+    state_dict = all_state_dict.pop("state_dict")
+    converted_state_dict = {}
+
+    # Patch embeddings.
+    converted_state_dict["patch_embed.proj.weight"] = state_dict.pop("x_embedder.proj.weight")
+    converted_state_dict["patch_embed.proj.bias"] = state_dict.pop("x_embedder.proj.bias")
+
+    # Caption projection.
+    converted_state_dict["caption_projection.linear_1.weight"] = state_dict.pop("y_embedder.y_proj.fc1.weight")
+    converted_state_dict["caption_projection.linear_1.bias"] = state_dict.pop("y_embedder.y_proj.fc1.bias")
+    converted_state_dict["caption_projection.linear_2.weight"] = state_dict.pop("y_embedder.y_proj.fc2.weight")
+    converted_state_dict["caption_projection.linear_2.bias"] = state_dict.pop("y_embedder.y_proj.fc2.bias")
+
+    # AdaLN-single LN
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.weight"] = state_dict.pop(
+        "t_embedder.mlp.0.weight"
+    )
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_1.bias"] = state_dict.pop("t_embedder.mlp.0.bias")
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.weight"] = state_dict.pop(
+        "t_embedder.mlp.2.weight"
+    )
+    converted_state_dict["time_embed.emb.timestep_embedder.linear_2.bias"] = state_dict.pop("t_embedder.mlp.2.bias")
+
+    # Shared norm.
+    converted_state_dict["time_embed.linear.weight"] = state_dict.pop("t_block.1.weight")
+    converted_state_dict["time_embed.linear.bias"] = state_dict.pop("t_block.1.bias")
+
+    # y norm
+    converted_state_dict["caption_norm.weight"] = state_dict.pop("attention_y_norm.weight")
+
+    flow_shift = 3.0
+    if args.model_type == "SanaMS_1600M_P1_D20":
+        layer_num = 20
+    elif args.model_type == "SanaMS_600M_P1_D28":
+        layer_num = 28
+    else:
+        raise ValueError(f"{args.model_type} is not supported.")
+
+    for depth in range(layer_num):
+        # Transformer blocks.
+        converted_state_dict[f"transformer_blocks.{depth}.scale_shift_table"] = state_dict.pop(
+            f"blocks.{depth}.scale_shift_table"
+        )
+
+        # Linear Attention is all you need 🤘
+        # Self attention.
+        q, k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.attn.qkv.weight"), 3, dim=0)
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.weight"] = v
+        # Projection.
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.weight"] = state_dict.pop(
+            f"blocks.{depth}.attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.bias"] = state_dict.pop(
+            f"blocks.{depth}.attn.proj.bias"
+        )
+
+        # Feed-forward.
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_inverted.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.inverted_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_inverted.bias"] = state_dict.pop(
+            f"blocks.{depth}.mlp.inverted_conv.conv.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_depth.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.depth_conv.conv.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_depth.bias"] = state_dict.pop(
+            f"blocks.{depth}.mlp.depth_conv.conv.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.conv_point.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.point_conv.conv.weight"
+        )
+
+        # Cross-attention.
+        q = state_dict.pop(f"blocks.{depth}.cross_attn.q_linear.weight")
+        q_bias = state_dict.pop(f"blocks.{depth}.cross_attn.q_linear.bias")
+        k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.cross_attn.kv_linear.weight"), 2, dim=0)
+        k_bias, v_bias = torch.chunk(state_dict.pop(f"blocks.{depth}.cross_attn.kv_linear.bias"), 2, dim=0)
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.bias"] = q_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.bias"] = k_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.weight"] = v
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.bias"] = v_bias
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.weight"] = state_dict.pop(
+            f"blocks.{depth}.cross_attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.bias"] = state_dict.pop(
+            f"blocks.{depth}.cross_attn.proj.bias"
+        )
+
+    # Final block.
+    converted_state_dict["proj_out.weight"] = state_dict.pop("final_layer.linear.weight")
+    converted_state_dict["proj_out.bias"] = state_dict.pop("final_layer.linear.bias")
+    converted_state_dict["scale_shift_table"] = state_dict.pop("final_layer.scale_shift_table")
+
+    # Transformer
+    with CTX():
+        transformer = SanaTransformer2DModel(
+            in_channels=32,
+            out_channels=32,
+            num_attention_heads=model_kwargs[args.model_type]["num_attention_heads"],
+            attention_head_dim=model_kwargs[args.model_type]["attention_head_dim"],
+            num_layers=model_kwargs[args.model_type]["num_layers"],
+            num_cross_attention_heads=model_kwargs[args.model_type]["num_cross_attention_heads"],
+            cross_attention_head_dim=model_kwargs[args.model_type]["cross_attention_head_dim"],
+            cross_attention_dim=model_kwargs[args.model_type]["cross_attention_dim"],
+            caption_channels=2304,
+            mlp_ratio=2.5,
+            attention_bias=False,
+            sample_size=args.image_size // 32,
+            patch_size=1,
+            norm_elementwise_affine=False,
+            norm_eps=1e-6,
+        )
+
+    if is_accelerate_available():
+        load_model_dict_into_meta(transformer, converted_state_dict)
+    else:
+        transformer.load_state_dict(converted_state_dict, strict=True, assign=True)
+
+    try:
+        state_dict.pop("y_embedder.y_embedding")
+        state_dict.pop("pos_embed")
+    except KeyError:
+        print("y_embedder.y_embedding or pos_embed not found in the state_dict")
+
+    assert len(state_dict) == 0, f"State dict is not empty, {state_dict.keys()}"
+
+    num_model_params = sum(p.numel() for p in transformer.parameters())
+    print(f"Total number of transformer parameters: {num_model_params}")
+
+    transformer = transformer.to(weight_dtype)
+
+    if not args.save_full_pipeline:
+        print(
+            colored(
+                f"Only saving transformer model of {args.model_type}. "
+                f"Set --save_full_pipeline to save the whole SanaPipeline",
+                "green",
+                attrs=["bold"],
+            )
+        )
+        transformer.save_pretrained(
+            os.path.join(args.dump_path, "transformer"), safe_serialization=True, max_shard_size="5GB", variant=variant
+        )
+    else:
+        print(colored(f"Saving the whole SanaPipeline containing {args.model_type}", "green", attrs=["bold"]))
+        # VAE
+        ae = AutoencoderDC.from_pretrained("mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers", torch_dtype=torch.float32)
+
+        # Text Encoder
+        text_encoder_model_path = "google/gemma-2-2b-it"
+        tokenizer = AutoTokenizer.from_pretrained(text_encoder_model_path)
+        tokenizer.padding_side = "right"
+        text_encoder = AutoModelForCausalLM.from_pretrained(
+            text_encoder_model_path, torch_dtype=torch.bfloat16
+        ).get_decoder()
+
+        # Scheduler
+        if args.scheduler_type == "flow-dpm_solver":
+            scheduler = DPMSolverMultistepScheduler(
+                flow_shift=flow_shift,
+                use_flow_sigmas=True,
+                prediction_type="flow_prediction",
+            )
+        elif args.scheduler_type == "flow-euler":
+            scheduler = FlowMatchEulerDiscreteScheduler(shift=flow_shift)
+        else:
+            raise ValueError(f"Scheduler type {args.scheduler_type} is not supported")
+
+        pipe = SanaPipeline(
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+            transformer=transformer,
+            vae=ae,
+            scheduler=scheduler,
+        )
+        pipe.save_pretrained(args.dump_path, safe_serialization=True, max_shard_size="5GB", variant=variant)
+
+
+DTYPE_MAPPING = {
+    "fp32": torch.float32,
+    "fp16": torch.float16,
+    "bf16": torch.bfloat16,
+}
+
+VARIANT_MAPPING = {
+    "fp32": None,
+    "fp16": "fp16",
+    "bf16": "bf16",
+}
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--orig_ckpt_path", default=None, type=str, required=False, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--image_size",
+        default=1024,
+        type=int,
+        choices=[512, 1024],
+        required=False,
+        help="Image size of pretrained model, 512 or 1024.",
+    )
+    parser.add_argument(
+        "--model_type", default="SanaMS_1600M_P1_D20", type=str, choices=["SanaMS_1600M_P1_D20", "SanaMS_600M_P1_D28"]
+    )
+    parser.add_argument(
+        "--scheduler_type", default="flow-dpm_solver", type=str, choices=["flow-dpm_solver", "flow-euler"]
+    )
+    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
+    parser.add_argument("--save_full_pipeline", action="store_true", help="save all the pipelien elemets in one.")
+    parser.add_argument("--dtype", default="fp32", type=str, choices=["fp32", "fp16", "bf16"], help="Weight dtype.")
+
+    args = parser.parse_args()
+
+    model_kwargs = {
+        "SanaMS_1600M_P1_D20": {
+            "num_attention_heads": 70,
+            "attention_head_dim": 32,
+            "num_cross_attention_heads": 20,
+            "cross_attention_head_dim": 112,
+            "cross_attention_dim": 2240,
+            "num_layers": 20,
+        },
+        "SanaMS_600M_P1_D28": {
+            "num_attention_heads": 36,
+            "attention_head_dim": 32,
+            "num_cross_attention_heads": 16,
+            "cross_attention_head_dim": 72,
+            "cross_attention_dim": 1152,
+            "num_layers": 28,
+        },
+    }
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    weight_dtype = DTYPE_MAPPING[args.dtype]
+    variant = VARIANT_MAPPING[args.dtype]
+
+    main(args)

src/diffusers/__init__.py

@@ -84,6 +84,7 @@ else:
             "AutoencoderKL",
             "AutoencoderKLAllegro",
             "AutoencoderKLCogVideoX",
+            "AutoencoderKLLTXVideo",
             "AutoencoderKLMochi",
             "AutoencoderKLTemporalDecoder",
             "AutoencoderOobleck",
@@ -92,6 +93,7 @@ else:
             "CogView3PlusTransformer2DModel",
             "ConsistencyDecoderVAE",
             "ControlNetModel",
+            "ControlNetUnionModel",
             "ControlNetXSAdapter",
             "DiTTransformer2DModel",
             "FluxControlNetModel",
@@ -103,6 +105,7 @@ else:
             "I2VGenXLUNet",
             "Kandinsky3UNet",
             "LatteTransformer3DModel",
+            "LTXVideoTransformer3DModel",
             "LuminaNextDiT2DModel",
             "MochiTransformer3DModel",
             "ModelMixin",
@@ -111,6 +114,7 @@ else:
             "MultiControlNetModel",
             "PixArtTransformer2DModel",
             "PriorTransformer",
+            "SanaTransformer2DModel",
             "SD3ControlNetModel",
             "SD3MultiControlNetModel",
             "SD3Transformer2DModel",
@@ -316,6 +320,8 @@ else:
             "LDMTextToImagePipeline",
             "LEditsPPPipelineStableDiffusion",
             "LEditsPPPipelineStableDiffusionXL",
+            "LTXImageToVideoPipeline",
+            "LTXPipeline",
             "LuminaText2ImgPipeline",
             "MarigoldDepthPipeline",
             "MarigoldNormalsPipeline",
@@ -327,6 +333,8 @@ else:
             "PixArtSigmaPAGPipeline",
             "PixArtSigmaPipeline",
             "ReduxImageEncoder",
+            "SanaPAGPipeline",
+            "SanaPipeline",
             "SemanticStableDiffusionPipeline",
             "ShapEImg2ImgPipeline",
             "ShapEPipeline",
@@ -340,6 +348,7 @@ else:
             "StableDiffusion3Img2ImgPipeline",
             "StableDiffusion3InpaintPipeline",
             "StableDiffusion3PAGImg2ImgPipeline",
+            "StableDiffusion3PAGImg2ImgPipeline",
             "StableDiffusion3PAGPipeline",
             "StableDiffusion3Pipeline",
             "StableDiffusionAdapterPipeline",
@@ -378,6 +387,9 @@ else:
             "StableDiffusionXLControlNetPAGImg2ImgPipeline",
             "StableDiffusionXLControlNetPAGPipeline",
             "StableDiffusionXLControlNetPipeline",
+            "StableDiffusionXLControlNetUnionImg2ImgPipeline",
+            "StableDiffusionXLControlNetUnionInpaintPipeline",
+            "StableDiffusionXLControlNetUnionPipeline",
             "StableDiffusionXLControlNetXSPipeline",
             "StableDiffusionXLImg2ImgPipeline",
             "StableDiffusionXLInpaintPipeline",
@@ -578,6 +590,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderKL,
             AutoencoderKLAllegro,
             AutoencoderKLCogVideoX,
+            AutoencoderKLLTXVideo,
             AutoencoderKLMochi,
             AutoencoderKLTemporalDecoder,
             AutoencoderOobleck,
@@ -586,6 +599,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             CogView3PlusTransformer2DModel,
             ConsistencyDecoderVAE,
             ControlNetModel,
+            ControlNetUnionModel,
             ControlNetXSAdapter,
             DiTTransformer2DModel,
             FluxControlNetModel,
@@ -597,6 +611,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             I2VGenXLUNet,
             Kandinsky3UNet,
             LatteTransformer3DModel,
+            LTXVideoTransformer3DModel,
             LuminaNextDiT2DModel,
             MochiTransformer3DModel,
             ModelMixin,
@@ -605,6 +620,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             MultiControlNetModel,
             PixArtTransformer2DModel,
             PriorTransformer,
+            SanaTransformer2DModel,
             SD3ControlNetModel,
             SD3MultiControlNetModel,
             SD3Transformer2DModel,
@@ -789,6 +805,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LDMTextToImagePipeline,
             LEditsPPPipelineStableDiffusion,
             LEditsPPPipelineStableDiffusionXL,
+            LTXImageToVideoPipeline,
+            LTXPipeline,
             LuminaText2ImgPipeline,
             MarigoldDepthPipeline,
             MarigoldNormalsPipeline,
@@ -800,6 +818,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             PixArtSigmaPAGPipeline,
             PixArtSigmaPipeline,
             ReduxImageEncoder,
+            SanaPAGPipeline,
+            SanaPipeline,
             SemanticStableDiffusionPipeline,
             ShapEImg2ImgPipeline,
             ShapEPipeline,
@@ -850,6 +870,9 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusionXLControlNetPAGImg2ImgPipeline,
             StableDiffusionXLControlNetPAGPipeline,
             StableDiffusionXLControlNetPipeline,
+            StableDiffusionXLControlNetUnionImg2ImgPipeline,
+            StableDiffusionXLControlNetUnionInpaintPipeline,
+            StableDiffusionXLControlNetUnionPipeline,
             StableDiffusionXLControlNetXSPipeline,
             StableDiffusionXLImg2ImgPipeline,
             StableDiffusionXLInpaintPipeline,

src/diffusers/loaders/lora_conversion_utils.py

@@ -665,251 +665,6 @@ def _convert_xlabs_flux_lora_to_diffusers(old_state_dict):
     return new_state_dict
 
 
-def _convert_non_diffusers_sd3_lora_to_diffusers(state_dict, prefix=None):
-    new_state_dict = {}
-
-    # in SD3 original implementation of AdaLayerNormContinuous, it split linear projection output into shift, scale;
-    # while in diffusers it split into scale, shift. Here we swap the linear projection weights in order to be able to use diffusers implementation
-    def swap_scale_shift(weight):
-        shift, scale = weight.chunk(2, dim=0)
-        new_weight = torch.cat([scale, shift], dim=0)
-        return new_weight
-
-    def calculate_scales(key):
-        lora_rank = state_dict[f"{key}.lora_down.weight"].shape[0]
-        alpha = state_dict.pop(key + ".alpha")
-        scale = alpha / lora_rank
-
-        # calculate scale_down and scale_up
-        scale_down = scale
-        scale_up = 1.0
-        while scale_down * 2 < scale_up:
-            scale_down *= 2
-            scale_up /= 2
-
-        return scale_down, scale_up
-
-    def weight_is_sparse(key, rank, num_splits, up_weight):
-        dims = [up_weight.shape[0] // num_splits] * num_splits
-
-        is_sparse = False
-        requested_rank = rank
-        if rank % num_splits == 0:
-            requested_rank = rank // num_splits
-            is_sparse = True
-            i = 0
-            for j in range(len(dims)):
-                for k in range(len(dims)):
-                    if j == k:
-                        continue
-                    is_sparse = is_sparse and torch.all(
-                        up_weight[i : i + dims[j], k * requested_rank : (k + 1) * requested_rank] == 0
-                    )
-                i += dims[j]
-            if is_sparse:
-                logger.info(f"weight is sparse: {key}")
-
-        return is_sparse, requested_rank
-
-    # handle only transformer blocks for now.
-    layers = set()
-    for k in state_dict:
-        if "joint_blocks" in k:
-            idx = int(k.split("_", 4)[-1].split("_", 1)[0])
-            layers.add(idx)
-    num_layers = max(layers) + 1
-
-    for i in range(num_layers):
-        # norms
-        for diffusers_key, orig_key in [
-            (f"transformer_blocks.{i}.norm1.linear", f"lora_unet_joint_blocks_{i}_x_block_adaLN_modulation_1")
-        ]:
-            scale_down, scale_up = calculate_scales(orig_key)
-            new_state_dict[f"{diffusers_key}.lora_A.weight"] = (
-                state_dict.pop(f"{orig_key}.lora_down.weight") * scale_down
-            )
-            new_state_dict[f"{diffusers_key}.lora_B.weight"] = state_dict.pop(f"{orig_key}.lora_up.weight") * scale_up
-
-        if not (i == num_layers - 1):
-            for diffusers_key, orig_key in [
-                (
-                    f"transformer_blocks.{i}.norm1_context.linear",
-                    f"lora_unet_joint_blocks_{i}_context_block_adaLN_modulation_1",
-                )
-            ]:
-                scale_down, scale_up = calculate_scales(orig_key)
-                new_state_dict[f"{diffusers_key}.lora_A.weight"] = (
-                    state_dict.pop(f"{orig_key}.lora_down.weight") * scale_down
-                )
-                new_state_dict[f"{diffusers_key}.lora_B.weight"] = (
-                    state_dict.pop(f"{orig_key}.lora_up.weight") * scale_up
-                )
-        else:
-            for diffusers_key, orig_key in [
-                (
-                    f"transformer_blocks.{i}.norm1_context.linear",
-                    f"lora_unet_joint_blocks_{i}_context_block_adaLN_modulation_1",
-                )
-            ]:
-                scale_down, scale_up = calculate_scales(orig_key)
-                new_state_dict[f"{diffusers_key}.lora_A.weight"] = (
-                    swap_scale_shift(state_dict.pop(f"{orig_key}.lora_down.weight")) * scale_down
-                )
-                new_state_dict[f"{diffusers_key}.lora_B.weight"] = (
-                    swap_scale_shift(state_dict.pop(f"{orig_key}.lora_up.weight")) * scale_up
-                )
-
-        # output projections
-        for diffusers_key, orig_key in [
-            (f"transformer_blocks.{i}.attn.to_out.0", f"lora_unet_joint_blocks_{i}_x_block_attn_proj")
-        ]:
-            scale_down, scale_up = calculate_scales(orig_key)
-            new_state_dict[f"{diffusers_key}.lora_A.weight"] = (
-                state_dict.pop(f"{orig_key}.lora_down.weight") * scale_down
-            )
-            new_state_dict[f"{diffusers_key}.lora_B.weight"] = state_dict.pop(f"{orig_key}.lora_up.weight") * scale_up
-        if not (i == num_layers - 1):
-            for diffusers_key, orig_key in [
-                (f"transformer_blocks.{i}.attn.to_add_out", f"lora_unet_joint_blocks_{i}_context_block_attn_proj")
-            ]:
-                scale_down, scale_up = calculate_scales(orig_key)
-                new_state_dict[f"{diffusers_key}.lora_A.weight"] = (
-                    state_dict.pop(f"{orig_key}.lora_down.weight") * scale_down
-                )
-                new_state_dict[f"{diffusers_key}.lora_B.weight"] = (
-                    state_dict.pop(f"{orig_key}.lora_up.weight") * scale_up
-                )
-
-        # ffs
-        for diffusers_key, orig_key in [
-            (f"transformer_blocks.{i}.ff.net.0.proj", f"lora_unet_joint_blocks_{i}_x_block_mlp_fc1")
-        ]:
-            scale_down, scale_up = calculate_scales(orig_key)
-            new_state_dict[f"{diffusers_key}.lora_A.weight"] = (
-                state_dict.pop(f"{orig_key}.lora_down.weight") * scale_down
-            )
-            new_state_dict[f"{diffusers_key}.lora_B.weight"] = state_dict.pop(f"{orig_key}.lora_up.weight") * scale_up
-
-        for diffusers_key, orig_key in [
-            (f"transformer_blocks.{i}.ff.net.2", f"lora_unet_joint_blocks_{i}_x_block_mlp_fc2")
-        ]:
-            scale_down, scale_up = calculate_scales(orig_key)
-            new_state_dict[f"{diffusers_key}.lora_A.weight"] = (
-                state_dict.pop(f"{orig_key}.lora_down.weight") * scale_down
-            )
-            new_state_dict[f"{diffusers_key}.lora_B.weight"] = state_dict.pop(f"{orig_key}.lora_up.weight") * scale_up
-
-        if not (i == num_layers - 1):
-            for diffusers_key, orig_key in [
-                (f"transformer_blocks.{i}.ff_context.net.0.proj", f"lora_unet_joint_blocks_{i}_context_block_mlp_fc1")
-            ]:
-                scale_down, scale_up = calculate_scales(orig_key)
-                new_state_dict[f"{diffusers_key}.lora_A.weight"] = (
-                    state_dict.pop(f"{orig_key}.lora_down.weight") * scale_down
-                )
-                new_state_dict[f"{diffusers_key}.lora_B.weight"] = (
-                    state_dict.pop(f"{orig_key}.lora_up.weight") * scale_up
-                )
-
-            for diffusers_key, orig_key in [
-                (f"transformer_blocks.{i}.ff_context.net.2", f"lora_unet_joint_blocks_{i}_context_block_mlp_fc2")
-            ]:
-                scale_down, scale_up = calculate_scales(orig_key)
-                new_state_dict[f"{diffusers_key}.lora_A.weight"] = (
-                    state_dict.pop(f"{orig_key}.lora_down.weight") * scale_down
-                )
-                new_state_dict[f"{diffusers_key}.lora_B.weight"] = (
-                    state_dict.pop(f"{orig_key}.lora_up.weight") * scale_up
-                )
-
-        # core transformer blocks.
-        # sample blocks.
-        scale_down, scale_up = calculate_scales(f"lora_unet_joint_blocks_{i}_x_block_attn_qkv")
-        is_sparse, requested_rank = weight_is_sparse(
-            key=f"lora_unet_joint_blocks_{i}_x_block_attn_qkv",
-            rank=state_dict[f"lora_unet_joint_blocks_{i}_x_block_attn_qkv.lora_down.weight"].shape[0],
-            num_splits=3,
-            up_weight=state_dict[f"lora_unet_joint_blocks_{i}_x_block_attn_qkv.lora_up.weight"],
-        )
-        num_splits = 3
-        sample_qkv_lora_down = (
-            state_dict.pop(f"lora_unet_joint_blocks_{i}_x_block_attn_qkv.lora_down.weight") * scale_down
-        )
-        sample_qkv_lora_up = state_dict.pop(f"lora_unet_joint_blocks_{i}_x_block_attn_qkv.lora_up.weight") * scale_up
-        dims = [sample_qkv_lora_up.shape[0] // num_splits] * num_splits  # 3 = num_splits
-        if not is_sparse:
-            for attn_k in ["to_q", "to_k", "to_v"]:
-                new_state_dict[f"transformer_blocks.{i}.attn.{attn_k}.lora_A.weight"] = sample_qkv_lora_down
-            for attn_k, v in zip(["to_q", "to_k", "to_v"], torch.split(sample_qkv_lora_up, dims, dim=0)):
-                new_state_dict[f"transformer_blocks.{i}.attn.{attn_k}.lora_B.weight"] = v
-        else:
-            # down_weight is chunked to each split
-            new_state_dict.update(
-                {
-                    f"transformer_blocks.{i}.attn.{k}.lora_A.weight": v
-                    for k, v in zip(["to_q", "to_k", "to_v"], torch.chunk(sample_qkv_lora_down, num_splits, dim=0))
-                }
-            )  # noqa: C416
-
-            # up_weight is sparse: only non-zero values are copied to each split
-            i = 0
-            for j, attn_k in enumerate(["to_q", "to_k", "to_v"]):
-                new_state_dict[f"transformer_blocks.{i}.attn.{attn_k}.lora_B.weight"] = sample_qkv_lora_up[
-                    i : i + dims[j], j * requested_rank : (j + 1) * requested_rank
-                ].contiguous()
-                i += dims[j]
-
-        # context blocks.
-        scale_down, scale_up = calculate_scales(f"lora_unet_joint_blocks_{i}_context_block_attn_qkv")
-        is_sparse, requested_rank = weight_is_sparse(
-            key=f"lora_unet_joint_blocks_{i}_context_block_attn_qkv",
-            rank=state_dict[f"lora_unet_joint_blocks_{i}_context_block_attn_qkv.lora_down.weight"].shape[0],
-            num_splits=3,
-            up_weight=state_dict[f"lora_unet_joint_blocks_{i}_context_block_attn_qkv.lora_up.weight"],
-        )
-        num_splits = 3
-        sample_qkv_lora_down = (
-            state_dict.pop(f"lora_unet_joint_blocks_{i}_context_block_attn_qkv.lora_down.weight") * scale_down
-        )
-        sample_qkv_lora_up = (
-            state_dict.pop(f"lora_unet_joint_blocks_{i}_context_block_attn_qkv.lora_up.weight") * scale_up
-        )
-        dims = [sample_qkv_lora_up.shape[0] // num_splits] * num_splits  # 3 = num_splits
-        if not is_sparse:
-            for attn_k in ["add_q_proj", "add_k_proj", "add_v_proj"]:
-                new_state_dict[f"transformer_blocks.{i}.attn.{attn_k}.lora_A.weight"] = sample_qkv_lora_down
-            for attn_k, v in zip(
-                ["add_q_proj", "add_k_proj", "add_v_proj"], torch.split(sample_qkv_lora_up, dims, dim=0)
-            ):
-                new_state_dict[f"transformer_blocks.{i}.attn.{attn_k}.lora_B.weight"] = v
-        else:
-            # down_weight is chunked to each split
-            new_state_dict.update(
-                {
-                    f"transformer_blocks.{i}.attn.{k}.lora_A.weight": v
-                    for k, v in zip(
-                        ["add_q_proj", "add_k_proj", "add_v_proj"],
-                        torch.chunk(sample_qkv_lora_down, num_splits, dim=0),
-                    )
-                }
-            )  # noqa: C416
-
-            # up_weight is sparse: only non-zero values are copied to each split
-            i = 0
-            for j, attn_k in enumerate(["add_q_proj", "add_k_proj", "add_v_proj"]):
-                new_state_dict[f"transformer_blocks.{i}.attn.{attn_k}.lora_B.weight"] = sample_qkv_lora_up[
-                    i : i + dims[j], j * requested_rank : (j + 1) * requested_rank
-                ].contiguous()
-                i += dims[j]
-
-    if len(state_dict) > 0:
-        raise ValueError(f"`state_dict` should be at this point but has: {list(state_dict.keys())}.")
-
-    prefix = prefix or "transformer"
-    new_state_dict = {f"{prefix}.{k}": v for k, v in new_state_dict.items()}
-    return new_state_dict
-
-
 def _convert_bfl_flux_control_lora_to_diffusers(original_state_dict):
     converted_state_dict = {}
     original_state_dict_keys = list(original_state_dict.keys())

src/diffusers/loaders/lora_pipeline.py

@@ -38,7 +38,6 @@ from .lora_conversion_utils import (
     _convert_bfl_flux_control_lora_to_diffusers,
     _convert_kohya_flux_lora_to_diffusers,
     _convert_non_diffusers_lora_to_diffusers,
-    _convert_non_diffusers_sd3_lora_to_diffusers,
     _convert_xlabs_flux_lora_to_diffusers,
     _maybe_map_sgm_blocks_to_diffusers,
 )
@@ -1240,27 +1239,6 @@ class SD3LoraLoaderMixin(LoraBaseMixin):
             logger.warning(warn_msg)
             state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}
 
-        is_non_diffusers = any("lora_unet" in k for k in state_dict)
-        if is_non_diffusers:
-            has_only_transformer = all(k.startswith("lora_unet") for k in state_dict)
-            if not has_only_transformer:
-                state_dict = {k: v for k, v in state_dict.items() if k.startswith("lora_unet")}
-                logger.warning(
-                    "Some keys in the LoRA checkpoint are not related to transformer blocks and we will filter them out during loading. Please open a new issue with the LoRA checkpoint you are trying to load with a reproducible snippet - https://github.com/huggingface/diffusers/issues/new."
-                )
-
-            all_joint_blocks = all("joint_blocks" in k for k in state_dict)
-            if not all_joint_blocks:
-                raise ValueError(
-                    "LoRAs containing only transformer blocks are supported at this point. Please open a new issue with the LoRA checkpoint you are trying to load with a reproducible snippet - https://github.com/huggingface/diffusers/issues/new."
-                )
-
-            has_dual_attention_layers = any("attn2" in k for k in state_dict)
-            if has_dual_attention_layers:
-                raise ValueError("LoRA state dicts with dual attention layers are not supported.")
-
-            state_dict = _convert_non_diffusers_sd3_lora_to_diffusers(state_dict, prefix=cls.transformer_name)
-
         return state_dict
 
     def load_lora_weights(
@@ -1305,11 +1283,12 @@ class SD3LoraLoaderMixin(LoraBaseMixin):
 
         # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
         state_dict = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
+
         is_correct_format = all("lora" in key for key in state_dict.keys())
         if not is_correct_format:
             raise ValueError("Invalid LoRA checkpoint.")
 
-        transformer_state_dict = {k: v for k, v in state_dict.items() if k.startswith(f"{self.transformer_name}.")}
+        transformer_state_dict = {k: v for k, v in state_dict.items() if "transformer." in k}
         if len(transformer_state_dict) > 0:
             self.load_lora_into_transformer(
                 state_dict,
@@ -1320,10 +1299,8 @@ class SD3LoraLoaderMixin(LoraBaseMixin):
                 _pipeline=self,
                 low_cpu_mem_usage=low_cpu_mem_usage,
             )
-        else:
-            logger.debug("No LoRA keys were found for the transformer.")
 
-        text_encoder_state_dict = {k: v for k, v in state_dict.items() if k.startswith(f"{self.text_encoder_name}.")}
+        text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
         if len(text_encoder_state_dict) > 0:
             self.load_lora_into_text_encoder(
                 text_encoder_state_dict,
@@ -1335,10 +1312,8 @@ class SD3LoraLoaderMixin(LoraBaseMixin):
                 _pipeline=self,
                 low_cpu_mem_usage=low_cpu_mem_usage,
             )
-        else:
-            logger.debug("No LoRA keys were found for the first text encoder.")
 
-        text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if k.startswith("text_encoder_2.")}
+        text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
         if len(text_encoder_2_state_dict) > 0:
             self.load_lora_into_text_encoder(
                 text_encoder_2_state_dict,
@@ -1350,8 +1325,6 @@ class SD3LoraLoaderMixin(LoraBaseMixin):
                 _pipeline=self,
                 low_cpu_mem_usage=low_cpu_mem_usage,
             )
-        else:
-            logger.debug("No LoRA keys were found for the second text encoder.")
 
     @classmethod
     def load_lora_into_transformer(
@@ -2340,7 +2313,7 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
         for name, module in transformer.named_modules():
             if isinstance(module, torch.nn.Linear):
                 module_weight = module.weight.data
-                module_bias = module.bias.data if hasattr(module, "bias") else None
+                module_bias = module.bias.data if module.bias is not None else None
                 bias = module_bias is not None
 
                 lora_A_weight_name = f"{name}.lora_A.weight"
@@ -2364,12 +2337,19 @@ class FluxLoraLoaderMixin(LoraBaseMixin):
                         f"this please open an issue at https://github.com/huggingface/diffusers/issues."
                     )
 
-                logger.debug(
+                debug_message = (
                     f'Expanding the nn.Linear input/output features for module="{name}" because the provided LoRA '
                     f"checkpoint contains higher number of features than expected. The number of input_features will be "
-                    f"expanded from {module_in_features} to {in_features}, and the number of output features will be "
-                    f"expanded from {module_out_features} to {out_features}."
+                    f"expanded from {module_in_features} to {in_features}"
                 )
+                if module_out_features != out_features:
+                    debug_message += (
+                        ", and the number of output features will be "
+                        f"expanded from {module_out_features} to {out_features}."
+                    )
+                else:
+                    debug_message += "."
+                logger.debug(debug_message)
 
                 has_param_with_shape_update = True
                 parent_module_name, _, current_module_name = name.rpartition(".")

src/diffusers/loaders/peft.py

@@ -205,6 +205,7 @@ class PeftAdapterMixin:
                 weights.
         """
         from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
+        from peft.tuners.tuners_utils import BaseTunerLayer
 
         cache_dir = kwargs.pop("cache_dir", None)
         force_download = kwargs.pop("force_download", False)
@@ -316,8 +317,22 @@ class PeftAdapterMixin:
             if is_peft_version(">=", "0.13.1"):
                 peft_kwargs["low_cpu_mem_usage"] = low_cpu_mem_usage
 
-            inject_adapter_in_model(lora_config, self, adapter_name=adapter_name, **peft_kwargs)
-            incompatible_keys = set_peft_model_state_dict(self, state_dict, adapter_name, **peft_kwargs)
+            # To handle scenarios where we cannot successfully set state dict. If it's unsucessful,
+            # we should also delete the `peft_config` associated to the `adapter_name`.
+            try:
+                inject_adapter_in_model(lora_config, self, adapter_name=adapter_name, **peft_kwargs)
+                incompatible_keys = set_peft_model_state_dict(self, state_dict, adapter_name, **peft_kwargs)
+            except RuntimeError as e:
+                for module in self.modules():
+                    if isinstance(module, BaseTunerLayer):
+                        active_adapters = module.active_adapters
+                        for active_adapter in active_adapters:
+                            if adapter_name in active_adapter:
+                                module.delete_adapter(adapter_name)
+
+                self.peft_config.pop(adapter_name)
+                logger.error(f"Loading {adapter_name} was unsucessful with the following error: \n{e}")
+                raise
 
             warn_msg = ""
             if incompatible_keys is not None:

src/diffusers/loaders/single_file_model.py

@@ -23,10 +23,13 @@ from ..utils import deprecate, is_accelerate_available, logging
 from .single_file_utils import (
     SingleFileComponentError,
     convert_animatediff_checkpoint_to_diffusers,
+    convert_autoencoder_dc_checkpoint_to_diffusers,
     convert_controlnet_checkpoint,
     convert_flux_transformer_checkpoint_to_diffusers,
     convert_ldm_unet_checkpoint,
     convert_ldm_vae_checkpoint,
+    convert_ltx_transformer_checkpoint_to_diffusers,
+    convert_ltx_vae_checkpoint_to_diffusers,
     convert_sd3_transformer_checkpoint_to_diffusers,
     convert_stable_cascade_unet_single_file_to_diffusers,
     create_controlnet_diffusers_config_from_ldm,
@@ -82,6 +85,15 @@ SINGLE_FILE_LOADABLE_CLASSES = {
         "checkpoint_mapping_fn": convert_flux_transformer_checkpoint_to_diffusers,
         "default_subfolder": "transformer",
     },
+    "LTXVideoTransformer3DModel": {
+        "checkpoint_mapping_fn": convert_ltx_transformer_checkpoint_to_diffusers,
+        "default_subfolder": "transformer",
+    },
+    "AutoencoderKLLTXVideo": {
+        "checkpoint_mapping_fn": convert_ltx_vae_checkpoint_to_diffusers,
+        "default_subfolder": "vae",
+    },
+    "AutoencoderDC": {"checkpoint_mapping_fn": convert_autoencoder_dc_checkpoint_to_diffusers},
 }
 
 

src/diffusers/loaders/single_file_utils.py

@@ -92,6 +92,14 @@ CHECKPOINT_KEY_NAMES = {
         "double_blocks.0.img_attn.norm.key_norm.scale",
         "model.diffusion_model.double_blocks.0.img_attn.norm.key_norm.scale",
     ],
+    "ltx-video": [
+        (
+            "model.diffusion_model.patchify_proj.weight",
+            "model.diffusion_model.transformer_blocks.27.scale_shift_table",
+        ),
+    ],
+    "autoencoder-dc": "decoder.stages.1.op_list.0.main.conv.conv.bias",
+    "autoencoder-dc-sana": "encoder.project_in.conv.bias",
 }
 
 DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
@@ -138,6 +146,11 @@ DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
     "animatediff_rgb": {"pretrained_model_name_or_path": "guoyww/animatediff-sparsectrl-rgb"},
     "flux-dev": {"pretrained_model_name_or_path": "black-forest-labs/FLUX.1-dev"},
     "flux-schnell": {"pretrained_model_name_or_path": "black-forest-labs/FLUX.1-schnell"},
+    "ltx-video": {"pretrained_model_name_or_path": "Lightricks/LTX-Video"},
+    "autoencoder-dc-f128c512": {"pretrained_model_name_or_path": "mit-han-lab/dc-ae-f128c512-mix-1.0-diffusers"},
+    "autoencoder-dc-f64c128": {"pretrained_model_name_or_path": "mit-han-lab/dc-ae-f64c128-mix-1.0-diffusers"},
+    "autoencoder-dc-f32c32": {"pretrained_model_name_or_path": "mit-han-lab/dc-ae-f32c32-mix-1.0-diffusers"},
+    "autoencoder-dc-f32c32-sana": {"pretrained_model_name_or_path": "mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers"},
 }
 
 # Use to configure model sample size when original config is provided
@@ -564,6 +577,26 @@ def infer_diffusers_model_type(checkpoint):
             model_type = "flux-dev"
         else:
             model_type = "flux-schnell"
+
+    elif any(all(key in checkpoint for key in key_list) for key_list in CHECKPOINT_KEY_NAMES["ltx-video"]):
+        model_type = "ltx-video"
+
+    elif CHECKPOINT_KEY_NAMES["autoencoder-dc"] in checkpoint:
+        encoder_key = "encoder.project_in.conv.conv.bias"
+        decoder_key = "decoder.project_in.main.conv.weight"
+
+        if CHECKPOINT_KEY_NAMES["autoencoder-dc-sana"] in checkpoint:
+            model_type = "autoencoder-dc-f32c32-sana"
+
+        elif checkpoint[encoder_key].shape[-1] == 64 and checkpoint[decoder_key].shape[1] == 32:
+            model_type = "autoencoder-dc-f32c32"
+
+        elif checkpoint[encoder_key].shape[-1] == 64 and checkpoint[decoder_key].shape[1] == 128:
+            model_type = "autoencoder-dc-f64c128"
+
+        else:
+            model_type = "autoencoder-dc-f128c512"
+
     else:
         model_type = "v1"
 
@@ -2198,3 +2231,165 @@ def convert_flux_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
     )
 
     return converted_state_dict
+
+
+def convert_ltx_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
+    converted_state_dict = {
+        key: checkpoint.pop(key) for key in list(checkpoint.keys()) if "model.diffusion_model." in key
+    }
+
+    TRANSFORMER_KEYS_RENAME_DICT = {
+        "model.diffusion_model.": "",
+        "patchify_proj": "proj_in",
+        "adaln_single": "time_embed",
+        "q_norm": "norm_q",
+        "k_norm": "norm_k",
+    }
+
+    TRANSFORMER_SPECIAL_KEYS_REMAP = {}
+
+    for key in list(converted_state_dict.keys()):
+        new_key = key
+        for replace_key, rename_key in TRANSFORMER_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        converted_state_dict[new_key] = converted_state_dict.pop(key)
+
+    for key in list(converted_state_dict.keys()):
+        for special_key, handler_fn_inplace in TRANSFORMER_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, converted_state_dict)
+
+    return converted_state_dict
+
+
+def convert_ltx_vae_checkpoint_to_diffusers(checkpoint, **kwargs):
+    converted_state_dict = {key: checkpoint.pop(key) for key in list(checkpoint.keys()) if "vae." in key}
+
+    def remove_keys_(key: str, state_dict):
+        state_dict.pop(key)
+
+    VAE_KEYS_RENAME_DICT = {
+        # common
+        "vae.": "",
+        # decoder
+        "up_blocks.0": "mid_block",
+        "up_blocks.1": "up_blocks.0",
+        "up_blocks.2": "up_blocks.1.upsamplers.0",
+        "up_blocks.3": "up_blocks.1",
+        "up_blocks.4": "up_blocks.2.conv_in",
+        "up_blocks.5": "up_blocks.2.upsamplers.0",
+        "up_blocks.6": "up_blocks.2",
+        "up_blocks.7": "up_blocks.3.conv_in",
+        "up_blocks.8": "up_blocks.3.upsamplers.0",
+        "up_blocks.9": "up_blocks.3",
+        # encoder
+        "down_blocks.0": "down_blocks.0",
+        "down_blocks.1": "down_blocks.0.downsamplers.0",
+        "down_blocks.2": "down_blocks.0.conv_out",
+        "down_blocks.3": "down_blocks.1",
+        "down_blocks.4": "down_blocks.1.downsamplers.0",
+        "down_blocks.5": "down_blocks.1.conv_out",
+        "down_blocks.6": "down_blocks.2",
+        "down_blocks.7": "down_blocks.2.downsamplers.0",
+        "down_blocks.8": "down_blocks.3",
+        "down_blocks.9": "mid_block",
+        # common
+        "conv_shortcut": "conv_shortcut.conv",
+        "res_blocks": "resnets",
+        "norm3.norm": "norm3",
+        "per_channel_statistics.mean-of-means": "latents_mean",
+        "per_channel_statistics.std-of-means": "latents_std",
+    }
+
+    VAE_SPECIAL_KEYS_REMAP = {
+        "per_channel_statistics.channel": remove_keys_,
+        "per_channel_statistics.mean-of-means": remove_keys_,
+        "per_channel_statistics.mean-of-stds": remove_keys_,
+    }
+
+    for key in list(converted_state_dict.keys()):
+        new_key = key
+        for replace_key, rename_key in VAE_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        converted_state_dict[new_key] = converted_state_dict.pop(key)
+
+    for key in list(converted_state_dict.keys()):
+        for special_key, handler_fn_inplace in VAE_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, converted_state_dict)
+
+    return converted_state_dict
+
+
+def convert_autoencoder_dc_checkpoint_to_diffusers(checkpoint, **kwargs):
+    converted_state_dict = {key: checkpoint.pop(key) for key in list(checkpoint.keys())}
+
+    def remap_qkv_(key: str, state_dict):
+        qkv = state_dict.pop(key)
+        q, k, v = torch.chunk(qkv, 3, dim=0)
+        parent_module, _, _ = key.rpartition(".qkv.conv.weight")
+        state_dict[f"{parent_module}.to_q.weight"] = q.squeeze()
+        state_dict[f"{parent_module}.to_k.weight"] = k.squeeze()
+        state_dict[f"{parent_module}.to_v.weight"] = v.squeeze()
+
+    def remap_proj_conv_(key: str, state_dict):
+        parent_module, _, _ = key.rpartition(".proj.conv.weight")
+        state_dict[f"{parent_module}.to_out.weight"] = state_dict.pop(key).squeeze()
+
+    AE_KEYS_RENAME_DICT = {
+        # common
+        "main.": "",
+        "op_list.": "",
+        "context_module": "attn",
+        "local_module": "conv_out",
+        # NOTE: The below two lines work because scales in the available configs only have a tuple length of 1
+        # If there were more scales, there would be more layers, so a loop would be better to handle this
+        "aggreg.0.0": "to_qkv_multiscale.0.proj_in",
+        "aggreg.0.1": "to_qkv_multiscale.0.proj_out",
+        "depth_conv.conv": "conv_depth",
+        "inverted_conv.conv": "conv_inverted",
+        "point_conv.conv": "conv_point",
+        "point_conv.norm": "norm",
+        "conv.conv.": "conv.",
+        "conv1.conv": "conv1",
+        "conv2.conv": "conv2",
+        "conv2.norm": "norm",
+        "proj.norm": "norm_out",
+        # encoder
+        "encoder.project_in.conv": "encoder.conv_in",
+        "encoder.project_out.0.conv": "encoder.conv_out",
+        "encoder.stages": "encoder.down_blocks",
+        # decoder
+        "decoder.project_in.conv": "decoder.conv_in",
+        "decoder.project_out.0": "decoder.norm_out",
+        "decoder.project_out.2.conv": "decoder.conv_out",
+        "decoder.stages": "decoder.up_blocks",
+    }
+
+    AE_F32C32_F64C128_F128C512_KEYS = {
+        "encoder.project_in.conv": "encoder.conv_in.conv",
+        "decoder.project_out.2.conv": "decoder.conv_out.conv",
+    }
+
+    AE_SPECIAL_KEYS_REMAP = {
+        "qkv.conv.weight": remap_qkv_,
+        "proj.conv.weight": remap_proj_conv_,
+    }
+    if "encoder.project_in.conv.bias" not in converted_state_dict:
+        AE_KEYS_RENAME_DICT.update(AE_F32C32_F64C128_F128C512_KEYS)
+
+    for key in list(converted_state_dict.keys()):
+        new_key = key[:]
+        for replace_key, rename_key in AE_KEYS_RENAME_DICT.items():
+            new_key = new_key.replace(replace_key, rename_key)
+        converted_state_dict[new_key] = converted_state_dict.pop(key)
+
+    for key in list(converted_state_dict.keys()):
+        for special_key, handler_fn_inplace in AE_SPECIAL_KEYS_REMAP.items():
+            if special_key not in key:
+                continue
+            handler_fn_inplace(key, converted_state_dict)
+
+    return converted_state_dict

src/diffusers/models/__init__.py

@@ -31,6 +31,7 @@ if is_torch_available():
     _import_structure["autoencoders.autoencoder_kl"] = ["AutoencoderKL"]
     _import_structure["autoencoders.autoencoder_kl_allegro"] = ["AutoencoderKLAllegro"]
     _import_structure["autoencoders.autoencoder_kl_cogvideox"] = ["AutoencoderKLCogVideoX"]
+    _import_structure["autoencoders.autoencoder_kl_ltx"] = ["AutoencoderKLLTXVideo"]
     _import_structure["autoencoders.autoencoder_kl_mochi"] = ["AutoencoderKLMochi"]
     _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
     _import_structure["autoencoders.autoencoder_oobleck"] = ["AutoencoderOobleck"]
@@ -45,6 +46,7 @@ if is_torch_available():
     ]
     _import_structure["controlnets.controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"]
     _import_structure["controlnets.controlnet_sparsectrl"] = ["SparseControlNetModel"]
+    _import_structure["controlnets.controlnet_union"] = ["ControlNetUnionModel"]
     _import_structure["controlnets.controlnet_xs"] = ["ControlNetXSAdapter", "UNetControlNetXSModel"]
     _import_structure["controlnets.multicontrolnet"] = ["MultiControlNetModel"]
     _import_structure["embeddings"] = ["ImageProjection"]
@@ -58,12 +60,14 @@ if is_torch_available():
     _import_structure["transformers.lumina_nextdit2d"] = ["LuminaNextDiT2DModel"]
     _import_structure["transformers.pixart_transformer_2d"] = ["PixArtTransformer2DModel"]
     _import_structure["transformers.prior_transformer"] = ["PriorTransformer"]
+    _import_structure["transformers.sana_transformer"] = ["SanaTransformer2DModel"]
     _import_structure["transformers.stable_audio_transformer"] = ["StableAudioDiTModel"]
     _import_structure["transformers.t5_film_transformer"] = ["T5FilmDecoder"]
     _import_structure["transformers.transformer_2d"] = ["Transformer2DModel"]
     _import_structure["transformers.transformer_allegro"] = ["AllegroTransformer3DModel"]
     _import_structure["transformers.transformer_cogview3plus"] = ["CogView3PlusTransformer2DModel"]
     _import_structure["transformers.transformer_flux"] = ["FluxTransformer2DModel"]
+    _import_structure["transformers.transformer_ltx"] = ["LTXVideoTransformer3DModel"]
     _import_structure["transformers.transformer_mochi"] = ["MochiTransformer3DModel"]
     _import_structure["transformers.transformer_sd3"] = ["SD3Transformer2DModel"]
     _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
@@ -93,6 +97,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderKL,
             AutoencoderKLAllegro,
             AutoencoderKLCogVideoX,
+            AutoencoderKLLTXVideo,
             AutoencoderKLMochi,
             AutoencoderKLTemporalDecoder,
             AutoencoderOobleck,
@@ -102,6 +107,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         )
         from .controlnets import (
             ControlNetModel,
+            ControlNetUnionModel,
             ControlNetXSAdapter,
             FluxControlNetModel,
             FluxMultiControlNetModel,
@@ -125,10 +131,12 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             FluxTransformer2DModel,
             HunyuanDiT2DModel,
             LatteTransformer3DModel,
+            LTXVideoTransformer3DModel,
             LuminaNextDiT2DModel,
             MochiTransformer3DModel,
             PixArtTransformer2DModel,
             PriorTransformer,
+            SanaTransformer2DModel,
             SD3Transformer2DModel,
             StableAudioDiTModel,
             T5FilmDecoder,

src/diffusers/models/activations.py

@@ -18,7 +18,7 @@ import torch.nn.functional as F
 from torch import nn
 
 from ..utils import deprecate
-from ..utils.import_utils import is_torch_npu_available
+from ..utils.import_utils import is_torch_npu_available, is_torch_version
 
 
 if is_torch_npu_available():
@@ -79,10 +79,10 @@ class GELU(nn.Module):
         self.approximate = approximate
 
     def gelu(self, gate: torch.Tensor) -> torch.Tensor:
-        if gate.device.type != "mps":
-            return F.gelu(gate, approximate=self.approximate)
-        # mps: gelu is not implemented for float16
-        return F.gelu(gate.to(dtype=torch.float32), approximate=self.approximate).to(dtype=gate.dtype)
+        if gate.device.type == "mps" and is_torch_version("<", "2.0.0"):
+            # fp16 gelu not supported on mps before torch 2.0
+            return F.gelu(gate.to(dtype=torch.float32), approximate=self.approximate).to(dtype=gate.dtype)
+        return F.gelu(gate, approximate=self.approximate)
 
     def forward(self, hidden_states):
         hidden_states = self.proj(hidden_states)
@@ -105,10 +105,10 @@ class GEGLU(nn.Module):
         self.proj = nn.Linear(dim_in, dim_out * 2, bias=bias)
 
     def gelu(self, gate: torch.Tensor) -> torch.Tensor:
-        if gate.device.type != "mps":
-            return F.gelu(gate)
-        # mps: gelu is not implemented for float16
-        return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
+        if gate.device.type == "mps" and is_torch_version("<", "2.0.0"):
+            # fp16 gelu not supported on mps before torch 2.0
+            return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
+        return F.gelu(gate)
 
     def forward(self, hidden_states, *args, **kwargs):
         if len(args) > 0 or kwargs.get("scale", None) is not None:

src/diffusers/models/attention_flax.py

@@ -216,8 +216,8 @@ class FlaxAttention(nn.Module):
             hidden_states = jax_memory_efficient_attention(
                 query_states, key_states, value_states, query_chunk_size=query_chunk_size, key_chunk_size=4096 * 4
             )
-
             hidden_states = hidden_states.transpose(1, 0, 2)
+            hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
         else:
             # compute attentions
             if self.split_head_dim:

src/diffusers/models/attention_processor.py

@@ -199,12 +199,16 @@ class Attention(nn.Module):
             self.norm_q = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
             self.norm_k = FP32LayerNorm(dim_head, elementwise_affine=False, bias=False, eps=eps)
         elif qk_norm == "layer_norm_across_heads":
-            # Lumina applys qk norm across all heads
+            # Lumina applies qk norm across all heads
             self.norm_q = nn.LayerNorm(dim_head * heads, eps=eps)
             self.norm_k = nn.LayerNorm(dim_head * kv_heads, eps=eps)
         elif qk_norm == "rms_norm":
             self.norm_q = RMSNorm(dim_head, eps=eps)
             self.norm_k = RMSNorm(dim_head, eps=eps)
+        elif qk_norm == "rms_norm_across_heads":
+            # LTX applies qk norm across all heads
+            self.norm_q = RMSNorm(dim_head * heads, eps=eps)
+            self.norm_k = RMSNorm(dim_head * kv_heads, eps=eps)
         elif qk_norm == "l2":
             self.norm_q = LpNorm(p=2, dim=-1, eps=eps)
             self.norm_k = LpNorm(p=2, dim=-1, eps=eps)
@@ -358,6 +362,14 @@ class Attention(nn.Module):
             self.processor,
             (IPAdapterAttnProcessor, IPAdapterAttnProcessor2_0, IPAdapterXFormersAttnProcessor),
         )
+        is_joint_processor = hasattr(self, "processor") and isinstance(
+            self.processor,
+            (
+                JointAttnProcessor2_0,
+                XFormersJointAttnProcessor,
+            ),
+        )
+
         if use_memory_efficient_attention_xformers:
             if is_added_kv_processor and is_custom_diffusion:
                 raise NotImplementedError(
@@ -420,6 +432,8 @@ class Attention(nn.Module):
                     processor.to(
                         device=self.processor.to_k_ip[0].weight.device, dtype=self.processor.to_k_ip[0].weight.dtype
                     )
+            elif is_joint_processor:
+                processor = XFormersJointAttnProcessor(attention_op=attention_op)
             else:
                 processor = XFormersAttnProcessor(attention_op=attention_op)
         else:
@@ -1685,6 +1699,91 @@ class FusedJointAttnProcessor2_0:
         return hidden_states, encoder_hidden_states
 
 
+class XFormersJointAttnProcessor:
+    r"""
+    Processor for implementing memory efficient attention using xFormers.
+
+    Args:
+        attention_op (`Callable`, *optional*, defaults to `None`):
+            The base
+            [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to
+            use as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best
+            operator.
+    """
+
+    def __init__(self, attention_op: Optional[Callable] = None):
+        self.attention_op = attention_op
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        residual = hidden_states
+
+        # `sample` projections.
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        query = attn.head_to_batch_dim(query).contiguous()
+        key = attn.head_to_batch_dim(key).contiguous()
+        value = attn.head_to_batch_dim(value).contiguous()
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # `context` projections.
+        if encoder_hidden_states is not None:
+            encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
+            encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+            encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+
+            encoder_hidden_states_query_proj = attn.head_to_batch_dim(encoder_hidden_states_query_proj).contiguous()
+            encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj).contiguous()
+            encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj).contiguous()
+
+            if attn.norm_added_q is not None:
+                encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
+            if attn.norm_added_k is not None:
+                encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
+
+            query = torch.cat([query, encoder_hidden_states_query_proj], dim=1)
+            key = torch.cat([key, encoder_hidden_states_key_proj], dim=1)
+            value = torch.cat([value, encoder_hidden_states_value_proj], dim=1)
+
+        hidden_states = xformers.ops.memory_efficient_attention(
+            query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
+        )
+        hidden_states = hidden_states.to(query.dtype)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        if encoder_hidden_states is not None:
+            # Split the attention outputs.
+            hidden_states, encoder_hidden_states = (
+                hidden_states[:, : residual.shape[1]],
+                hidden_states[:, residual.shape[1] :],
+            )
+            if not attn.context_pre_only:
+                encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if encoder_hidden_states is not None:
+            return hidden_states, encoder_hidden_states
+        else:
+            return hidden_states
+
+
 class AllegroAttnProcessor2_0:
     r"""
     Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). This is
@@ -5342,6 +5441,165 @@ class FluxSingleAttnProcessor2_0(FluxAttnProcessor2_0):
         super().__init__()
 
 
+class SanaLinearAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product linear attention.
+    """
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        original_dtype = hidden_states.dtype
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = query.transpose(1, 2).unflatten(1, (attn.heads, -1))
+        key = key.transpose(1, 2).unflatten(1, (attn.heads, -1)).transpose(2, 3)
+        value = value.transpose(1, 2).unflatten(1, (attn.heads, -1))
+
+        query = F.relu(query)
+        key = F.relu(key)
+
+        query, key, value = query.float(), key.float(), value.float()
+
+        value = F.pad(value, (0, 0, 0, 1), mode="constant", value=1.0)
+        scores = torch.matmul(value, key)
+        hidden_states = torch.matmul(scores, query)
+
+        hidden_states = hidden_states[:, :, :-1] / (hidden_states[:, :, -1:] + 1e-15)
+        hidden_states = hidden_states.flatten(1, 2).transpose(1, 2)
+        hidden_states = hidden_states.to(original_dtype)
+
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if original_dtype == torch.float16:
+            hidden_states = hidden_states.clip(-65504, 65504)
+
+        return hidden_states
+
+
+class PAGCFGSanaLinearAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product linear attention.
+    """
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        original_dtype = hidden_states.dtype
+
+        hidden_states_uncond, hidden_states_org, hidden_states_ptb = hidden_states.chunk(3)
+        hidden_states_org = torch.cat([hidden_states_uncond, hidden_states_org])
+
+        query = attn.to_q(hidden_states_org)
+        key = attn.to_k(hidden_states_org)
+        value = attn.to_v(hidden_states_org)
+
+        query = query.transpose(1, 2).unflatten(1, (attn.heads, -1))
+        key = key.transpose(1, 2).unflatten(1, (attn.heads, -1)).transpose(2, 3)
+        value = value.transpose(1, 2).unflatten(1, (attn.heads, -1))
+
+        query = F.relu(query)
+        key = F.relu(key)
+
+        query, key, value = query.float(), key.float(), value.float()
+
+        value = F.pad(value, (0, 0, 0, 1), mode="constant", value=1.0)
+        scores = torch.matmul(value, key)
+        hidden_states_org = torch.matmul(scores, query)
+
+        hidden_states_org = hidden_states_org[:, :, :-1] / (hidden_states_org[:, :, -1:] + 1e-15)
+        hidden_states_org = hidden_states_org.flatten(1, 2).transpose(1, 2)
+        hidden_states_org = hidden_states_org.to(original_dtype)
+
+        hidden_states_org = attn.to_out[0](hidden_states_org)
+        hidden_states_org = attn.to_out[1](hidden_states_org)
+
+        # perturbed path (identity attention)
+        hidden_states_ptb = attn.to_v(hidden_states_ptb).to(original_dtype)
+
+        hidden_states_ptb = attn.to_out[0](hidden_states_ptb)
+        hidden_states_ptb = attn.to_out[1](hidden_states_ptb)
+
+        hidden_states = torch.cat([hidden_states_org, hidden_states_ptb])
+
+        if original_dtype == torch.float16:
+            hidden_states = hidden_states.clip(-65504, 65504)
+
+        return hidden_states
+
+
+class PAGIdentitySanaLinearAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product linear attention.
+    """
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        original_dtype = hidden_states.dtype
+
+        hidden_states_org, hidden_states_ptb = hidden_states.chunk(2)
+
+        query = attn.to_q(hidden_states_org)
+        key = attn.to_k(hidden_states_org)
+        value = attn.to_v(hidden_states_org)
+
+        query = query.transpose(1, 2).unflatten(1, (attn.heads, -1))
+        key = key.transpose(1, 2).unflatten(1, (attn.heads, -1)).transpose(2, 3)
+        value = value.transpose(1, 2).unflatten(1, (attn.heads, -1))
+
+        query = F.relu(query)
+        key = F.relu(key)
+
+        query, key, value = query.float(), key.float(), value.float()
+
+        value = F.pad(value, (0, 0, 0, 1), mode="constant", value=1.0)
+        scores = torch.matmul(value, key)
+        hidden_states_org = torch.matmul(scores, query)
+
+        if hidden_states_org.dtype in [torch.float16, torch.bfloat16]:
+            hidden_states_org = hidden_states_org.float()
+
+        hidden_states_org = hidden_states_org[:, :, :-1] / (hidden_states_org[:, :, -1:] + 1e-15)
+        hidden_states_org = hidden_states_org.flatten(1, 2).transpose(1, 2)
+        hidden_states_org = hidden_states_org.to(original_dtype)
+
+        hidden_states_org = attn.to_out[0](hidden_states_org)
+        hidden_states_org = attn.to_out[1](hidden_states_org)
+
+        # perturbed path (identity attention)
+        hidden_states_ptb = attn.to_v(hidden_states_ptb).to(original_dtype)
+
+        hidden_states_ptb = attn.to_out[0](hidden_states_ptb)
+        hidden_states_ptb = attn.to_out[1](hidden_states_ptb)
+
+        hidden_states = torch.cat([hidden_states_org, hidden_states_ptb])
+
+        if original_dtype == torch.float16:
+            hidden_states = hidden_states.clip(-65504, 65504)
+
+        return hidden_states
+
+
 ADDED_KV_ATTENTION_PROCESSORS = (
     AttnAddedKVProcessor,
     SlicedAttnAddedKVProcessor,
@@ -5394,6 +5652,12 @@ AttentionProcessor = Union[
     CustomDiffusionAttnProcessor2_0,
     SlicedAttnProcessor,
     SlicedAttnAddedKVProcessor,
+    SanaLinearAttnProcessor2_0,
+    PAGCFGSanaLinearAttnProcessor2_0,
+    PAGIdentitySanaLinearAttnProcessor2_0,
+    SanaMultiscaleLinearAttention,
+    SanaMultiscaleAttnProcessor2_0,
+    SanaMultiscaleAttentionProjection,
     IPAdapterAttnProcessor,
     IPAdapterAttnProcessor2_0,
     IPAdapterXFormersAttnProcessor,

src/diffusers/models/autoencoders/__init__.py

@@ -3,6 +3,7 @@ from .autoencoder_dc import AutoencoderDC
 from .autoencoder_kl import AutoencoderKL
 from .autoencoder_kl_allegro import AutoencoderKLAllegro
 from .autoencoder_kl_cogvideox import AutoencoderKLCogVideoX
+from .autoencoder_kl_ltx import AutoencoderKLLTXVideo
 from .autoencoder_kl_mochi import AutoencoderKLMochi
 from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
 from .autoencoder_oobleck import AutoencoderOobleck

src/diffusers/models/autoencoders/autoencoder_dc.py

@@ -26,39 +26,10 @@ from ..activations import get_activation
 from ..attention_processor import SanaMultiscaleLinearAttention
 from ..modeling_utils import ModelMixin
 from ..normalization import RMSNorm, get_normalization
+from ..transformers.sana_transformer import GLUMBConv
 from .vae import DecoderOutput, EncoderOutput
 
 
-class GLUMBConv(nn.Module):
-    def __init__(self, in_channels: int, out_channels: int) -> None:
-        super().__init__()
-
-        hidden_channels = 4 * in_channels
-
-        self.nonlinearity = nn.SiLU()
-
-        self.conv_inverted = nn.Conv2d(in_channels, hidden_channels * 2, 1, 1, 0)
-        self.conv_depth = nn.Conv2d(hidden_channels * 2, hidden_channels * 2, 3, 1, 1, groups=hidden_channels * 2)
-        self.conv_point = nn.Conv2d(hidden_channels, out_channels, 1, 1, 0, bias=False)
-        self.norm = RMSNorm(out_channels, eps=1e-5, elementwise_affine=True, bias=True)
-
-    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        residual = hidden_states
-
-        hidden_states = self.conv_inverted(hidden_states)
-        hidden_states = self.nonlinearity(hidden_states)
-
-        hidden_states = self.conv_depth(hidden_states)
-        hidden_states, gate = torch.chunk(hidden_states, 2, dim=1)
-        hidden_states = hidden_states * self.nonlinearity(gate)
-
-        hidden_states = self.conv_point(hidden_states)
-        # move channel to the last dimension so we apply RMSnorm across channel dimension
-        hidden_states = self.norm(hidden_states.movedim(1, -1)).movedim(-1, 1)
-
-        return hidden_states + residual
-
-
 class ResBlock(nn.Module):
     def __init__(
         self,
@@ -115,6 +86,7 @@ class EfficientViTBlock(nn.Module):
         self.conv_out = GLUMBConv(
             in_channels=in_channels,
             out_channels=in_channels,
+            norm_type="rms_norm",
         )
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:

src/diffusers/models/controlnets/__init__.py

@@ -15,6 +15,7 @@ if is_torch_available():
         SparseControlNetModel,
         SparseControlNetOutput,
     )
+    from .controlnet_union import ControlNetUnionModel
     from .controlnet_xs import ControlNetXSAdapter, ControlNetXSOutput, UNetControlNetXSModel
     from .multicontrolnet import MultiControlNetModel
 

src/diffusers/models/controlnets/controlnet_union.py

@@ -0,0 +1,832 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders.single_file_model import FromOriginalModelMixin
+from ...utils import logging
+from ..attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from ..embeddings import TextImageTimeEmbedding, TextTimeEmbedding, TimestepEmbedding, Timesteps
+from ..modeling_utils import ModelMixin
+from ..unets.unet_2d_blocks import (
+    CrossAttnDownBlock2D,
+    DownBlock2D,
+    UNetMidBlock2DCrossAttn,
+    get_down_block,
+)
+from ..unets.unet_2d_condition import UNet2DConditionModel
+from .controlnet import ControlNetConditioningEmbedding, ControlNetOutput, zero_module
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class QuickGELU(nn.Module):
+    """
+    Applies GELU approximation that is fast but somewhat inaccurate. See: https://github.com/hendrycks/GELUs
+    """
+
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        return input * torch.sigmoid(1.702 * input)
+
+
+class ResidualAttentionMlp(nn.Module):
+    def __init__(self, d_model: int):
+        super().__init__()
+        self.c_fc = nn.Linear(d_model, d_model * 4)
+        self.gelu = QuickGELU()
+        self.c_proj = nn.Linear(d_model * 4, d_model)
+
+    def forward(self, x: torch.Tensor):
+        x = self.c_fc(x)
+        x = self.gelu(x)
+        x = self.c_proj(x)
+        return x
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
+        super().__init__()
+        self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ln_1 = nn.LayerNorm(d_model)
+        self.mlp = ResidualAttentionMlp(d_model)
+        self.ln_2 = nn.LayerNorm(d_model)
+        self.attn_mask = attn_mask
+
+    def attention(self, x: torch.Tensor):
+        self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
+        return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
+
+    def forward(self, x: torch.Tensor):
+        x = x + self.attention(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class ControlNetUnionModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+    """
+    A ControlNetUnion model.
+
+    Args:
+        in_channels (`int`, defaults to 4):
+            The number of channels in the input sample.
+        flip_sin_to_cos (`bool`, defaults to `True`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, defaults to 0):
+            The frequency shift to apply to the time embedding.
+        down_block_types (`tuple[str]`, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        only_cross_attention (`Union[bool, Tuple[bool]]`, defaults to `False`):
+        block_out_channels (`tuple[int]`, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, defaults to 2):
+            The number of layers per block.
+        downsample_padding (`int`, defaults to 1):
+            The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, defaults to 1):
+            The scale factor to use for the mid block.
+        act_fn (`str`, defaults to "silu"):
+            The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups to use for the normalization. If None, normalization and activation layers is skipped
+            in post-processing.
+        norm_eps (`float`, defaults to 1e-5):
+            The epsilon to use for the normalization.
+        cross_attention_dim (`int`, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`Union[int, Tuple[int]]`, defaults to 8):
+            The dimension of the attention heads.
+        use_linear_projection (`bool`, defaults to `False`):
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from None,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        num_class_embeds (`int`, *optional*, defaults to 0):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        upcast_attention (`bool`, defaults to `False`):
+        resnet_time_scale_shift (`str`, defaults to `"default"`):
+            Time scale shift config for ResNet blocks (see `ResnetBlock2D`). Choose from `default` or `scale_shift`.
+        projection_class_embeddings_input_dim (`int`, *optional*, defaults to `None`):
+            The dimension of the `class_labels` input when `class_embed_type="projection"`. Required when
+            `class_embed_type="projection"`.
+        controlnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
+            The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
+        conditioning_embedding_out_channels (`tuple[int]`, *optional*, defaults to `(48, 96, 192, 384)`):
+            The tuple of output channel for each block in the `conditioning_embedding` layer.
+        global_pool_conditions (`bool`, defaults to `False`):
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 4,
+        conditioning_channels: int = 3,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str, ...] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
+        layers_per_block: int = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: int = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int, ...]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        controlnet_conditioning_channel_order: str = "rgb",
+        conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (48, 96, 192, 384),
+        global_pool_conditions: bool = False,
+        addition_embed_type_num_heads: int = 64,
+        num_control_type: int = 6,
+        num_trans_channel: int = 320,
+        num_trans_head: int = 8,
+        num_trans_layer: int = 1,
+        num_proj_channel: int = 320,
+    ):
+        super().__init__()
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        # input
+        conv_in_kernel = 3
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+        )
+
+        # time
+        time_embed_dim = block_out_channels[0] * 4
+        self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+        timestep_input_dim = block_out_channels[0]
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+        )
+
+        if encoder_hid_dim_type is not None:
+            raise ValueError(f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None.")
+        else:
+            self.encoder_hid_proj = None
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kandinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+            self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+
+        elif addition_embed_type is not None:
+            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+        # control net conditioning embedding
+        self.controlnet_cond_embedding = ControlNetConditioningEmbedding(
+            conditioning_embedding_channels=block_out_channels[0],
+            block_out_channels=conditioning_embedding_out_channels,
+            conditioning_channels=conditioning_channels,
+        )
+
+        task_scale_factor = num_trans_channel**0.5
+        self.task_embedding = nn.Parameter(task_scale_factor * torch.randn(num_control_type, num_trans_channel))
+        self.transformer_layes = nn.ModuleList(
+            [ResidualAttentionBlock(num_trans_channel, num_trans_head) for _ in range(num_trans_layer)]
+        )
+        self.spatial_ch_projs = zero_module(nn.Linear(num_trans_channel, num_proj_channel))
+        self.control_type_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+        self.control_add_embedding = TimestepEmbedding(addition_time_embed_dim * num_control_type, time_embed_dim)
+
+        self.down_blocks = nn.ModuleList([])
+        self.controlnet_down_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        # down
+        output_channel = block_out_channels[0]
+
+        controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+        controlnet_block = zero_module(controlnet_block)
+        self.controlnet_down_blocks.append(controlnet_block)
+
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block,
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                num_attention_heads=num_attention_heads[i],
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+                downsample_padding=downsample_padding,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+            )
+            self.down_blocks.append(down_block)
+
+            for _ in range(layers_per_block):
+                controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+                controlnet_block = zero_module(controlnet_block)
+                self.controlnet_down_blocks.append(controlnet_block)
+
+            if not is_final_block:
+                controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+                controlnet_block = zero_module(controlnet_block)
+                self.controlnet_down_blocks.append(controlnet_block)
+
+        # mid
+        mid_block_channel = block_out_channels[-1]
+
+        controlnet_block = nn.Conv2d(mid_block_channel, mid_block_channel, kernel_size=1)
+        controlnet_block = zero_module(controlnet_block)
+        self.controlnet_mid_block = controlnet_block
+
+        self.mid_block = UNetMidBlock2DCrossAttn(
+            transformer_layers_per_block=transformer_layers_per_block[-1],
+            in_channels=mid_block_channel,
+            temb_channels=time_embed_dim,
+            resnet_eps=norm_eps,
+            resnet_act_fn=act_fn,
+            output_scale_factor=mid_block_scale_factor,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads[-1],
+            resnet_groups=norm_num_groups,
+            use_linear_projection=use_linear_projection,
+            upcast_attention=upcast_attention,
+        )
+
+    @classmethod
+    def from_unet(
+        cls,
+        unet: UNet2DConditionModel,
+        controlnet_conditioning_channel_order: str = "rgb",
+        conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+        load_weights_from_unet: bool = True,
+    ):
+        r"""
+        Instantiate a [`ControlNetUnionModel`] from [`UNet2DConditionModel`].
+
+        Parameters:
+            unet (`UNet2DConditionModel`):
+                The UNet model weights to copy to the [`ControlNetUnionModel`]. All configuration options are also
+                copied where applicable.
+        """
+        transformer_layers_per_block = (
+            unet.config.transformer_layers_per_block if "transformer_layers_per_block" in unet.config else 1
+        )
+        encoder_hid_dim = unet.config.encoder_hid_dim if "encoder_hid_dim" in unet.config else None
+        encoder_hid_dim_type = unet.config.encoder_hid_dim_type if "encoder_hid_dim_type" in unet.config else None
+        addition_embed_type = unet.config.addition_embed_type if "addition_embed_type" in unet.config else None
+        addition_time_embed_dim = (
+            unet.config.addition_time_embed_dim if "addition_time_embed_dim" in unet.config else None
+        )
+
+        controlnet = cls(
+            encoder_hid_dim=encoder_hid_dim,
+            encoder_hid_dim_type=encoder_hid_dim_type,
+            addition_embed_type=addition_embed_type,
+            addition_time_embed_dim=addition_time_embed_dim,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=unet.config.in_channels,
+            flip_sin_to_cos=unet.config.flip_sin_to_cos,
+            freq_shift=unet.config.freq_shift,
+            down_block_types=unet.config.down_block_types,
+            only_cross_attention=unet.config.only_cross_attention,
+            block_out_channels=unet.config.block_out_channels,
+            layers_per_block=unet.config.layers_per_block,
+            downsample_padding=unet.config.downsample_padding,
+            mid_block_scale_factor=unet.config.mid_block_scale_factor,
+            act_fn=unet.config.act_fn,
+            norm_num_groups=unet.config.norm_num_groups,
+            norm_eps=unet.config.norm_eps,
+            cross_attention_dim=unet.config.cross_attention_dim,
+            attention_head_dim=unet.config.attention_head_dim,
+            num_attention_heads=unet.config.num_attention_heads,
+            use_linear_projection=unet.config.use_linear_projection,
+            class_embed_type=unet.config.class_embed_type,
+            num_class_embeds=unet.config.num_class_embeds,
+            upcast_attention=unet.config.upcast_attention,
+            resnet_time_scale_shift=unet.config.resnet_time_scale_shift,
+            projection_class_embeddings_input_dim=unet.config.projection_class_embeddings_input_dim,
+            controlnet_conditioning_channel_order=controlnet_conditioning_channel_order,
+            conditioning_embedding_out_channels=conditioning_embedding_out_channels,
+        )
+
+        if load_weights_from_unet:
+            controlnet.conv_in.load_state_dict(unet.conv_in.state_dict())
+            controlnet.time_proj.load_state_dict(unet.time_proj.state_dict())
+            controlnet.time_embedding.load_state_dict(unet.time_embedding.state_dict())
+
+            if controlnet.class_embedding:
+                controlnet.class_embedding.load_state_dict(unet.class_embedding.state_dict())
+
+            controlnet.down_blocks.load_state_dict(unet.down_blocks.state_dict(), strict=False)
+            controlnet.mid_block.load_state_dict(unet.mid_block.state_dict(), strict=False)
+
+        return controlnet
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attention_slice
+    def set_attention_slice(self, slice_size: Union[str, int, List[int]]) -> None:
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+        if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D)):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        controlnet_cond: List[torch.Tensor],
+        control_type: torch.Tensor,
+        control_type_idx: List[int],
+        conditioning_scale: float = 1.0,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guess_mode: bool = False,
+        return_dict: bool = True,
+    ) -> Union[ControlNetOutput, Tuple[Tuple[torch.Tensor, ...], torch.Tensor]]:
+        """
+        The [`ControlNetUnionModel`] forward method.
+
+        Args:
+            sample (`torch.Tensor`):
+                The noisy input tensor.
+            timestep (`Union[torch.Tensor, float, int]`):
+                The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.Tensor`):
+                The encoder hidden states.
+            controlnet_cond (`List[torch.Tensor]`):
+                The conditional input tensors.
+            control_type (`torch.Tensor`):
+                A tensor of shape `(batch, num_control_type)` with values `0` or `1` depending on whether the control
+                type is used.
+            control_type_idx (`List[int]`):
+                The indices of `control_type`.
+            conditioning_scale (`float`, defaults to `1.0`):
+                The scale factor for ControlNet outputs.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
+                Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
+                timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
+                embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            added_cond_kwargs (`dict`):
+                Additional conditions for the Stable Diffusion XL UNet.
+            cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
+                A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
+            guess_mode (`bool`, defaults to `False`):
+                In this mode, the ControlNet encoder tries its best to recognize the input content of the input even if
+                you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
+            return_dict (`bool`, defaults to `True`):
+                Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.controlnet.ControlNetOutput`] **or** `tuple`:
+                If `return_dict` is `True`, a [`~models.controlnet.ControlNetOutput`] is returned, otherwise a tuple is
+                returned where the first element is the sample tensor.
+        """
+        # check channel order
+        channel_order = self.config.controlnet_conditioning_channel_order
+
+        if channel_order != "rgb":
+            raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}")
+
+        # prepare attention_mask
+        if attention_mask is not None:
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+            emb = emb + class_emb
+
+        if self.config.addition_embed_type is not None:
+            if self.config.addition_embed_type == "text":
+                aug_emb = self.add_embedding(encoder_hidden_states)
+
+            elif self.config.addition_embed_type == "text_time":
+                if "text_embeds" not in added_cond_kwargs:
+                    raise ValueError(
+                        f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                    )
+                text_embeds = added_cond_kwargs.get("text_embeds")
+                if "time_ids" not in added_cond_kwargs:
+                    raise ValueError(
+                        f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                    )
+                time_ids = added_cond_kwargs.get("time_ids")
+                time_embeds = self.add_time_proj(time_ids.flatten())
+                time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+
+                add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+                add_embeds = add_embeds.to(emb.dtype)
+                aug_emb = self.add_embedding(add_embeds)
+
+        control_embeds = self.control_type_proj(control_type.flatten())
+        control_embeds = control_embeds.reshape((t_emb.shape[0], -1))
+        control_embeds = control_embeds.to(emb.dtype)
+        control_emb = self.control_add_embedding(control_embeds)
+        emb = emb + control_emb
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        inputs = []
+        condition_list = []
+
+        for cond, control_idx in zip(controlnet_cond, control_type_idx):
+            condition = self.controlnet_cond_embedding(cond)
+            feat_seq = torch.mean(condition, dim=(2, 3))
+            feat_seq = feat_seq + self.task_embedding[control_idx]
+            inputs.append(feat_seq.unsqueeze(1))
+            condition_list.append(condition)
+
+        condition = sample
+        feat_seq = torch.mean(condition, dim=(2, 3))
+        inputs.append(feat_seq.unsqueeze(1))
+        condition_list.append(condition)
+
+        x = torch.cat(inputs, dim=1)
+        for layer in self.transformer_layes:
+            x = layer(x)
+
+        controlnet_cond_fuser = sample * 0.0
+        for idx, condition in enumerate(condition_list[:-1]):
+            alpha = self.spatial_ch_projs(x[:, idx])
+            alpha = alpha.unsqueeze(-1).unsqueeze(-1)
+            controlnet_cond_fuser += condition + alpha
+
+        sample = sample + controlnet_cond_fuser
+
+        # 3. down
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+            down_block_res_samples += res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            sample = self.mid_block(
+                sample,
+                emb,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=attention_mask,
+                cross_attention_kwargs=cross_attention_kwargs,
+            )
+
+        # 5. Control net blocks
+        controlnet_down_block_res_samples = ()
+
+        for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
+            down_block_res_sample = controlnet_block(down_block_res_sample)
+            controlnet_down_block_res_samples = controlnet_down_block_res_samples + (down_block_res_sample,)
+
+        down_block_res_samples = controlnet_down_block_res_samples
+
+        mid_block_res_sample = self.controlnet_mid_block(sample)
+
+        # 6. scaling
+        if guess_mode and not self.config.global_pool_conditions:
+            scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1, device=sample.device)  # 0.1 to 1.0
+            scales = scales * conditioning_scale
+            down_block_res_samples = [sample * scale for sample, scale in zip(down_block_res_samples, scales)]
+            mid_block_res_sample = mid_block_res_sample * scales[-1]  # last one
+        else:
+            down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples]
+            mid_block_res_sample = mid_block_res_sample * conditioning_scale
+
+        if self.config.global_pool_conditions:
+            down_block_res_samples = [
+                torch.mean(sample, dim=(2, 3), keepdim=True) for sample in down_block_res_samples
+            ]
+            mid_block_res_sample = torch.mean(mid_block_res_sample, dim=(2, 3), keepdim=True)
+
+        if not return_dict:
+            return (down_block_res_samples, mid_block_res_sample)
+
+        return ControlNetOutput(
+            down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample
+        )

src/diffusers/models/embeddings.py

@@ -84,6 +84,78 @@ def get_3d_sincos_pos_embed(
     temporal_size: int,
     spatial_interpolation_scale: float = 1.0,
     temporal_interpolation_scale: float = 1.0,
+    device: Optional[torch.device] = None,
+    output_type: str = "np",
+) -> torch.Tensor:
+    r"""
+    Creates 3D sinusoidal positional embeddings.
+
+    Args:
+        embed_dim (`int`):
+            The embedding dimension of inputs. It must be divisible by 16.
+        spatial_size (`int` or `Tuple[int, int]`):
+            The spatial dimension of positional embeddings. If an integer is provided, the same size is applied to both
+            spatial dimensions (height and width).
+        temporal_size (`int`):
+            The temporal dimension of postional embeddings (number of frames).
+        spatial_interpolation_scale (`float`, defaults to 1.0):
+            Scale factor for spatial grid interpolation.
+        temporal_interpolation_scale (`float`, defaults to 1.0):
+            Scale factor for temporal grid interpolation.
+
+    Returns:
+        `torch.Tensor`:
+            The 3D sinusoidal positional embeddings of shape `[temporal_size, spatial_size[0] * spatial_size[1],
+            embed_dim]`.
+    """
+    if output_type == "np":
+        return _get_3d_sincos_pos_embed_np(
+            embed_dim=embed_dim,
+            spatial_size=spatial_size,
+            temporal_size=temporal_size,
+            spatial_interpolation_scale=spatial_interpolation_scale,
+            temporal_interpolation_scale=temporal_interpolation_scale,
+        )
+    if embed_dim % 4 != 0:
+        raise ValueError("`embed_dim` must be divisible by 4")
+    if isinstance(spatial_size, int):
+        spatial_size = (spatial_size, spatial_size)
+
+    embed_dim_spatial = 3 * embed_dim // 4
+    embed_dim_temporal = embed_dim // 4
+
+    # 1. Spatial
+    grid_h = torch.arange(spatial_size[1], device=device, dtype=torch.float32) / spatial_interpolation_scale
+    grid_w = torch.arange(spatial_size[0], device=device, dtype=torch.float32) / spatial_interpolation_scale
+    grid = torch.meshgrid(grid_w, grid_h, indexing="xy")  # here w goes first
+    grid = torch.stack(grid, dim=0)
+
+    grid = grid.reshape([2, 1, spatial_size[1], spatial_size[0]])
+    pos_embed_spatial = get_2d_sincos_pos_embed_from_grid(embed_dim_spatial, grid, output_type="pt")
+
+    # 2. Temporal
+    grid_t = torch.arange(temporal_size, device=device, dtype=torch.float32) / temporal_interpolation_scale
+    pos_embed_temporal = get_1d_sincos_pos_embed_from_grid(embed_dim_temporal, grid_t, output_type="pt")
+
+    # 3. Concat
+    pos_embed_spatial = pos_embed_spatial[None, :, :]
+    pos_embed_spatial = pos_embed_spatial.repeat_interleave(temporal_size, dim=0)  # [T, H*W, D // 4 * 3]
+
+    pos_embed_temporal = pos_embed_temporal[:, None, :]
+    pos_embed_temporal = pos_embed_temporal.repeat_interleave(
+        spatial_size[0] * spatial_size[1], dim=1
+    )  # [T, H*W, D // 4]
+
+    pos_embed = torch.concat([pos_embed_temporal, pos_embed_spatial], dim=-1)  # [T, H*W, D]
+    return pos_embed
+
+
+def _get_3d_sincos_pos_embed_np(
+    embed_dim: int,
+    spatial_size: Union[int, Tuple[int, int]],
+    temporal_size: int,
+    spatial_interpolation_scale: float = 1.0,
+    temporal_interpolation_scale: float = 1.0,
 ) -> np.ndarray:
     r"""
     Creates 3D sinusoidal positional embeddings.
@@ -106,6 +178,12 @@ def get_3d_sincos_pos_embed(
             The 3D sinusoidal positional embeddings of shape `[temporal_size, spatial_size[0] * spatial_size[1],
             embed_dim]`.
     """
+    deprecation_message = (
+        "`get_3d_sincos_pos_embed` uses `torch` and supports `device`."
+        " `from_numpy` is no longer required."
+        "  Pass `output_type='pt' to use the new version now."
+    )
+    deprecate("output_type=='np'", "0.33.0", deprecation_message, standard_warn=False)
     if embed_dim % 4 != 0:
         raise ValueError("`embed_dim` must be divisible by 4")
     if isinstance(spatial_size, int):
@@ -139,6 +217,143 @@ def get_3d_sincos_pos_embed(
 
 
 def get_2d_sincos_pos_embed(
+    embed_dim,
+    grid_size,
+    cls_token=False,
+    extra_tokens=0,
+    interpolation_scale=1.0,
+    base_size=16,
+    device: Optional[torch.device] = None,
+    output_type: str = "np",
+):
+    """
+    Creates 2D sinusoidal positional embeddings.
+
+    Args:
+        embed_dim (`int`):
+            The embedding dimension.
+        grid_size (`int`):
+            The size of the grid height and width.
+        cls_token (`bool`, defaults to `False`):
+            Whether or not to add a classification token.
+        extra_tokens (`int`, defaults to `0`):
+            The number of extra tokens to add.
+        interpolation_scale (`float`, defaults to `1.0`):
+            The scale of the interpolation.
+
+    Returns:
+        pos_embed (`torch.Tensor`):
+            Shape is either `[grid_size * grid_size, embed_dim]` if not using cls_token, or `[1 + grid_size*grid_size,
+            embed_dim]` if using cls_token
+    """
+    if output_type == "np":
+        deprecation_message = (
+            "`get_2d_sincos_pos_embed` uses `torch` and supports `device`."
+            " `from_numpy` is no longer required."
+            "  Pass `output_type='pt' to use the new version now."
+        )
+        deprecate("output_type=='np'", "0.33.0", deprecation_message, standard_warn=False)
+        return get_2d_sincos_pos_embed_np(
+            embed_dim=embed_dim,
+            grid_size=grid_size,
+            cls_token=cls_token,
+            extra_tokens=extra_tokens,
+            interpolation_scale=interpolation_scale,
+            base_size=base_size,
+        )
+    if isinstance(grid_size, int):
+        grid_size = (grid_size, grid_size)
+
+    grid_h = (
+        torch.arange(grid_size[0], device=device, dtype=torch.float32)
+        / (grid_size[0] / base_size)
+        / interpolation_scale
+    )
+    grid_w = (
+        torch.arange(grid_size[1], device=device, dtype=torch.float32)
+        / (grid_size[1] / base_size)
+        / interpolation_scale
+    )
+    grid = torch.meshgrid(grid_w, grid_h, indexing="xy")  # here w goes first
+    grid = torch.stack(grid, dim=0)
+
+    grid = grid.reshape([2, 1, grid_size[1], grid_size[0]])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid, output_type=output_type)
+    if cls_token and extra_tokens > 0:
+        pos_embed = torch.concat([torch.zeros([extra_tokens, embed_dim]), pos_embed], dim=0)
+    return pos_embed
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid, output_type="np"):
+    r"""
+    This function generates 2D sinusoidal positional embeddings from a grid.
+
+    Args:
+        embed_dim (`int`): The embedding dimension.
+        grid (`torch.Tensor`): Grid of positions with shape `(H * W,)`.
+
+    Returns:
+        `torch.Tensor`: The 2D sinusoidal positional embeddings with shape `(H * W, embed_dim)`
+    """
+    if output_type == "np":
+        deprecation_message = (
+            "`get_2d_sincos_pos_embed_from_grid` uses `torch` and supports `device`."
+            " `from_numpy` is no longer required."
+            "  Pass `output_type='pt' to use the new version now."
+        )
+        deprecate("output_type=='np'", "0.33.0", deprecation_message, standard_warn=False)
+        return get_2d_sincos_pos_embed_from_grid_np(
+            embed_dim=embed_dim,
+            grid=grid,
+        )
+    if embed_dim % 2 != 0:
+        raise ValueError("embed_dim must be divisible by 2")
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0], output_type=output_type)  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1], output_type=output_type)  # (H*W, D/2)
+
+    emb = torch.concat([emb_h, emb_w], dim=1)  # (H*W, D)
+    return emb
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos, output_type="np"):
+    """
+    This function generates 1D positional embeddings from a grid.
+
+    Args:
+        embed_dim (`int`): The embedding dimension `D`
+        pos (`torch.Tensor`): 1D tensor of positions with shape `(M,)`
+
+    Returns:
+        `torch.Tensor`: Sinusoidal positional embeddings of shape `(M, D)`.
+    """
+    if output_type == "np":
+        deprecation_message = (
+            "`get_1d_sincos_pos_embed_from_grid` uses `torch` and supports `device`."
+            " `from_numpy` is no longer required."
+            "  Pass `output_type='pt' to use the new version now."
+        )
+        deprecate("output_type=='np'", "0.33.0", deprecation_message, standard_warn=False)
+        return get_1d_sincos_pos_embed_from_grid_np(embed_dim=embed_dim, pos=pos)
+    if embed_dim % 2 != 0:
+        raise ValueError("embed_dim must be divisible by 2")
+
+    omega = torch.arange(embed_dim // 2, device=pos.device, dtype=torch.float64)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = torch.outer(pos, omega)  # (M, D/2), outer product
+
+    emb_sin = torch.sin(out)  # (M, D/2)
+    emb_cos = torch.cos(out)  # (M, D/2)
+
+    emb = torch.concat([emb_sin, emb_cos], dim=1)  # (M, D)
+    return emb
+
+
+def get_2d_sincos_pos_embed_np(
     embed_dim, grid_size, cls_token=False, extra_tokens=0, interpolation_scale=1.0, base_size=16
 ):
     """
@@ -170,13 +385,13 @@ def get_2d_sincos_pos_embed(
     grid = np.stack(grid, axis=0)
 
     grid = grid.reshape([2, 1, grid_size[1], grid_size[0]])
-    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    pos_embed = get_2d_sincos_pos_embed_from_grid_np(embed_dim, grid)
     if cls_token and extra_tokens > 0:
         pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
     return pos_embed
 
 
-def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+def get_2d_sincos_pos_embed_from_grid_np(embed_dim, grid):
     r"""
     This function generates 2D sinusoidal positional embeddings from a grid.
 
@@ -191,14 +406,14 @@ def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
         raise ValueError("embed_dim must be divisible by 2")
 
     # use half of dimensions to encode grid_h
-    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
-    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+    emb_h = get_1d_sincos_pos_embed_from_grid_np(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid_np(embed_dim // 2, grid[1])  # (H*W, D/2)
 
     emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)
     return emb
 
 
-def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+def get_1d_sincos_pos_embed_from_grid_np(embed_dim, pos):
     """
     This function generates 1D positional embeddings from a grid.
 
@@ -288,10 +503,14 @@ class PatchEmbed(nn.Module):
             self.pos_embed = None
         elif pos_embed_type == "sincos":
             pos_embed = get_2d_sincos_pos_embed(
-                embed_dim, grid_size, base_size=self.base_size, interpolation_scale=self.interpolation_scale
+                embed_dim,
+                grid_size,
+                base_size=self.base_size,
+                interpolation_scale=self.interpolation_scale,
+                output_type="pt",
             )
             persistent = True if pos_embed_max_size else False
-            self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=persistent)
+            self.register_buffer("pos_embed", pos_embed.float().unsqueeze(0), persistent=persistent)
         else:
             raise ValueError(f"Unsupported pos_embed_type: {pos_embed_type}")
 
@@ -341,8 +560,10 @@ class PatchEmbed(nn.Module):
                     grid_size=(height, width),
                     base_size=self.base_size,
                     interpolation_scale=self.interpolation_scale,
+                    device=latent.device,
+                    output_type="pt",
                 )
-                pos_embed = torch.from_numpy(pos_embed).float().unsqueeze(0).to(latent.device)
+                pos_embed = pos_embed.float().unsqueeze(0)
             else:
                 pos_embed = self.pos_embed
 
@@ -453,7 +674,9 @@ class CogVideoXPatchEmbed(nn.Module):
             pos_embedding = self._get_positional_embeddings(sample_height, sample_width, sample_frames)
             self.register_buffer("pos_embedding", pos_embedding, persistent=persistent)
 
-    def _get_positional_embeddings(self, sample_height: int, sample_width: int, sample_frames: int) -> torch.Tensor:
+    def _get_positional_embeddings(
+        self, sample_height: int, sample_width: int, sample_frames: int, device: Optional[torch.device] = None
+    ) -> torch.Tensor:
         post_patch_height = sample_height // self.patch_size
         post_patch_width = sample_width // self.patch_size
         post_time_compression_frames = (sample_frames - 1) // self.temporal_compression_ratio + 1
@@ -465,8 +688,10 @@ class CogVideoXPatchEmbed(nn.Module):
             post_time_compression_frames,
             self.spatial_interpolation_scale,
             self.temporal_interpolation_scale,
+            device=device,
+            output_type="pt",
         )
-        pos_embedding = torch.from_numpy(pos_embedding).flatten(0, 1)
+        pos_embedding = pos_embedding.flatten(0, 1)
         joint_pos_embedding = torch.zeros(
             1, self.max_text_seq_length + num_patches, self.embed_dim, requires_grad=False
         )
@@ -521,8 +746,10 @@ class CogVideoXPatchEmbed(nn.Module):
                 or self.sample_width != width
                 or self.sample_frames != pre_time_compression_frames
             ):
-                pos_embedding = self._get_positional_embeddings(height, width, pre_time_compression_frames)
-                pos_embedding = pos_embedding.to(embeds.device, dtype=embeds.dtype)
+                pos_embedding = self._get_positional_embeddings(
+                    height, width, pre_time_compression_frames, device=embeds.device
+                )
+                pos_embedding = pos_embedding.to(dtype=embeds.dtype)
             else:
                 pos_embedding = self.pos_embedding
 
@@ -552,9 +779,11 @@ class CogView3PlusPatchEmbed(nn.Module):
         # Linear projection for text embeddings
         self.text_proj = nn.Linear(text_hidden_size, hidden_size)
 
-        pos_embed = get_2d_sincos_pos_embed(hidden_size, pos_embed_max_size, base_size=pos_embed_max_size)
+        pos_embed = get_2d_sincos_pos_embed(
+            hidden_size, pos_embed_max_size, base_size=pos_embed_max_size, output_type="pt"
+        )
         pos_embed = pos_embed.reshape(pos_embed_max_size, pos_embed_max_size, hidden_size)
-        self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float(), persistent=False)
+        self.register_buffer("pos_embed", pos_embed.float(), persistent=False)
 
     def forward(self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor) -> torch.Tensor:
         batch_size, channel, height, width = hidden_states.shape

src/diffusers/models/transformers/__init__.py

@@ -11,12 +11,14 @@ if is_torch_available():
     from .lumina_nextdit2d import LuminaNextDiT2DModel
     from .pixart_transformer_2d import PixArtTransformer2DModel
     from .prior_transformer import PriorTransformer
+    from .sana_transformer import SanaTransformer2DModel
     from .stable_audio_transformer import StableAudioDiTModel
     from .t5_film_transformer import T5FilmDecoder
     from .transformer_2d import Transformer2DModel
     from .transformer_allegro import AllegroTransformer3DModel
     from .transformer_cogview3plus import CogView3PlusTransformer2DModel
     from .transformer_flux import FluxTransformer2DModel
+    from .transformer_ltx import LTXVideoTransformer3DModel
     from .transformer_mochi import MochiTransformer3DModel
     from .transformer_sd3 import SD3Transformer2DModel
     from .transformer_temporal import TransformerTemporalModel

src/diffusers/models/transformers/latte_transformer_3d.py

@@ -156,9 +156,9 @@ class LatteTransformer3DModel(ModelMixin, ConfigMixin):
 
         # define temporal positional embedding
         temp_pos_embed = get_1d_sincos_pos_embed_from_grid(
-            inner_dim, torch.arange(0, video_length).unsqueeze(1)
+            inner_dim, torch.arange(0, video_length).unsqueeze(1), output_type="pt"
         )  # 1152 hidden size
-        self.register_buffer("temp_pos_embed", torch.from_numpy(temp_pos_embed).float().unsqueeze(0), persistent=False)
+        self.register_buffer("temp_pos_embed", temp_pos_embed.float().unsqueeze(0), persistent=False)
 
         self.gradient_checkpointing = False
 

src/diffusers/models/transformers/sana_transformer.py

@@ -0,0 +1,465 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import is_torch_version, logging
+from ..attention_processor import (
+    Attention,
+    AttentionProcessor,
+    AttnProcessor2_0,
+    SanaLinearAttnProcessor2_0,
+)
+from ..embeddings import PatchEmbed, PixArtAlphaTextProjection
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormSingle, RMSNorm
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class GLUMBConv(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        expand_ratio: float = 4,
+        norm_type: Optional[str] = None,
+        residual_connection: bool = True,
+    ) -> None:
+        super().__init__()
+
+        hidden_channels = int(expand_ratio * in_channels)
+        self.norm_type = norm_type
+        self.residual_connection = residual_connection
+
+        self.nonlinearity = nn.SiLU()
+        self.conv_inverted = nn.Conv2d(in_channels, hidden_channels * 2, 1, 1, 0)
+        self.conv_depth = nn.Conv2d(hidden_channels * 2, hidden_channels * 2, 3, 1, 1, groups=hidden_channels * 2)
+        self.conv_point = nn.Conv2d(hidden_channels, out_channels, 1, 1, 0, bias=False)
+
+        self.norm = None
+        if norm_type == "rms_norm":
+            self.norm = RMSNorm(out_channels, eps=1e-5, elementwise_affine=True, bias=True)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if self.residual_connection:
+            residual = hidden_states
+
+        hidden_states = self.conv_inverted(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv_depth(hidden_states)
+        hidden_states, gate = torch.chunk(hidden_states, 2, dim=1)
+        hidden_states = hidden_states * self.nonlinearity(gate)
+
+        hidden_states = self.conv_point(hidden_states)
+
+        if self.norm_type == "rms_norm":
+            # move channel to the last dimension so we apply RMSnorm across channel dimension
+            hidden_states = self.norm(hidden_states.movedim(1, -1)).movedim(-1, 1)
+
+        if self.residual_connection:
+            hidden_states = hidden_states + residual
+
+        return hidden_states
+
+
+class SanaTransformerBlock(nn.Module):
+    r"""
+    Transformer block introduced in [Sana](https://huggingface.co/papers/2410.10629).
+    """
+
+    def __init__(
+        self,
+        dim: int = 2240,
+        num_attention_heads: int = 70,
+        attention_head_dim: int = 32,
+        dropout: float = 0.0,
+        num_cross_attention_heads: Optional[int] = 20,
+        cross_attention_head_dim: Optional[int] = 112,
+        cross_attention_dim: Optional[int] = 2240,
+        attention_bias: bool = True,
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-6,
+        attention_out_bias: bool = True,
+        mlp_ratio: float = 2.5,
+    ) -> None:
+        super().__init__()
+
+        # 1. Self Attention
+        self.norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=norm_eps)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=None,
+            processor=SanaLinearAttnProcessor2_0(),
+        )
+
+        # 2. Cross Attention
+        if cross_attention_dim is not None:
+            self.norm2 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_cross_attention_heads,
+                dim_head=cross_attention_head_dim,
+                dropout=dropout,
+                bias=True,
+                out_bias=attention_out_bias,
+                processor=AttnProcessor2_0(),
+            )
+
+        # 3. Feed-forward
+        self.ff = GLUMBConv(dim, dim, mlp_ratio, norm_type=None, residual_connection=False)
+
+        self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        height: int = None,
+        width: int = None,
+    ) -> torch.Tensor:
+        batch_size = hidden_states.shape[0]
+
+        # 1. Modulation
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+            self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+        ).chunk(6, dim=1)
+
+        # 2. Self Attention
+        norm_hidden_states = self.norm1(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+        norm_hidden_states = norm_hidden_states.to(hidden_states.dtype)
+
+        attn_output = self.attn1(norm_hidden_states)
+        hidden_states = hidden_states + gate_msa * attn_output
+
+        # 3. Cross Attention
+        if self.attn2 is not None:
+            attn_output = self.attn2(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        norm_hidden_states = norm_hidden_states.unflatten(1, (height, width)).permute(0, 3, 1, 2)
+        ff_output = self.ff(norm_hidden_states)
+        ff_output = ff_output.flatten(2, 3).permute(0, 2, 1)
+        hidden_states = hidden_states + gate_mlp * ff_output
+
+        return hidden_states
+
+
+class SanaTransformer2DModel(ModelMixin, ConfigMixin):
+    r"""
+    A 2D Transformer model introduced in [Sana](https://huggingface.co/papers/2410.10629) family of models.
+
+    Args:
+        in_channels (`int`, defaults to `32`):
+            The number of channels in the input.
+        out_channels (`int`, *optional*, defaults to `32`):
+            The number of channels in the output.
+        num_attention_heads (`int`, defaults to `70`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, defaults to `32`):
+            The number of channels in each head.
+        num_layers (`int`, defaults to `20`):
+            The number of layers of Transformer blocks to use.
+        num_cross_attention_heads (`int`, *optional*, defaults to `20`):
+            The number of heads to use for cross-attention.
+        cross_attention_head_dim (`int`, *optional*, defaults to `112`):
+            The number of channels in each head for cross-attention.
+        cross_attention_dim (`int`, *optional*, defaults to `2240`):
+            The number of channels in the cross-attention output.
+        caption_channels (`int`, defaults to `2304`):
+            The number of channels in the caption embeddings.
+        mlp_ratio (`float`, defaults to `2.5`):
+            The expansion ratio to use in the GLUMBConv layer.
+        dropout (`float`, defaults to `0.0`):
+            The dropout probability.
+        attention_bias (`bool`, defaults to `False`):
+            Whether to use bias in the attention layer.
+        sample_size (`int`, defaults to `32`):
+            The base size of the input latent.
+        patch_size (`int`, defaults to `1`):
+            The size of the patches to use in the patch embedding layer.
+        norm_elementwise_affine (`bool`, defaults to `False`):
+            Whether to use elementwise affinity in the normalization layer.
+        norm_eps (`float`, defaults to `1e-6`):
+            The epsilon value for the normalization layer.
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["SanaTransformerBlock", "PatchEmbed"]
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 32,
+        out_channels: Optional[int] = 32,
+        num_attention_heads: int = 70,
+        attention_head_dim: int = 32,
+        num_layers: int = 20,
+        num_cross_attention_heads: Optional[int] = 20,
+        cross_attention_head_dim: Optional[int] = 112,
+        cross_attention_dim: Optional[int] = 2240,
+        caption_channels: int = 2304,
+        mlp_ratio: float = 2.5,
+        dropout: float = 0.0,
+        attention_bias: bool = False,
+        sample_size: int = 32,
+        patch_size: int = 1,
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-6,
+    ) -> None:
+        super().__init__()
+
+        out_channels = out_channels or in_channels
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # 1. Patch Embedding
+        self.patch_embed = PatchEmbed(
+            height=sample_size,
+            width=sample_size,
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=inner_dim,
+            interpolation_scale=None,
+            pos_embed_type=None,
+        )
+
+        # 2. Additional condition embeddings
+        self.time_embed = AdaLayerNormSingle(inner_dim)
+
+        self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+        self.caption_norm = RMSNorm(inner_dim, eps=1e-5, elementwise_affine=True)
+
+        # 3. Transformer blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                SanaTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    num_cross_attention_heads=num_cross_attention_heads,
+                    cross_attention_head_dim=cross_attention_head_dim,
+                    cross_attention_dim=cross_attention_dim,
+                    attention_bias=attention_bias,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    mlp_ratio=mlp_ratio,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 4. Output blocks
+        self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+
+        self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+        self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * out_channels)
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[Tuple[torch.Tensor, ...], Transformer2DModelOutput]:
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 1. Input
+        batch_size, num_channels, height, width = hidden_states.shape
+        p = self.config.patch_size
+        post_patch_height, post_patch_width = height // p, width // p
+
+        hidden_states = self.patch_embed(hidden_states)
+
+        timestep, embedded_timestep = self.time_embed(
+            timestep, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+        )
+
+        encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+        encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        encoder_hidden_states = self.caption_norm(encoder_hidden_states)
+
+        # 2. Transformer blocks
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+
+            def create_custom_forward(module, return_dict=None):
+                def custom_forward(*inputs):
+                    if return_dict is not None:
+                        return module(*inputs, return_dict=return_dict)
+                    else:
+                        return module(*inputs)
+
+                return custom_forward
+
+            ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+
+            for block in self.transformer_blocks:
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    post_patch_height,
+                    post_patch_width,
+                    **ckpt_kwargs,
+                )
+
+        else:
+            for block in self.transformer_blocks:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    post_patch_height,
+                    post_patch_width,
+                )
+
+        # 3. Normalization
+        shift, scale = (
+            self.scale_shift_table[None] + embedded_timestep[:, None].to(self.scale_shift_table.device)
+        ).chunk(2, dim=1)
+        hidden_states = self.norm_out(hidden_states)
+
+        # 4. Modulation
+        hidden_states = hidden_states * (1 + scale) + shift
+        hidden_states = self.proj_out(hidden_states)
+
+        # 5. Unpatchify
+        hidden_states = hidden_states.reshape(
+            batch_size, post_patch_height, post_patch_width, self.config.patch_size, self.config.patch_size, -1
+        )
+        hidden_states = hidden_states.permute(0, 5, 1, 3, 2, 4)
+        output = hidden_states.reshape(batch_size, -1, post_patch_height * p, post_patch_width * p)
+
+        if not return_dict:
+            return (output,)
+        return Transformer2DModelOutput(sample=output)

src/diffusers/models/transformers/transformer_ltx.py

@@ -0,0 +1,449 @@
+# Copyright 2024 The Genmo team and The HuggingFace Team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Any, Dict, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalModelMixin
+from ...utils import is_torch_version, logging
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention import FeedForward
+from ..attention_processor import Attention
+from ..embeddings import PixArtAlphaTextProjection
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormSingle, RMSNorm
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class LTXAttentionProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). This is
+    used in the LTX model. It applies a normalization layer and rotary embedding on the query and key vector.
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "LTXAttentionProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+            )
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.norm_q(query)
+        key = attn.norm_k(key)
+
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb)
+            key = apply_rotary_emb(key, image_rotary_emb)
+
+        query = query.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        key = key.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        value = value.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+        hidden_states = hidden_states.transpose(1, 2).flatten(2, 3)
+        hidden_states = hidden_states.to(query.dtype)
+
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+        return hidden_states
+
+
+class LTXRotaryPosEmbed(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        base_num_frames: int = 20,
+        base_height: int = 2048,
+        base_width: int = 2048,
+        patch_size: int = 1,
+        patch_size_t: int = 1,
+        theta: float = 10000.0,
+    ) -> None:
+        super().__init__()
+
+        self.dim = dim
+        self.base_num_frames = base_num_frames
+        self.base_height = base_height
+        self.base_width = base_width
+        self.patch_size = patch_size
+        self.patch_size_t = patch_size_t
+        self.theta = theta
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        num_frames: int,
+        height: int,
+        width: int,
+        rope_interpolation_scale: Optional[Tuple[torch.Tensor, float, float]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        batch_size = hidden_states.size(0)
+
+        # Always compute rope in fp32
+        grid_h = torch.arange(height, dtype=torch.float32, device=hidden_states.device)
+        grid_w = torch.arange(width, dtype=torch.float32, device=hidden_states.device)
+        grid_f = torch.arange(num_frames, dtype=torch.float32, device=hidden_states.device)
+        grid = torch.meshgrid(grid_f, grid_h, grid_w, indexing="ij")
+        grid = torch.stack(grid, dim=0)
+        grid = grid.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1)
+
+        if rope_interpolation_scale is not None:
+            grid[:, 0:1] = grid[:, 0:1] * rope_interpolation_scale[0] * self.patch_size_t / self.base_num_frames
+            grid[:, 1:2] = grid[:, 1:2] * rope_interpolation_scale[1] * self.patch_size / self.base_height
+            grid[:, 2:3] = grid[:, 2:3] * rope_interpolation_scale[2] * self.patch_size / self.base_width
+
+        grid = grid.flatten(2, 4).transpose(1, 2)
+
+        start = 1.0
+        end = self.theta
+        freqs = self.theta ** torch.linspace(
+            math.log(start, self.theta),
+            math.log(end, self.theta),
+            self.dim // 6,
+            device=hidden_states.device,
+            dtype=torch.float32,
+        )
+        freqs = freqs * math.pi / 2.0
+        freqs = freqs * (grid.unsqueeze(-1) * 2 - 1)
+        freqs = freqs.transpose(-1, -2).flatten(2)
+
+        cos_freqs = freqs.cos().repeat_interleave(2, dim=-1)
+        sin_freqs = freqs.sin().repeat_interleave(2, dim=-1)
+
+        if self.dim % 6 != 0:
+            cos_padding = torch.ones_like(cos_freqs[:, :, : self.dim % 6])
+            sin_padding = torch.zeros_like(cos_freqs[:, :, : self.dim % 6])
+            cos_freqs = torch.cat([cos_padding, cos_freqs], dim=-1)
+            sin_freqs = torch.cat([sin_padding, sin_freqs], dim=-1)
+
+        return cos_freqs, sin_freqs
+
+
+@maybe_allow_in_graph
+class LTXTransformerBlock(nn.Module):
+    r"""
+    Transformer block used in [LTX](https://huggingface.co/Lightricks/LTX-Video).
+
+    Args:
+        dim (`int`):
+            The number of channels in the input and output.
+        num_attention_heads (`int`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`):
+            The number of channels in each head.
+        qk_norm (`str`, defaults to `"rms_norm"`):
+            The normalization layer to use.
+        activation_fn (`str`, defaults to `"gelu-approximate"`):
+            Activation function to use in feed-forward.
+        eps (`float`, defaults to `1e-6`):
+            Epsilon value for normalization layers.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        cross_attention_dim: int,
+        qk_norm: str = "rms_norm_across_heads",
+        activation_fn: str = "gelu-approximate",
+        attention_bias: bool = True,
+        attention_out_bias: bool = True,
+        eps: float = 1e-6,
+        elementwise_affine: bool = False,
+    ):
+        super().__init__()
+
+        self.norm1 = RMSNorm(dim, eps=eps, elementwise_affine=elementwise_affine)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            kv_heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            bias=attention_bias,
+            cross_attention_dim=None,
+            out_bias=attention_out_bias,
+            qk_norm=qk_norm,
+            processor=LTXAttentionProcessor2_0(),
+        )
+
+        self.norm2 = RMSNorm(dim, eps=eps, elementwise_affine=elementwise_affine)
+        self.attn2 = Attention(
+            query_dim=dim,
+            cross_attention_dim=cross_attention_dim,
+            heads=num_attention_heads,
+            kv_heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            bias=attention_bias,
+            out_bias=attention_out_bias,
+            qk_norm=qk_norm,
+            processor=LTXAttentionProcessor2_0(),
+        )
+
+        self.ff = FeedForward(dim, activation_fn=activation_fn)
+
+        self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size = hidden_states.size(0)
+        norm_hidden_states = self.norm1(hidden_states)
+
+        num_ada_params = self.scale_shift_table.shape[0]
+        ada_values = self.scale_shift_table[None, None] + temb.reshape(batch_size, temb.size(1), num_ada_params, -1)
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ada_values.unbind(dim=2)
+        norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+
+        attn_hidden_states = self.attn1(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=None,
+            image_rotary_emb=image_rotary_emb,
+        )
+        hidden_states = hidden_states + attn_hidden_states * gate_msa
+
+        attn_hidden_states = self.attn2(
+            hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            image_rotary_emb=None,
+            attention_mask=encoder_attention_mask,
+        )
+        hidden_states = hidden_states + attn_hidden_states
+        norm_hidden_states = self.norm2(hidden_states) * (1 + scale_mlp) + shift_mlp
+
+        ff_output = self.ff(norm_hidden_states)
+        hidden_states = hidden_states + ff_output * gate_mlp
+
+        return hidden_states
+
+
+@maybe_allow_in_graph
+class LTXVideoTransformer3DModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+    r"""
+    A Transformer model for video-like data used in [LTX](https://huggingface.co/Lightricks/LTX-Video).
+
+    Args:
+        in_channels (`int`, defaults to `128`):
+            The number of channels in the input.
+        out_channels (`int`, defaults to `128`):
+            The number of channels in the output.
+        patch_size (`int`, defaults to `1`):
+            The size of the spatial patches to use in the patch embedding layer.
+        patch_size_t (`int`, defaults to `1`):
+            The size of the tmeporal patches to use in the patch embedding layer.
+        num_attention_heads (`int`, defaults to `32`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, defaults to `64`):
+            The number of channels in each head.
+        cross_attention_dim (`int`, defaults to `2048 `):
+            The number of channels for cross attention heads.
+        num_layers (`int`, defaults to `28`):
+            The number of layers of Transformer blocks to use.
+        activation_fn (`str`, defaults to `"gelu-approximate"`):
+            Activation function to use in feed-forward.
+        qk_norm (`str`, defaults to `"rms_norm_across_heads"`):
+            The normalization layer to use.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 128,
+        out_channels: int = 128,
+        patch_size: int = 1,
+        patch_size_t: int = 1,
+        num_attention_heads: int = 32,
+        attention_head_dim: int = 64,
+        cross_attention_dim: int = 2048,
+        num_layers: int = 28,
+        activation_fn: str = "gelu-approximate",
+        qk_norm: str = "rms_norm_across_heads",
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-6,
+        caption_channels: int = 4096,
+        attention_bias: bool = True,
+        attention_out_bias: bool = True,
+    ) -> None:
+        super().__init__()
+
+        out_channels = out_channels or in_channels
+        inner_dim = num_attention_heads * attention_head_dim
+
+        self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+        self.time_embed = AdaLayerNormSingle(inner_dim, use_additional_conditions=False)
+
+        self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+
+        self.rope = LTXRotaryPosEmbed(
+            dim=inner_dim,
+            base_num_frames=20,
+            base_height=2048,
+            base_width=2048,
+            patch_size=patch_size,
+            patch_size_t=patch_size_t,
+            theta=10000.0,
+        )
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                LTXTransformerBlock(
+                    dim=inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    cross_attention_dim=cross_attention_dim,
+                    qk_norm=qk_norm,
+                    activation_fn=activation_fn,
+                    attention_bias=attention_bias,
+                    attention_out_bias=attention_out_bias,
+                    eps=norm_eps,
+                    elementwise_affine=norm_elementwise_affine,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        self.norm_out = nn.LayerNorm(inner_dim, eps=1e-6, elementwise_affine=False)
+        self.proj_out = nn.Linear(inner_dim, out_channels)
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_attention_mask: torch.Tensor,
+        num_frames: int,
+        height: int,
+        width: int,
+        rope_interpolation_scale: Optional[Tuple[float, float, float]] = None,
+        return_dict: bool = True,
+    ) -> torch.Tensor:
+        image_rotary_emb = self.rope(hidden_states, num_frames, height, width, rope_interpolation_scale)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        batch_size = hidden_states.size(0)
+        hidden_states = self.proj_in(hidden_states)
+
+        temb, embedded_timestep = self.time_embed(
+            timestep.flatten(),
+            batch_size=batch_size,
+            hidden_dtype=hidden_states.dtype,
+        )
+
+        temb = temb.view(batch_size, -1, temb.size(-1))
+        embedded_timestep = embedded_timestep.view(batch_size, -1, embedded_timestep.size(-1))
+
+        encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+        encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.size(-1))
+
+        for block in self.transformer_blocks:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    encoder_hidden_states,
+                    temb,
+                    image_rotary_emb,
+                    encoder_attention_mask,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    temb=temb,
+                    image_rotary_emb=image_rotary_emb,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+
+        scale_shift_values = self.scale_shift_table[None, None] + embedded_timestep[:, :, None]
+        shift, scale = scale_shift_values[:, :, 0], scale_shift_values[:, :, 1]
+
+        hidden_states = self.norm_out(hidden_states)
+        hidden_states = hidden_states * (1 + scale) + shift
+        output = self.proj_out(hidden_states)
+
+        if not return_dict:
+            return (output,)
+        return Transformer2DModelOutput(sample=output)
+
+
+def apply_rotary_emb(x, freqs):
+    cos, sin = freqs
+    x_real, x_imag = x.unflatten(2, (-1, 2)).unbind(-1)  # [B, S, H, D // 2]
+    x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(2)
+    out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+    return out

src/diffusers/models/unets/unet_3d_blocks.py

@@ -1375,6 +1375,7 @@ class UpBlockSpatioTemporal(nn.Module):
         res_hidden_states_tuple: Tuple[torch.Tensor, ...],
         temb: Optional[torch.Tensor] = None,
         image_only_indicator: Optional[torch.Tensor] = None,
+        upsample_size: Optional[int] = None,
     ) -> torch.Tensor:
         for resnet in self.resnets:
             # pop res hidden states
@@ -1415,7 +1416,7 @@ class UpBlockSpatioTemporal(nn.Module):
 
         if self.upsamplers is not None:
             for upsampler in self.upsamplers:
-                hidden_states = upsampler(hidden_states)
+                hidden_states = upsampler(hidden_states, upsample_size)
 
         return hidden_states
 
@@ -1485,6 +1486,7 @@ class CrossAttnUpBlockSpatioTemporal(nn.Module):
         temb: Optional[torch.Tensor] = None,
         encoder_hidden_states: Optional[torch.Tensor] = None,
         image_only_indicator: Optional[torch.Tensor] = None,
+        upsample_size: Optional[int] = None,
     ) -> torch.Tensor:
         for resnet, attn in zip(self.resnets, self.attentions):
             # pop res hidden states
@@ -1533,6 +1535,6 @@ class CrossAttnUpBlockSpatioTemporal(nn.Module):
 
         if self.upsamplers is not None:
             for upsampler in self.upsamplers:
-                hidden_states = upsampler(hidden_states)
+                hidden_states = upsampler(hidden_states, upsample_size)
 
         return hidden_states

src/diffusers/models/unets/unet_spatio_temporal_condition.py

@@ -382,6 +382,20 @@ class UNetSpatioTemporalConditionModel(ModelMixin, ConfigMixin, UNet2DConditionL
                 If `return_dict` is True, an [`~models.unet_slatio_temporal.UNetSpatioTemporalConditionOutput`] is
                 returned, otherwise a `tuple` is returned where the first element is the sample tensor.
         """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
         # 1. time
         timesteps = timestep
         if not torch.is_tensor(timesteps):
@@ -457,15 +471,23 @@ class UNetSpatioTemporalConditionModel(ModelMixin, ConfigMixin, UNet2DConditionL
 
         # 5. up
         for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
             res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
             down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
 
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
             if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
                 sample = upsample_block(
                     hidden_states=sample,
                     temb=emb,
                     res_hidden_states_tuple=res_samples,
                     encoder_hidden_states=encoder_hidden_states,
+                    upsample_size=upsample_size,
                     image_only_indicator=image_only_indicator,
                 )
             else:
@@ -473,6 +495,7 @@ class UNetSpatioTemporalConditionModel(ModelMixin, ConfigMixin, UNet2DConditionL
                     hidden_states=sample,
                     temb=emb,
                     res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
                     image_only_indicator=image_only_indicator,
                 )
 

src/diffusers/pipelines/__init__.py

@@ -162,6 +162,9 @@ else:
             "StableDiffusionXLControlNetImg2ImgPipeline",
             "StableDiffusionXLControlNetInpaintPipeline",
             "StableDiffusionXLControlNetPipeline",
+            "StableDiffusionXLControlNetUnionPipeline",
+            "StableDiffusionXLControlNetUnionInpaintPipeline",
+            "StableDiffusionXLControlNetUnionImg2ImgPipeline",
         ]
     )
     _import_structure["pag"].extend(
@@ -182,6 +185,7 @@ else:
             "StableDiffusionXLControlNetPAGPipeline",
             "StableDiffusionXLPAGImg2ImgPipeline",
             "PixArtSigmaPAGPipeline",
+            "SanaPAGPipeline",
         ]
     )
     _import_structure["controlnet_xs"].extend(
@@ -247,6 +251,7 @@ else:
         ]
     )
     _import_structure["latte"] = ["LattePipeline"]
+    _import_structure["ltx"] = ["LTXPipeline", "LTXImageToVideoPipeline"]
     _import_structure["lumina"] = ["LuminaText2ImgPipeline"]
     _import_structure["marigold"].extend(
         [
@@ -259,6 +264,7 @@ else:
     _import_structure["paint_by_example"] = ["PaintByExamplePipeline"]
     _import_structure["pia"] = ["PIAPipeline"]
     _import_structure["pixart_alpha"] = ["PixArtAlphaPipeline", "PixArtSigmaPipeline"]
+    _import_structure["sana"] = ["SanaPipeline"]
     _import_structure["semantic_stable_diffusion"] = ["SemanticStableDiffusionPipeline"]
     _import_structure["shap_e"] = ["ShapEImg2ImgPipeline", "ShapEPipeline"]
     _import_structure["stable_audio"] = [
@@ -496,6 +502,9 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusionXLControlNetImg2ImgPipeline,
             StableDiffusionXLControlNetInpaintPipeline,
             StableDiffusionXLControlNetPipeline,
+            StableDiffusionXLControlNetUnionImg2ImgPipeline,
+            StableDiffusionXLControlNetUnionInpaintPipeline,
+            StableDiffusionXLControlNetUnionPipeline,
         )
         from .controlnet_hunyuandit import (
             HunyuanDiTControlNetPipeline,
@@ -579,6 +588,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LEditsPPPipelineStableDiffusion,
             LEditsPPPipelineStableDiffusionXL,
         )
+        from .ltx import LTXImageToVideoPipeline, LTXPipeline
         from .lumina import LuminaText2ImgPipeline
         from .marigold import (
             MarigoldDepthPipeline,
@@ -591,6 +601,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             HunyuanDiTPAGPipeline,
             KolorsPAGPipeline,
             PixArtSigmaPAGPipeline,
+            SanaPAGPipeline,
             StableDiffusion3PAGImg2ImgPipeline,
             StableDiffusion3PAGPipeline,
             StableDiffusionControlNetPAGInpaintPipeline,
@@ -607,6 +618,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .paint_by_example import PaintByExamplePipeline
         from .pia import PIAPipeline
         from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline
+        from .sana import SanaPipeline
         from .semantic_stable_diffusion import SemanticStableDiffusionPipeline
         from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline
         from .stable_audio import StableAudioPipeline, StableAudioProjectionModel

src/diffusers/pipelines/controlnet/__init__.py

@@ -1,80 +1,86 @@
-from typing import TYPE_CHECKING
-
-from ...utils import (
-    DIFFUSERS_SLOW_IMPORT,
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    get_objects_from_module,
-    is_flax_available,
-    is_torch_available,
-    is_transformers_available,
-)
-
-
-_dummy_objects = {}
-_import_structure = {}
-
-try:
-    if not (is_transformers_available() and is_torch_available()):
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
-
-    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
-else:
-    _import_structure["multicontrolnet"] = ["MultiControlNetModel"]
-    _import_structure["pipeline_controlnet"] = ["StableDiffusionControlNetPipeline"]
-    _import_structure["pipeline_controlnet_blip_diffusion"] = ["BlipDiffusionControlNetPipeline"]
-    _import_structure["pipeline_controlnet_img2img"] = ["StableDiffusionControlNetImg2ImgPipeline"]
-    _import_structure["pipeline_controlnet_inpaint"] = ["StableDiffusionControlNetInpaintPipeline"]
-    _import_structure["pipeline_controlnet_inpaint_sd_xl"] = ["StableDiffusionXLControlNetInpaintPipeline"]
-    _import_structure["pipeline_controlnet_sd_xl"] = ["StableDiffusionXLControlNetPipeline"]
-    _import_structure["pipeline_controlnet_sd_xl_img2img"] = ["StableDiffusionXLControlNetImg2ImgPipeline"]
-try:
-    if not (is_transformers_available() and is_flax_available()):
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    from ...utils import dummy_flax_and_transformers_objects  # noqa F403
-
-    _dummy_objects.update(get_objects_from_module(dummy_flax_and_transformers_objects))
-else:
-    _import_structure["pipeline_flax_controlnet"] = ["FlaxStableDiffusionControlNetPipeline"]
-
-
-if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
-    try:
-        if not (is_transformers_available() and is_torch_available()):
-            raise OptionalDependencyNotAvailable()
-
-    except OptionalDependencyNotAvailable:
-        from ...utils.dummy_torch_and_transformers_objects import *
-    else:
-        from .multicontrolnet import MultiControlNetModel
-        from .pipeline_controlnet import StableDiffusionControlNetPipeline
-        from .pipeline_controlnet_blip_diffusion import BlipDiffusionControlNetPipeline
-        from .pipeline_controlnet_img2img import StableDiffusionControlNetImg2ImgPipeline
-        from .pipeline_controlnet_inpaint import StableDiffusionControlNetInpaintPipeline
-        from .pipeline_controlnet_inpaint_sd_xl import StableDiffusionXLControlNetInpaintPipeline
-        from .pipeline_controlnet_sd_xl import StableDiffusionXLControlNetPipeline
-        from .pipeline_controlnet_sd_xl_img2img import StableDiffusionXLControlNetImg2ImgPipeline
-
-    try:
-        if not (is_transformers_available() and is_flax_available()):
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        from ...utils.dummy_flax_and_transformers_objects import *  # noqa F403
-    else:
-        from .pipeline_flax_controlnet import FlaxStableDiffusionControlNetPipeline
-
-
-else:
-    import sys
-
-    sys.modules[__name__] = _LazyModule(
-        __name__,
-        globals()["__file__"],
-        _import_structure,
-        module_spec=__spec__,
-    )
-    for name, value in _dummy_objects.items():
-        setattr(sys.modules[__name__], name, value)
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_flax_available,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["multicontrolnet"] = ["MultiControlNetModel"]
+    _import_structure["pipeline_controlnet"] = ["StableDiffusionControlNetPipeline"]
+    _import_structure["pipeline_controlnet_blip_diffusion"] = ["BlipDiffusionControlNetPipeline"]
+    _import_structure["pipeline_controlnet_img2img"] = ["StableDiffusionControlNetImg2ImgPipeline"]
+    _import_structure["pipeline_controlnet_inpaint"] = ["StableDiffusionControlNetInpaintPipeline"]
+    _import_structure["pipeline_controlnet_inpaint_sd_xl"] = ["StableDiffusionXLControlNetInpaintPipeline"]
+    _import_structure["pipeline_controlnet_sd_xl"] = ["StableDiffusionXLControlNetPipeline"]
+    _import_structure["pipeline_controlnet_sd_xl_img2img"] = ["StableDiffusionXLControlNetImg2ImgPipeline"]
+    _import_structure["pipeline_controlnet_union_inpaint_sd_xl"] = ["StableDiffusionXLControlNetUnionInpaintPipeline"]
+    _import_structure["pipeline_controlnet_union_sd_xl"] = ["StableDiffusionXLControlNetUnionPipeline"]
+    _import_structure["pipeline_controlnet_union_sd_xl_img2img"] = ["StableDiffusionXLControlNetUnionImg2ImgPipeline"]
+try:
+    if not (is_transformers_available() and is_flax_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_flax_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_flax_and_transformers_objects))
+else:
+    _import_structure["pipeline_flax_controlnet"] = ["FlaxStableDiffusionControlNetPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *
+    else:
+        from .multicontrolnet import MultiControlNetModel
+        from .pipeline_controlnet import StableDiffusionControlNetPipeline
+        from .pipeline_controlnet_blip_diffusion import BlipDiffusionControlNetPipeline
+        from .pipeline_controlnet_img2img import StableDiffusionControlNetImg2ImgPipeline
+        from .pipeline_controlnet_inpaint import StableDiffusionControlNetInpaintPipeline
+        from .pipeline_controlnet_inpaint_sd_xl import StableDiffusionXLControlNetInpaintPipeline
+        from .pipeline_controlnet_sd_xl import StableDiffusionXLControlNetPipeline
+        from .pipeline_controlnet_sd_xl_img2img import StableDiffusionXLControlNetImg2ImgPipeline
+        from .pipeline_controlnet_union_inpaint_sd_xl import StableDiffusionXLControlNetUnionInpaintPipeline
+        from .pipeline_controlnet_union_sd_xl import StableDiffusionXLControlNetUnionPipeline
+        from .pipeline_controlnet_union_sd_xl_img2img import StableDiffusionXLControlNetUnionImg2ImgPipeline
+
+    try:
+        if not (is_transformers_available() and is_flax_available()):
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_flax_and_transformers_objects import *  # noqa F403
+    else:
+        from .pipeline_flax_controlnet import FlaxStableDiffusionControlNetPipeline
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

src/diffusers/pipelines/ltx/__init__.py

@@ -0,0 +1,50 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_ltx"] = ["LTXPipeline"]
+    _import_structure["pipeline_ltx_image2video"] = ["LTXImageToVideoPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *
+    else:
+        from .pipeline_ltx import LTXPipeline
+        from .pipeline_ltx_image2video import LTXImageToVideoPipeline
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

src/diffusers/pipelines/ltx/pipeline_ltx.py

@@ -0,0 +1,755 @@
+# Copyright 2024 Black Forest Labs and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import T5EncoderModel, T5TokenizerFast
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...loaders import FromSingleFileMixin
+from ...models.autoencoders import AutoencoderKLLTXVideo
+from ...models.transformers import LTXVideoTransformer3DModel
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+from ...utils import is_torch_xla_available, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ...video_processor import VideoProcessor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import LTXPipelineOutput
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import LTXPipeline
+        >>> from diffusers.utils import export_to_video
+
+        >>> pipe = LTXPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+
+        >>> prompt = "A woman with long brown hair and light skin smiles at another woman with long blonde hair. The woman with brown hair wears a black jacket and has a small, barely noticeable mole on her right cheek. The camera angle is a close-up, focused on the woman with brown hair's face. The lighting is warm and natural, likely from the setting sun, casting a soft glow on the scene. The scene appears to be real-life footage"
+        >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
+
+        >>> video = pipe(
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     width=704,
+        ...     height=480,
+        ...     num_frames=161,
+        ...     num_inference_steps=50,
+        ... ).frames[0]
+        >>> export_to_video(video, "output.mp4", fps=24)
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.16,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class LTXPipeline(DiffusionPipeline, FromSingleFileMixin):
+    r"""
+    Pipeline for text-to-video generation.
+
+    Reference: https://github.com/Lightricks/LTX-Video
+
+    Args:
+        transformer ([`LTXVideoTransformer3DModel`]):
+            Conditional Transformer architecture to denoise the encoded video latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLLTXVideo`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer (`T5TokenizerFast`):
+            Second Tokenizer of class
+            [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKLLTXVideo,
+        text_encoder: T5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: LTXVideoTransformer3DModel,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+
+        self.vae_spatial_compression_ratio = self.vae.spatial_compression_ratio if hasattr(self, "vae") else 32
+        self.vae_temporal_compression_ratio = self.vae.temporal_compression_ratio if hasattr(self, "vae") else 8
+        self.transformer_spatial_patch_size = self.transformer.config.patch_size if hasattr(self, "transformer") else 1
+        self.transformer_temporal_patch_size = (
+            self.transformer.config.patch_size_t if hasattr(self, "transformer") else 1
+        )
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio)
+        self.tokenizer_max_length = (
+            self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 128
+        )
+
+    # Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline._get_t5_prompt_embeds with 256->128
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 128,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        prompt_attention_mask = text_inputs.attention_mask
+        prompt_attention_mask = prompt_attention_mask.bool().to(device)
+
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        prompt_attention_mask = prompt_attention_mask.view(batch_size, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1)
+
+        return prompt_embeds, prompt_attention_mask
+
+    # Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline.encode_prompt with 256->128
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 128,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        callback_on_step_end_tensor_inputs=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_attention_mask=None,
+        negative_prompt_attention_mask=None,
+    ):
+        if height % 32 != 0 or width % 32 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt_embeds is not None and prompt_attention_mask is None:
+            raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
+
+        if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
+            raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+            if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
+                raise ValueError(
+                    "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
+                    f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
+                    f" {negative_prompt_attention_mask.shape}."
+                )
+
+    @staticmethod
+    def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor:
+        # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p].
+        # The patch dimensions are then permuted and collapsed into the channel dimension of shape:
+        # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor).
+        # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features
+        batch_size, num_channels, num_frames, height, width = latents.shape
+        post_patch_num_frames = num_frames // patch_size_t
+        post_patch_height = height // patch_size
+        post_patch_width = width // patch_size
+        latents = latents.reshape(
+            batch_size,
+            -1,
+            post_patch_num_frames,
+            patch_size_t,
+            post_patch_height,
+            patch_size,
+            post_patch_width,
+            patch_size,
+        )
+        latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3)
+        return latents
+
+    @staticmethod
+    def _unpack_latents(
+        latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1
+    ) -> torch.Tensor:
+        # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions)
+        # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of
+        # what happens in the `_pack_latents` method.
+        batch_size = latents.size(0)
+        latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size)
+        latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3)
+        return latents
+
+    @staticmethod
+    def _normalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Normalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = (latents - latents_mean) * scaling_factor / latents_std
+        return latents
+
+    @staticmethod
+    def _denormalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Denormalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = latents * latents_std / scaling_factor + latents_mean
+        return latents
+
+    def prepare_latents(
+        self,
+        batch_size: int = 1,
+        num_channels_latents: int = 128,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        height = height // self.vae_spatial_compression_ratio
+        width = width // self.vae_spatial_compression_ratio
+        num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1
+
+        shape = (batch_size, num_channels_latents, num_frames, height, width)
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        latents = self._pack_latents(
+            latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+        )
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        frame_rate: int = 25,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        guidance_scale: float = 3,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 128,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, defaults to `512`):
+                The height in pixels of the generated image. This is set to 480 by default for the best results.
+            width (`int`, defaults to `704`):
+                The width in pixels of the generated image. This is set to 848 by default for the best results.
+            num_frames (`int`, defaults to `161`):
+                The number of video frames to generate
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, defaults to `3 `):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            prompt_attention_mask (`torch.Tensor`, *optional*):
+                Pre-generated attention mask for text embeddings.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
+                Pre-generated attention mask for negative text embeddings.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.ltx.LTXPipelineOutput`] instead of a plain tuple.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to `128 `):
+                Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.ltx.LTXPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] is returned, otherwise a `tuple` is
+                returned where the first element is a list with the generated images.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt=prompt,
+            height=height,
+            width=width,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Prepare text embeddings
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # 5. Prepare timesteps
+        latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1
+        latent_height = height // self.vae_spatial_compression_ratio
+        latent_width = width // self.vae_spatial_compression_ratio
+        video_sequence_length = latent_num_frames * latent_height * latent_width
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+        mu = calculate_shift(
+            video_sequence_length,
+            self.scheduler.config.base_image_seq_len,
+            self.scheduler.config.max_image_seq_len,
+            self.scheduler.config.base_shift,
+            self.scheduler.config.max_shift,
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas=sigmas,
+            mu=mu,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Prepare micro-conditions
+        latent_frame_rate = frame_rate / self.vae_temporal_compression_ratio
+        rope_interpolation_scale = (
+            1 / latent_frame_rate,
+            self.vae_spatial_compression_ratio,
+            self.vae_spatial_compression_ratio,
+        )
+
+        # 7. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+                latent_model_input = latent_model_input.to(prompt_embeds.dtype)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    encoder_attention_mask=prompt_attention_mask,
+                    num_frames=latent_num_frames,
+                    height=latent_height,
+                    width=latent_width,
+                    rope_interpolation_scale=rope_interpolation_scale,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if output_type == "latent":
+            video = latents
+        else:
+            latents = self._unpack_latents(
+                latents,
+                latent_num_frames,
+                latent_height,
+                latent_width,
+                self.transformer_spatial_patch_size,
+                self.transformer_temporal_patch_size,
+            )
+            latents = self._denormalize_latents(
+                latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
+            )
+            latents = latents.to(prompt_embeds.dtype)
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return LTXPipelineOutput(frames=video)

src/diffusers/pipelines/ltx/pipeline_ltx_image2video.py

@@ -0,0 +1,851 @@
+# Copyright 2024 Black Forest Labs and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import T5EncoderModel, T5TokenizerFast
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...image_processor import PipelineImageInput
+from ...loaders import FromSingleFileMixin
+from ...models.autoencoders import AutoencoderKLLTXVideo
+from ...models.transformers import LTXVideoTransformer3DModel
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+from ...utils import is_torch_xla_available, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ...video_processor import VideoProcessor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import LTXPipelineOutput
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import LTXImageToVideoPipeline
+        >>> from diffusers.utils import export_to_video, load_image
+
+        >>> pipe = LTXImageToVideoPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/a-r-r-o-w/tiny-meme-dataset-captioned/resolve/main/images/8.png"
+        ... )
+        >>> prompt = "A young girl stands calmly in the foreground, looking directly at the camera, as a house fire rages in the background. Flames engulf the structure, with smoke billowing into the air. Firefighters in protective gear rush to the scene, a fire truck labeled '38' visible behind them. The girl's neutral expression contrasts sharply with the chaos of the fire, creating a poignant and emotionally charged scene."
+        >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
+
+        >>> video = pipe(
+        ...     image=image,
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     width=704,
+        ...     height=480,
+        ...     num_frames=161,
+        ...     num_inference_steps=50,
+        ... ).frames[0]
+        >>> export_to_video(video, "output.mp4", fps=24)
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.16,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class LTXImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin):
+    r"""
+    Pipeline for image-to-video generation.
+
+    Reference: https://github.com/Lightricks/LTX-Video
+
+    Args:
+        transformer ([`LTXVideoTransformer3DModel`]):
+            Conditional Transformer architecture to denoise the encoded video latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLLTXVideo`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer (`T5TokenizerFast`):
+            Second Tokenizer of class
+            [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKLLTXVideo,
+        text_encoder: T5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: LTXVideoTransformer3DModel,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+
+        self.vae_spatial_compression_ratio = self.vae.spatial_compression_ratio if hasattr(self, "vae") else 32
+        self.vae_temporal_compression_ratio = self.vae.temporal_compression_ratio if hasattr(self, "vae") else 8
+        self.transformer_spatial_patch_size = self.transformer.config.patch_size if hasattr(self, "transformer") else 1
+        self.transformer_temporal_patch_size = (
+            self.transformer.config.patch_size_t if hasattr(self, "transformer") else 1
+        )
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio)
+        self.tokenizer_max_length = (
+            self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 128
+        )
+
+        self.default_height = 512
+        self.default_width = 704
+        self.default_frames = 121
+
+    # Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline._get_t5_prompt_embeds with 256->128
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 128,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        prompt_attention_mask = text_inputs.attention_mask
+        prompt_attention_mask = prompt_attention_mask.bool().to(device)
+
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        prompt_attention_mask = prompt_attention_mask.view(batch_size, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1)
+
+        return prompt_embeds, prompt_attention_mask
+
+    # Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline.encode_prompt with 256->128
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 128,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        callback_on_step_end_tensor_inputs=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_attention_mask=None,
+        negative_prompt_attention_mask=None,
+    ):
+        if height % 32 != 0 or width % 32 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt_embeds is not None and prompt_attention_mask is None:
+            raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
+
+        if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
+            raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+            if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
+                raise ValueError(
+                    "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
+                    f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
+                    f" {negative_prompt_attention_mask.shape}."
+                )
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._pack_latents
+    def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor:
+        # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p].
+        # The patch dimensions are then permuted and collapsed into the channel dimension of shape:
+        # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor).
+        # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features
+        batch_size, num_channels, num_frames, height, width = latents.shape
+        post_patch_num_frames = num_frames // patch_size_t
+        post_patch_height = height // patch_size
+        post_patch_width = width // patch_size
+        latents = latents.reshape(
+            batch_size,
+            -1,
+            post_patch_num_frames,
+            patch_size_t,
+            post_patch_height,
+            patch_size,
+            post_patch_width,
+            patch_size,
+        )
+        latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3)
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._unpack_latents
+    def _unpack_latents(
+        latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1
+    ) -> torch.Tensor:
+        # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions)
+        # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of
+        # what happens in the `_pack_latents` method.
+        batch_size = latents.size(0)
+        latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size)
+        latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3)
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._normalize_latents
+    def _normalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Normalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = (latents - latents_mean) * scaling_factor / latents_std
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._denormalize_latents
+    def _denormalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Denormalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = latents * latents_std / scaling_factor + latents_mean
+        return latents
+
+    def prepare_latents(
+        self,
+        image: Optional[torch.Tensor] = None,
+        batch_size: int = 1,
+        num_channels_latents: int = 128,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        height = height // self.vae_spatial_compression_ratio
+        width = width // self.vae_spatial_compression_ratio
+        num_frames = (
+            (num_frames - 1) // self.vae_temporal_compression_ratio + 1 if latents is None else latents.size(2)
+        )
+
+        shape = (batch_size, num_channels_latents, num_frames, height, width)
+        mask_shape = (batch_size, 1, num_frames, height, width)
+
+        if latents is not None:
+            conditioning_mask = latents.new_zeros(shape)
+            conditioning_mask[:, :, 0] = 1.0
+            conditioning_mask = self._pack_latents(
+                conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+            )
+            return latents.to(device=device, dtype=dtype), conditioning_mask
+
+        if isinstance(generator, list):
+            if len(generator) != batch_size:
+                raise ValueError(
+                    f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                    f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+                )
+
+            init_latents = [
+                retrieve_latents(self.vae.encode(image[i].unsqueeze(0).unsqueeze(2)), generator[i])
+                for i in range(batch_size)
+            ]
+        else:
+            init_latents = [
+                retrieve_latents(self.vae.encode(img.unsqueeze(0).unsqueeze(2)), generator) for img in image
+            ]
+
+        init_latents = torch.cat(init_latents, dim=0).to(dtype)
+        init_latents = self._normalize_latents(init_latents, self.vae.latents_mean, self.vae.latents_std)
+        init_latents = init_latents.repeat(1, 1, num_frames, 1, 1)
+        conditioning_mask = torch.zeros(mask_shape, device=device, dtype=dtype)
+        conditioning_mask[:, :, 0] = 1.0
+
+        noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        latents = init_latents * conditioning_mask + noise * (1 - conditioning_mask)
+
+        conditioning_mask = self._pack_latents(
+            conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+        ).squeeze(-1)
+        latents = self._pack_latents(
+            latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+        )
+
+        return latents, conditioning_mask
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        image: PipelineImageInput = None,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        frame_rate: int = 25,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        guidance_scale: float = 3,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 128,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            image (`PipelineImageInput`):
+                The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`.
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, defaults to `512`):
+                The height in pixels of the generated image. This is set to 480 by default for the best results.
+            width (`int`, defaults to `704`):
+                The width in pixels of the generated image. This is set to 848 by default for the best results.
+            num_frames (`int`, defaults to `161`):
+                The number of video frames to generate
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, defaults to `3 `):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            prompt_attention_mask (`torch.Tensor`, *optional*):
+                Pre-generated attention mask for text embeddings.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
+                Pre-generated attention mask for negative text embeddings.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.ltx.LTXPipelineOutput`] instead of a plain tuple.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to `128 `):
+                Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.ltx.LTXPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] is returned, otherwise a `tuple` is
+                returned where the first element is a list with the generated images.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt=prompt,
+            height=height,
+            width=width,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Prepare text embeddings
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
+
+        # 4. Prepare latent variables
+        if latents is None:
+            image = self.video_processor.preprocess(image, height=height, width=width)
+            image = image.to(device=device, dtype=prompt_embeds.dtype)
+
+        num_channels_latents = self.transformer.config.in_channels
+        latents, conditioning_mask = self.prepare_latents(
+            image,
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        if self.do_classifier_free_guidance:
+            conditioning_mask = torch.cat([conditioning_mask, conditioning_mask])
+
+        # 5. Prepare timesteps
+        latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1
+        latent_height = height // self.vae_spatial_compression_ratio
+        latent_width = width // self.vae_spatial_compression_ratio
+        video_sequence_length = latent_num_frames * latent_height * latent_width
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+        mu = calculate_shift(
+            video_sequence_length,
+            self.scheduler.config.base_image_seq_len,
+            self.scheduler.config.max_image_seq_len,
+            self.scheduler.config.base_shift,
+            self.scheduler.config.max_shift,
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas=sigmas,
+            mu=mu,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Prepare micro-conditions
+        latent_frame_rate = frame_rate / self.vae_temporal_compression_ratio
+        rope_interpolation_scale = (
+            1 / latent_frame_rate,
+            self.vae_spatial_compression_ratio,
+            self.vae_spatial_compression_ratio,
+        )
+
+        # 7. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+                latent_model_input = latent_model_input.to(prompt_embeds.dtype)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+                timestep = timestep.unsqueeze(-1) * (1 - conditioning_mask)
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    encoder_attention_mask=prompt_attention_mask,
+                    num_frames=latent_num_frames,
+                    height=latent_height,
+                    width=latent_width,
+                    rope_interpolation_scale=rope_interpolation_scale,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                    timestep, _ = timestep.chunk(2)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                noise_pred = self._unpack_latents(
+                    noise_pred,
+                    latent_num_frames,
+                    latent_height,
+                    latent_width,
+                    self.transformer_spatial_patch_size,
+                    self.transformer_temporal_patch_size,
+                )
+                latents = self._unpack_latents(
+                    latents,
+                    latent_num_frames,
+                    latent_height,
+                    latent_width,
+                    self.transformer_spatial_patch_size,
+                    self.transformer_temporal_patch_size,
+                )
+
+                noise_pred = noise_pred[:, :, 1:]
+                noise_latents = latents[:, :, 1:]
+                pred_latents = self.scheduler.step(noise_pred, t, noise_latents, return_dict=False)[0]
+
+                latents = torch.cat([latents[:, :, :1], pred_latents], dim=2)
+                latents = self._pack_latents(
+                    latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+                )
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if output_type == "latent":
+            video = latents
+        else:
+            latents = self._unpack_latents(
+                latents,
+                latent_num_frames,
+                latent_height,
+                latent_width,
+                self.transformer_spatial_patch_size,
+                self.transformer_temporal_patch_size,
+            )
+            latents = self._denormalize_latents(
+                latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
+            )
+            latents = latents.to(prompt_embeds.dtype)
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return LTXPipelineOutput(frames=video)

src/diffusers/pipelines/ltx/pipeline_output.py

@@ -0,0 +1,20 @@
+from dataclasses import dataclass
+
+import torch
+
+from diffusers.utils import BaseOutput
+
+
+@dataclass
+class LTXPipelineOutput(BaseOutput):
+    r"""
+    Output class for LTX pipelines.
+
+    Args:
+        frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
+            List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing
+            denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
+            `(batch_size, num_frames, channels, height, width)`.
+    """
+
+    frames: torch.Tensor

src/diffusers/pipelines/pag/__init__.py

@@ -29,6 +29,7 @@ else:
     _import_structure["pipeline_pag_hunyuandit"] = ["HunyuanDiTPAGPipeline"]
     _import_structure["pipeline_pag_kolors"] = ["KolorsPAGPipeline"]
     _import_structure["pipeline_pag_pixart_sigma"] = ["PixArtSigmaPAGPipeline"]
+    _import_structure["pipeline_pag_sana"] = ["SanaPAGPipeline"]
     _import_structure["pipeline_pag_sd"] = ["StableDiffusionPAGPipeline"]
     _import_structure["pipeline_pag_sd_3"] = ["StableDiffusion3PAGPipeline"]
     _import_structure["pipeline_pag_sd_3_img2img"] = ["StableDiffusion3PAGImg2ImgPipeline"]
@@ -55,6 +56,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .pipeline_pag_hunyuandit import HunyuanDiTPAGPipeline
         from .pipeline_pag_kolors import KolorsPAGPipeline
         from .pipeline_pag_pixart_sigma import PixArtSigmaPAGPipeline
+        from .pipeline_pag_sana import SanaPAGPipeline
         from .pipeline_pag_sd import StableDiffusionPAGPipeline
         from .pipeline_pag_sd_3 import StableDiffusion3PAGPipeline
         from .pipeline_pag_sd_3_img2img import StableDiffusion3PAGImg2ImgPipeline

src/diffusers/pipelines/pag/pipeline_pag_sana.py

@@ -0,0 +1,887 @@
+# Copyright 2024 PixArt-Sigma Authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+import inspect
+import re
+import urllib.parse as ul
+from typing import Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...image_processor import PixArtImageProcessor
+from ...models import AutoencoderDC, SanaTransformer2DModel
+from ...models.attention_processor import PAGCFGSanaLinearAttnProcessor2_0, PAGIdentitySanaLinearAttnProcessor2_0
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+from ...utils import (
+    BACKENDS_MAPPING,
+    is_bs4_available,
+    is_ftfy_available,
+    logging,
+    replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ..pixart_alpha.pipeline_pixart_alpha import (
+    ASPECT_RATIO_512_BIN,
+    ASPECT_RATIO_1024_BIN,
+)
+from ..pixart_alpha.pipeline_pixart_sigma import ASPECT_RATIO_2048_BIN
+from .pag_utils import PAGMixin
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+if is_bs4_available():
+    from bs4 import BeautifulSoup
+
+if is_ftfy_available():
+    import ftfy
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import SanaPAGPipeline
+
+        >>> pipe = SanaPAGPipeline.from_pretrained(
+        ...     "Efficient-Large-Model/Sana_1600M_1024px_diffusers",
+        ...     pag_applied_layers=["transformer_blocks.8"],
+        ...     torch_dtype=torch.float32,
+        ... )
+        >>> pipe.to("cuda")
+        >>> pipe.text_encoder.to(torch.bfloat16)
+        >>> pipe.transformer = pipe.transformer.to(torch.float16)
+
+        >>> image = pipe(prompt='a cyberpunk cat with a neon sign that says "Sana"')[0]
+        >>> image[0].save("output.png")
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class SanaPAGPipeline(DiffusionPipeline, PAGMixin):
+    r"""
+    Pipeline for text-to-image generation using [Sana](https://huggingface.co/papers/2410.10629). This pipeline
+    supports the use of [Perturbed Attention Guidance
+    (PAG)](https://huggingface.co/docs/diffusers/main/en/using-diffusers/pag).
+    """
+
+    # fmt: off
+    bad_punct_regex = re.compile(r"[" + "#®•©™&@·º½¾¿¡§~" + r"\)" + r"\(" + r"\]" + r"\[" + r"\}" + r"\{" + r"\|" + "\\" + r"\/" + r"\*" + r"]{1,}")
+    # fmt: on
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        tokenizer: AutoTokenizer,
+        text_encoder: AutoModelForCausalLM,
+        vae: AutoencoderDC,
+        transformer: SanaTransformer2DModel,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        pag_applied_layers: Union[str, List[str]] = "transformer_blocks.0",
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+        )
+
+        self.vae_scale_factor = 2 ** (len(self.vae.config.encoder_block_out_channels) - 1)
+        self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+        self.set_pag_applied_layers(
+            pag_applied_layers,
+            pag_attn_processors=(PAGCFGSanaLinearAttnProcessor2_0(), PAGIdentitySanaLinearAttnProcessor2_0()),
+        )
+
+    # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: str = "",
+        num_images_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        clean_caption: bool = False,
+        max_sequence_length: int = 300,
+        complex_human_instruction: Optional[List[str]] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds`
+                instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For
+                PixArt-Alpha, this should be "".
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                whether to use classifier free guidance or not
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                number of images that should be generated per prompt
+            device: (`torch.device`, *optional*):
+                torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. For Sana, it's should be the embeddings of the "" string.
+            clean_caption (`bool`, defaults to `False`):
+                If `True`, the function will preprocess and clean the provided caption before encoding.
+            max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt.
+            complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`):
+                If `complex_human_instruction` is not empty, the function will use the complex Human instruction for
+                the prompt.
+        """
+
+        if device is None:
+            device = self._execution_device
+
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        self.tokenizer.padding_side = "right"
+
+        # See Section 3.1. of the paper.
+        max_length = max_sequence_length
+        select_index = [0] + list(range(-max_length + 1, 0))
+
+        if prompt_embeds is None:
+            prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
+
+            # prepare complex human instruction
+            if not complex_human_instruction:
+                max_length_all = max_length
+            else:
+                chi_prompt = "\n".join(complex_human_instruction)
+                prompt = [chi_prompt + p for p in prompt]
+                num_chi_prompt_tokens = len(self.tokenizer.encode(chi_prompt))
+                max_length_all = num_chi_prompt_tokens + max_length - 2
+
+            text_inputs = self.tokenizer(
+                prompt,
+                padding="max_length",
+                max_length=max_length_all,
+                truncation=True,
+                add_special_tokens=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids
+
+            prompt_attention_mask = text_inputs.attention_mask
+            prompt_attention_mask = prompt_attention_mask.to(device)
+
+            prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask)
+            prompt_embeds = prompt_embeds[0][:, select_index]
+            prompt_attention_mask = prompt_attention_mask[:, select_index]
+
+        if self.transformer is not None:
+            dtype = self.transformer.dtype
+        elif self.text_encoder is not None:
+            dtype = self.text_encoder.dtype
+        else:
+            dtype = None
+
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+        prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            uncond_tokens = [negative_prompt] * batch_size if isinstance(negative_prompt, str) else negative_prompt
+            uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
+            max_length = prompt_embeds.shape[1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_attention_mask=True,
+                add_special_tokens=True,
+                return_tensors="pt",
+            )
+            negative_prompt_attention_mask = uncond_input.attention_mask
+            negative_prompt_attention_mask = negative_prompt_attention_mask.to(device)
+
+            negative_prompt_embeds = self.text_encoder(
+                uncond_input.input_ids.to(device), attention_mask=negative_prompt_attention_mask
+            )
+            negative_prompt_embeds = negative_prompt_embeds[0]
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1)
+            negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1)
+        else:
+            negative_prompt_embeds = None
+            negative_prompt_attention_mask = None
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        callback_on_step_end_tensor_inputs=None,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_attention_mask=None,
+        negative_prompt_attention_mask=None,
+    ):
+        if height % 32 != 0 or width % 32 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and prompt_attention_mask is None:
+            raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
+
+        if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
+            raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+            if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
+                raise ValueError(
+                    "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
+                    f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
+                    f" {negative_prompt_attention_mask.shape}."
+                )
+
+    # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline._text_preprocessing
+    def _text_preprocessing(self, text, clean_caption=False):
+        if clean_caption and not is_bs4_available():
+            logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
+            logger.warning("Setting `clean_caption` to False...")
+            clean_caption = False
+
+        if clean_caption and not is_ftfy_available():
+            logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
+            logger.warning("Setting `clean_caption` to False...")
+            clean_caption = False
+
+        if not isinstance(text, (tuple, list)):
+            text = [text]
+
+        def process(text: str):
+            if clean_caption:
+                text = self._clean_caption(text)
+                text = self._clean_caption(text)
+            else:
+                text = text.lower().strip()
+            return text
+
+        return [process(t) for t in text]
+
+    # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline._clean_caption
+    def _clean_caption(self, caption):
+        caption = str(caption)
+        caption = ul.unquote_plus(caption)
+        caption = caption.strip().lower()
+        caption = re.sub("<person>", "person", caption)
+        # urls:
+        caption = re.sub(
+            r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))",  # noqa
+            "",
+            caption,
+        )  # regex for urls
+        caption = re.sub(
+            r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))",  # noqa
+            "",
+            caption,
+        )  # regex for urls
+        # html:
+        caption = BeautifulSoup(caption, features="html.parser").text
+
+        # @<nickname>
+        caption = re.sub(r"@[\w\d]+\b", "", caption)
+
+        # 31C0—31EF CJK Strokes
+        # 31F0—31FF Katakana Phonetic Extensions
+        # 3200—32FF Enclosed CJK Letters and Months
+        # 3300—33FF CJK Compatibility
+        # 3400—4DBF CJK Unified Ideographs Extension A
+        # 4DC0—4DFF Yijing Hexagram Symbols
+        # 4E00—9FFF CJK Unified Ideographs
+        caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
+        caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
+        caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
+        caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
+        caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
+        caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
+        caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
+        #######################################################
+
+        # все виды тире / all types of dash --> "-"
+        caption = re.sub(
+            r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+",  # noqa
+            "-",
+            caption,
+        )
+
+        # кавычки к одному стандарту
+        caption = re.sub(r"[`´«»“”¨]", '"', caption)
+        caption = re.sub(r"[‘’]", "'", caption)
+
+        # &quot;
+        caption = re.sub(r"&quot;?", "", caption)
+        # &amp
+        caption = re.sub(r"&amp", "", caption)
+
+        # ip adresses:
+        caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
+
+        # article ids:
+        caption = re.sub(r"\d:\d\d\s+$", "", caption)
+
+        # \n
+        caption = re.sub(r"\\n", " ", caption)
+
+        # "#123"
+        caption = re.sub(r"#\d{1,3}\b", "", caption)
+        # "#12345.."
+        caption = re.sub(r"#\d{5,}\b", "", caption)
+        # "123456.."
+        caption = re.sub(r"\b\d{6,}\b", "", caption)
+        # filenames:
+        caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
+
+        #
+        caption = re.sub(r"[\"\']{2,}", r'"', caption)  # """AUSVERKAUFT"""
+        caption = re.sub(r"[\.]{2,}", r" ", caption)  # """AUSVERKAUFT"""
+
+        caption = re.sub(self.bad_punct_regex, r" ", caption)  # ***AUSVERKAUFT***, #AUSVERKAUFT
+        caption = re.sub(r"\s+\.\s+", r" ", caption)  # " . "
+
+        # this-is-my-cute-cat / this_is_my_cute_cat
+        regex2 = re.compile(r"(?:\-|\_)")
+        if len(re.findall(regex2, caption)) > 3:
+            caption = re.sub(regex2, " ", caption)
+
+        caption = ftfy.fix_text(caption)
+        caption = html.unescape(html.unescape(caption))
+
+        caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption)  # jc6640
+        caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption)  # jc6640vc
+        caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption)  # 6640vc231
+
+        caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
+        caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
+        caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
+        caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
+        caption = re.sub(r"\bpage\s+\d+\b", "", caption)
+
+        caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption)  # j2d1a2a...
+
+        caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
+
+        caption = re.sub(r"\b\s+\:\s+", r": ", caption)
+        caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
+        caption = re.sub(r"\s+", " ", caption)
+
+        caption.strip()
+
+        caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
+        caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
+        caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
+        caption = re.sub(r"^\.\S+$", "", caption)
+
+        return caption.strip()
+
+    # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        shape = (
+            batch_size,
+            num_channels_latents,
+            int(height) // self.vae_scale_factor,
+            int(width) // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: str = "",
+        num_inference_steps: int = 20,
+        timesteps: List[int] = None,
+        sigmas: List[float] = None,
+        guidance_scale: float = 4.5,
+        num_images_per_prompt: Optional[int] = 1,
+        height: int = 1024,
+        width: int = 1024,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        clean_caption: bool = True,
+        use_resolution_binning: bool = True,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 300,
+        complex_human_instruction: List[str] = [
+            "Given a user prompt, generate an 'Enhanced prompt' that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:",
+            "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.",
+            "- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.",
+            "Here are examples of how to transform or refine prompts:",
+            "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.",
+            "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.",
+            "Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:",
+            "User Prompt: ",
+        ],
+        pag_scale: float = 3.0,
+        pag_adaptive_scale: float = 0.0,
+    ) -> Union[ImagePipelineOutput, Tuple]:
+        """
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            num_inference_steps (`int`, *optional*, defaults to 20):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to 4.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size):
+                The width in pixels of the generated image.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            negative_prompt_attention_mask (`torch.Tensor`, *optional*):
+                Pre-generated attention mask for negative text embeddings.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+            clean_caption (`bool`, *optional*, defaults to `True`):
+                Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
+                be installed. If the dependencies are not installed, the embeddings will be created from the raw
+                prompt.
+            use_resolution_binning (`bool` defaults to `True`):
+                If set to `True`, the requested height and width are first mapped to the closest resolutions using
+                `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to
+                the requested resolution. Useful for generating non-square images.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to 300): Maximum sequence length to use with the `prompt`.
+            complex_human_instruction (`List[str]`, *optional*):
+                Instructions for complex human attention:
+                https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55.
+            pag_scale (`float`, *optional*, defaults to 3.0):
+                The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention
+                guidance will not be used.
+            pag_adaptive_scale (`float`, *optional*, defaults to 0.0):
+                The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is
+                used.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.ImagePipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+                returned where the first element is a list with the generated images
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        if use_resolution_binning:
+            if self.transformer.config.sample_size == 64:
+                aspect_ratio_bin = ASPECT_RATIO_2048_BIN
+            elif self.transformer.config.sample_size == 32:
+                aspect_ratio_bin = ASPECT_RATIO_1024_BIN
+            elif self.transformer.config.sample_size == 16:
+                aspect_ratio_bin = ASPECT_RATIO_512_BIN
+            else:
+                raise ValueError("Invalid sample size")
+            orig_height, orig_width = height, width
+            height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin)
+
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            callback_on_step_end_tensor_inputs,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_attention_mask,
+        )
+
+        self._pag_scale = pag_scale
+        self._pag_adaptive_scale = pag_adaptive_scale
+        self._guidance_scale = guidance_scale
+        self._interrupt = False
+
+        # 2. Default height and width to transformer
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Encode input prompt
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt,
+            self.do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            device=device,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            clean_caption=clean_caption,
+            max_sequence_length=max_sequence_length,
+            complex_human_instruction=complex_human_instruction,
+        )
+
+        if self.do_perturbed_attention_guidance:
+            prompt_embeds = self._prepare_perturbed_attention_guidance(
+                prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance
+            )
+            prompt_attention_mask = self._prepare_perturbed_attention_guidance(
+                prompt_attention_mask, negative_prompt_attention_mask, self.do_classifier_free_guidance
+            )
+        elif self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler, num_inference_steps, device, timesteps, sigmas
+        )
+
+        # 5. Prepare latents.
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            latent_channels,
+            height,
+            width,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+        if self.do_perturbed_attention_guidance:
+            original_attn_proc = self.transformer.attn_processors
+            self._set_pag_attn_processor(
+                pag_applied_layers=self.pag_applied_layers,
+                do_classifier_free_guidance=self.do_classifier_free_guidance,
+            )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # expand the latents if we are doing classifier free guidance, perturbed-attention guidance, or both
+                latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0]))
+                latent_model_input = latent_model_input.to(prompt_embeds.dtype)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0]).to(latents.dtype)
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    encoder_attention_mask=prompt_attention_mask,
+                    timestep=timestep,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                # perform guidance
+                if self.do_perturbed_attention_guidance:
+                    noise_pred = self._apply_perturbed_attention_guidance(
+                        noise_pred, self.do_classifier_free_guidance, guidance_scale, timestep
+                    )
+                elif self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute previous image: x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                # call the callback, if provided
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+        if output_type == "latent":
+            image = latents
+        else:
+            latents = latents.to(self.vae.dtype)
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            if use_resolution_binning:
+                image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height)
+
+        if not output_type == "latent":
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if self.do_perturbed_attention_guidance:
+            self.transformer.set_attn_processor(original_attn_proc)
+
+        if not return_dict:
+            return (image,)
+
+        return ImagePipelineOutput(images=image)

src/diffusers/pipelines/pag/pipeline_pag_sd_3_img2img.py

@@ -549,6 +549,8 @@ class StableDiffusion3PAGImg2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin,
         prompt,
         prompt_2,
         prompt_3,
+        height,
+        width,
         strength,
         negative_prompt=None,
         negative_prompt_2=None,
@@ -560,6 +562,15 @@ class StableDiffusion3PAGImg2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin,
         callback_on_step_end_tensor_inputs=None,
         max_sequence_length=None,
     ):
+        if (
+            height % (self.vae_scale_factor * self.patch_size) != 0
+            or width % (self.vae_scale_factor * self.patch_size) != 0
+        ):
+            raise ValueError(
+                f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}."
+                f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}."
+            )
+
         if strength < 0 or strength > 1:
             raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
 
@@ -730,6 +741,8 @@ class StableDiffusion3PAGImg2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin,
         prompt: Union[str, List[str]] = None,
         prompt_2: Optional[Union[str, List[str]]] = None,
         prompt_3: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
         image: PipelineImageInput = None,
         strength: float = 0.6,
         num_inference_steps: int = 50,
@@ -860,11 +873,15 @@ class StableDiffusion3PAGImg2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin,
             [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a
             `tuple`. When returning a tuple, the first element is a list with the generated images.
         """
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(
             prompt,
             prompt_2,
             prompt_3,
+            height,
+            width,
             strength,
             negative_prompt=negative_prompt,
             negative_prompt_2=negative_prompt_2,
@@ -933,7 +950,7 @@ class StableDiffusion3PAGImg2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin,
             pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0)
 
         # 3. Preprocess image
-        image = self.image_processor.preprocess(image)
+        image = self.image_processor.preprocess(image, height=height, width=width)
 
         # 4. Prepare timesteps
         timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, sigmas=sigmas)

src/diffusers/pipelines/sana/__init__.py

@@ -0,0 +1,47 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_sana"] = ["SanaPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *
+    else:
+        from .pipeline_sana import SanaPipeline
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

src/diffusers/pipelines/sana/pipeline_output.py

@@ -0,0 +1,21 @@
+from dataclasses import dataclass
+from typing import List, Union
+
+import numpy as np
+import PIL.Image
+
+from ...utils import BaseOutput
+
+
+@dataclass
+class SanaPipelineOutput(BaseOutput):
+    """
+    Output class for Sana pipelines.
+
+    Args:
+        images (`List[PIL.Image.Image]` or `np.ndarray`)
+            List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+            num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+    """
+
+    images: Union[List[PIL.Image.Image], np.ndarray]

src/diffusers/pipelines/sana/pipeline_sana.py

@@ -0,0 +1,852 @@
+# Copyright 2024 PixArt-Sigma Authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+import inspect
+import re
+import urllib.parse as ul
+from typing import Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...image_processor import PixArtImageProcessor
+from ...models import AutoencoderDC, SanaTransformer2DModel
+from ...schedulers import DPMSolverMultistepScheduler
+from ...utils import (
+    BACKENDS_MAPPING,
+    is_bs4_available,
+    is_ftfy_available,
+    logging,
+    replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from ..pixart_alpha.pipeline_pixart_alpha import (
+    ASPECT_RATIO_512_BIN,
+    ASPECT_RATIO_1024_BIN,
+)
+from ..pixart_alpha.pipeline_pixart_sigma import ASPECT_RATIO_2048_BIN
+from .pipeline_output import SanaPipelineOutput
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+if is_bs4_available():
+    from bs4 import BeautifulSoup
+
+if is_ftfy_available():
+    import ftfy
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import SanaPipeline
+
+        >>> pipe = SanaPipeline.from_pretrained(
+        ...     "Efficient-Large-Model/Sana_1600M_1024px_diffusers", torch_dtype=torch.float32
+        ... )
+        >>> pipe.to("cuda")
+        >>> pipe.text_encoder.to(torch.bfloat16)
+        >>> pipe.transformer = pipe.transformer.to(torch.float16)
+
+        >>> image = pipe(prompt='a cyberpunk cat with a neon sign that says "Sana"')[0]
+        >>> image[0].save("output.png")
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class SanaPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using [Sana](https://huggingface.co/papers/2410.10629).
+    """
+
+    # fmt: off
+    bad_punct_regex = re.compile(r"[" + "#®•©™&@·º½¾¿¡§~" + r"\)" + r"\(" + r"\]" + r"\[" + r"\}" + r"\{" + r"\|" + "\\" + r"\/" + r"\*" + r"]{1,}")
+    # fmt: on
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        tokenizer: AutoTokenizer,
+        text_encoder: AutoModelForCausalLM,
+        vae: AutoencoderDC,
+        transformer: SanaTransformer2DModel,
+        scheduler: DPMSolverMultistepScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+        )
+
+        self.vae_scale_factor = (
+            2 ** (len(self.vae.config.encoder_block_out_channels) - 1)
+            if hasattr(self, "vae") and self.vae is not None
+            else 32
+        )
+        self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: str = "",
+        num_images_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        clean_caption: bool = False,
+        max_sequence_length: int = 300,
+        complex_human_instruction: Optional[List[str]] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds`
+                instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For
+                PixArt-Alpha, this should be "".
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                whether to use classifier free guidance or not
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                number of images that should be generated per prompt
+            device: (`torch.device`, *optional*):
+                torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. For Sana, it's should be the embeddings of the "" string.
+            clean_caption (`bool`, defaults to `False`):
+                If `True`, the function will preprocess and clean the provided caption before encoding.
+            max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt.
+            complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`):
+                If `complex_human_instruction` is not empty, the function will use the complex Human instruction for
+                the prompt.
+        """
+
+        if device is None:
+            device = self._execution_device
+
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        self.tokenizer.padding_side = "right"
+
+        # See Section 3.1. of the paper.
+        max_length = max_sequence_length
+        select_index = [0] + list(range(-max_length + 1, 0))
+
+        if prompt_embeds is None:
+            prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
+
+            # prepare complex human instruction
+            if not complex_human_instruction:
+                max_length_all = max_length
+            else:
+                chi_prompt = "\n".join(complex_human_instruction)
+                prompt = [chi_prompt + p for p in prompt]
+                num_chi_prompt_tokens = len(self.tokenizer.encode(chi_prompt))
+                max_length_all = num_chi_prompt_tokens + max_length - 2
+
+            text_inputs = self.tokenizer(
+                prompt,
+                padding="max_length",
+                max_length=max_length_all,
+                truncation=True,
+                add_special_tokens=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids
+
+            prompt_attention_mask = text_inputs.attention_mask
+            prompt_attention_mask = prompt_attention_mask.to(device)
+
+            prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask)
+            prompt_embeds = prompt_embeds[0][:, select_index]
+            prompt_attention_mask = prompt_attention_mask[:, select_index]
+
+        if self.transformer is not None:
+            dtype = self.transformer.dtype
+        elif self.text_encoder is not None:
+            dtype = self.text_encoder.dtype
+        else:
+            dtype = None
+
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+        prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            uncond_tokens = [negative_prompt] * batch_size if isinstance(negative_prompt, str) else negative_prompt
+            uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
+            max_length = prompt_embeds.shape[1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_attention_mask=True,
+                add_special_tokens=True,
+                return_tensors="pt",
+            )
+            negative_prompt_attention_mask = uncond_input.attention_mask
+            negative_prompt_attention_mask = negative_prompt_attention_mask.to(device)
+
+            negative_prompt_embeds = self.text_encoder(
+                uncond_input.input_ids.to(device), attention_mask=negative_prompt_attention_mask
+            )
+            negative_prompt_embeds = negative_prompt_embeds[0]
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1)
+            negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1)
+        else:
+            negative_prompt_embeds = None
+            negative_prompt_attention_mask = None
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        callback_on_step_end_tensor_inputs=None,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_attention_mask=None,
+        negative_prompt_attention_mask=None,
+    ):
+        if height % 32 != 0 or width % 32 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and prompt_attention_mask is None:
+            raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
+
+        if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
+            raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+            if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
+                raise ValueError(
+                    "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
+                    f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
+                    f" {negative_prompt_attention_mask.shape}."
+                )
+
+    # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
+    def _text_preprocessing(self, text, clean_caption=False):
+        if clean_caption and not is_bs4_available():
+            logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
+            logger.warning("Setting `clean_caption` to False...")
+            clean_caption = False
+
+        if clean_caption and not is_ftfy_available():
+            logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
+            logger.warning("Setting `clean_caption` to False...")
+            clean_caption = False
+
+        if not isinstance(text, (tuple, list)):
+            text = [text]
+
+        def process(text: str):
+            if clean_caption:
+                text = self._clean_caption(text)
+                text = self._clean_caption(text)
+            else:
+                text = text.lower().strip()
+            return text
+
+        return [process(t) for t in text]
+
+    # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption
+    def _clean_caption(self, caption):
+        caption = str(caption)
+        caption = ul.unquote_plus(caption)
+        caption = caption.strip().lower()
+        caption = re.sub("<person>", "person", caption)
+        # urls:
+        caption = re.sub(
+            r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))",  # noqa
+            "",
+            caption,
+        )  # regex for urls
+        caption = re.sub(
+            r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))",  # noqa
+            "",
+            caption,
+        )  # regex for urls
+        # html:
+        caption = BeautifulSoup(caption, features="html.parser").text
+
+        # @<nickname>
+        caption = re.sub(r"@[\w\d]+\b", "", caption)
+
+        # 31C0—31EF CJK Strokes
+        # 31F0—31FF Katakana Phonetic Extensions
+        # 3200—32FF Enclosed CJK Letters and Months
+        # 3300—33FF CJK Compatibility
+        # 3400—4DBF CJK Unified Ideographs Extension A
+        # 4DC0—4DFF Yijing Hexagram Symbols
+        # 4E00—9FFF CJK Unified Ideographs
+        caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
+        caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
+        caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
+        caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
+        caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
+        caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
+        caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
+        #######################################################
+
+        # все виды тире / all types of dash --> "-"
+        caption = re.sub(
+            r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+",  # noqa
+            "-",
+            caption,
+        )
+
+        # кавычки к одному стандарту
+        caption = re.sub(r"[`´«»“”¨]", '"', caption)
+        caption = re.sub(r"[‘’]", "'", caption)
+
+        # &quot;
+        caption = re.sub(r"&quot;?", "", caption)
+        # &amp
+        caption = re.sub(r"&amp", "", caption)
+
+        # ip adresses:
+        caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
+
+        # article ids:
+        caption = re.sub(r"\d:\d\d\s+$", "", caption)
+
+        # \n
+        caption = re.sub(r"\\n", " ", caption)
+
+        # "#123"
+        caption = re.sub(r"#\d{1,3}\b", "", caption)
+        # "#12345.."
+        caption = re.sub(r"#\d{5,}\b", "", caption)
+        # "123456.."
+        caption = re.sub(r"\b\d{6,}\b", "", caption)
+        # filenames:
+        caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
+
+        #
+        caption = re.sub(r"[\"\']{2,}", r'"', caption)  # """AUSVERKAUFT"""
+        caption = re.sub(r"[\.]{2,}", r" ", caption)  # """AUSVERKAUFT"""
+
+        caption = re.sub(self.bad_punct_regex, r" ", caption)  # ***AUSVERKAUFT***, #AUSVERKAUFT
+        caption = re.sub(r"\s+\.\s+", r" ", caption)  # " . "
+
+        # this-is-my-cute-cat / this_is_my_cute_cat
+        regex2 = re.compile(r"(?:\-|\_)")
+        if len(re.findall(regex2, caption)) > 3:
+            caption = re.sub(regex2, " ", caption)
+
+        caption = ftfy.fix_text(caption)
+        caption = html.unescape(html.unescape(caption))
+
+        caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption)  # jc6640
+        caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption)  # jc6640vc
+        caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption)  # 6640vc231
+
+        caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
+        caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
+        caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
+        caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
+        caption = re.sub(r"\bpage\s+\d+\b", "", caption)
+
+        caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption)  # j2d1a2a...
+
+        caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
+
+        caption = re.sub(r"\b\s+\:\s+", r": ", caption)
+        caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
+        caption = re.sub(r"\s+", " ", caption)
+
+        caption.strip()
+
+        caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
+        caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
+        caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
+        caption = re.sub(r"^\.\S+$", "", caption)
+
+        return caption.strip()
+
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        shape = (
+            batch_size,
+            num_channels_latents,
+            int(height) // self.vae_scale_factor,
+            int(width) // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: str = "",
+        num_inference_steps: int = 20,
+        timesteps: List[int] = None,
+        sigmas: List[float] = None,
+        guidance_scale: float = 4.5,
+        num_images_per_prompt: Optional[int] = 1,
+        height: int = 1024,
+        width: int = 1024,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        clean_caption: bool = True,
+        use_resolution_binning: bool = True,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 300,
+        complex_human_instruction: List[str] = [
+            "Given a user prompt, generate an 'Enhanced prompt' that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:",
+            "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.",
+            "- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.",
+            "Here are examples of how to transform or refine prompts:",
+            "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.",
+            "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.",
+            "Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:",
+            "User Prompt: ",
+        ],
+    ) -> Union[SanaPipelineOutput, Tuple]:
+        """
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            num_inference_steps (`int`, *optional*, defaults to 20):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to 4.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size):
+                The width in pixels of the generated image.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            negative_prompt_attention_mask (`torch.Tensor`, *optional*):
+                Pre-generated attention mask for negative text embeddings.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+            clean_caption (`bool`, *optional*, defaults to `True`):
+                Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
+                be installed. If the dependencies are not installed, the embeddings will be created from the raw
+                prompt.
+            use_resolution_binning (`bool` defaults to `True`):
+                If set to `True`, the requested height and width are first mapped to the closest resolutions using
+                `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to
+                the requested resolution. Useful for generating non-square images.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to `300`):
+                Maximum sequence length to use with the `prompt`.
+            complex_human_instruction (`List[str]`, *optional*):
+                Instructions for complex human attention:
+                https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is a list with the generated images
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        if use_resolution_binning:
+            if self.transformer.config.sample_size == 64:
+                aspect_ratio_bin = ASPECT_RATIO_2048_BIN
+            elif self.transformer.config.sample_size == 32:
+                aspect_ratio_bin = ASPECT_RATIO_1024_BIN
+            elif self.transformer.config.sample_size == 16:
+                aspect_ratio_bin = ASPECT_RATIO_512_BIN
+            else:
+                raise ValueError("Invalid sample size")
+            orig_height, orig_width = height, width
+            height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin)
+
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            callback_on_step_end_tensor_inputs,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_attention_mask,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._interrupt = False
+
+        # 2. Default height and width to transformer
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Encode input prompt
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt,
+            self.do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            device=device,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            clean_caption=clean_caption,
+            max_sequence_length=max_sequence_length,
+            complex_human_instruction=complex_human_instruction,
+        )
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler, num_inference_steps, device, timesteps, sigmas
+        )
+
+        # 5. Prepare latents.
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            latent_channels,
+            height,
+            width,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+                latent_model_input = latent_model_input.to(prompt_embeds.dtype)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0]).to(latents.dtype)
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    encoder_attention_mask=prompt_attention_mask,
+                    timestep=timestep,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # learned sigma
+                if self.transformer.config.out_channels // 2 == latent_channels:
+                    noise_pred = noise_pred.chunk(2, dim=1)[0]
+                else:
+                    noise_pred = noise_pred
+
+                # compute previous image: x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+        if output_type == "latent":
+            image = latents
+        else:
+            latents = latents.to(self.vae.dtype)
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            if use_resolution_binning:
+                image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height)
+
+        if not output_type == "latent":
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return SanaPipelineOutput(images=image)

src/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_img2img.py

@@ -218,6 +218,9 @@ class StableDiffusion3Img2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin, Fro
         )
         self.tokenizer_max_length = self.tokenizer.model_max_length
         self.default_sample_size = self.transformer.config.sample_size
+        self.patch_size = (
+            self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2
+        )
 
     # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_t5_prompt_embeds
     def _get_t5_prompt_embeds(
@@ -531,6 +534,8 @@ class StableDiffusion3Img2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin, Fro
         prompt,
         prompt_2,
         prompt_3,
+        height,
+        width,
         strength,
         negative_prompt=None,
         negative_prompt_2=None,
@@ -542,6 +547,15 @@ class StableDiffusion3Img2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin, Fro
         callback_on_step_end_tensor_inputs=None,
         max_sequence_length=None,
     ):
+        if (
+            height % (self.vae_scale_factor * self.patch_size) != 0
+            or width % (self.vae_scale_factor * self.patch_size) != 0
+        ):
+            raise ValueError(
+                f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}."
+                f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}."
+            )
+
         if strength < 0 or strength > 1:
             raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
 
@@ -710,6 +724,8 @@ class StableDiffusion3Img2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin, Fro
         prompt: Union[str, List[str]] = None,
         prompt_2: Optional[Union[str, List[str]]] = None,
         prompt_3: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
         image: PipelineImageInput = None,
         strength: float = 0.6,
         num_inference_steps: int = 50,
@@ -824,12 +840,16 @@ class StableDiffusion3Img2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin, Fro
             [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a
             `tuple`. When returning a tuple, the first element is a list with the generated images.
         """
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
 
         # 1. Check inputs. Raise error if not correct
         self.check_inputs(
             prompt,
             prompt_2,
             prompt_3,
+            height,
+            width,
             strength,
             negative_prompt=negative_prompt,
             negative_prompt_2=negative_prompt_2,
@@ -890,7 +910,7 @@ class StableDiffusion3Img2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin, Fro
             pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0)
 
         # 3. Preprocess image
-        image = self.image_processor.preprocess(image)
+        image = self.image_processor.preprocess(image, height=height, width=width)
 
         # 4. Prepare timesteps
         timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, sigmas=sigmas)

src/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_inpaint.py

@@ -224,6 +224,9 @@ class StableDiffusion3InpaintPipeline(DiffusionPipeline, SD3LoraLoaderMixin, Fro
         )
         self.tokenizer_max_length = self.tokenizer.model_max_length
         self.default_sample_size = self.transformer.config.sample_size
+        self.patch_size = (
+            self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2
+        )
 
     # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_t5_prompt_embeds
     def _get_t5_prompt_embeds(
@@ -538,6 +541,8 @@ class StableDiffusion3InpaintPipeline(DiffusionPipeline, SD3LoraLoaderMixin, Fro
         prompt,
         prompt_2,
         prompt_3,
+        height,
+        width,
         strength,
         negative_prompt=None,
         negative_prompt_2=None,
@@ -549,6 +554,15 @@ class StableDiffusion3InpaintPipeline(DiffusionPipeline, SD3LoraLoaderMixin, Fro
         callback_on_step_end_tensor_inputs=None,
         max_sequence_length=None,
     ):
+        if (
+            height % (self.vae_scale_factor * self.patch_size) != 0
+            or width % (self.vae_scale_factor * self.patch_size) != 0
+        ):
+            raise ValueError(
+                f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}."
+                f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}."
+            )
+
         if strength < 0 or strength > 1:
             raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
 
@@ -953,6 +967,8 @@ class StableDiffusion3InpaintPipeline(DiffusionPipeline, SD3LoraLoaderMixin, Fro
             prompt,
             prompt_2,
             prompt_3,
+            height,
+            width,
             strength,
             negative_prompt=negative_prompt,
             negative_prompt_2=negative_prompt_2,

src/diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py

@@ -446,13 +446,14 @@ class StableDiffusionGLIGENTextImagePipeline(DiffusionPipeline, StableDiffusionM
             extra_step_kwargs["generator"] = generator
         return extra_step_kwargs
 
-    # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.check_inputs
     def check_inputs(
         self,
         prompt,
         height,
         width,
         callback_steps,
+        gligen_images,
+        gligen_phrases,
         negative_prompt=None,
         prompt_embeds=None,
         negative_prompt_embeds=None,
@@ -499,6 +500,13 @@ class StableDiffusionGLIGENTextImagePipeline(DiffusionPipeline, StableDiffusionM
                     f" {negative_prompt_embeds.shape}."
                 )
 
+        if gligen_images is not None and gligen_phrases is not None:
+            if len(gligen_images) != len(gligen_phrases):
+                raise ValueError(
+                    "`gligen_images` and `gligen_phrases` must have the same length when both are provided, but"
+                    f" got: `gligen_images` with length {len(gligen_images)} != `gligen_phrases` with length {len(gligen_phrases)}."
+                )
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
     def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
         shape = (
@@ -814,6 +822,8 @@ class StableDiffusionGLIGENTextImagePipeline(DiffusionPipeline, StableDiffusionM
             height,
             width,
             callback_steps,
+            gligen_images,
+            gligen_phrases,
             negative_prompt,
             prompt_embeds,
             negative_prompt_embeds,

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py

@@ -237,11 +237,8 @@ class StableDiffusionXLPipeline(
     _callback_tensor_inputs = [
         "latents",
         "prompt_embeds",
-        "negative_prompt_embeds",
         "add_text_embeds",
         "add_time_ids",
-        "negative_pooled_prompt_embeds",
-        "negative_add_time_ids",
     ]
 
     def __init__(
@@ -1243,13 +1240,8 @@ class StableDiffusionXLPipeline(
 
                     latents = callback_outputs.pop("latents", latents)
                     prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
-                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
                     add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
-                    negative_pooled_prompt_embeds = callback_outputs.pop(
-                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
-                    )
                     add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
-                    negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)
 
                 # call the callback, if provided
                 if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py

@@ -257,11 +257,8 @@ class StableDiffusionXLImg2ImgPipeline(
     _callback_tensor_inputs = [
         "latents",
         "prompt_embeds",
-        "negative_prompt_embeds",
         "add_text_embeds",
         "add_time_ids",
-        "negative_pooled_prompt_embeds",
-        "add_neg_time_ids",
     ]
 
     def __init__(
@@ -1438,13 +1435,8 @@ class StableDiffusionXLImg2ImgPipeline(
 
                     latents = callback_outputs.pop("latents", latents)
                     prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
-                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
                     add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
-                    negative_pooled_prompt_embeds = callback_outputs.pop(
-                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
-                    )
                     add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
-                    add_neg_time_ids = callback_outputs.pop("add_neg_time_ids", add_neg_time_ids)
 
                 # call the callback, if provided
                 if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):

src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py

@@ -285,11 +285,8 @@ class StableDiffusionXLInpaintPipeline(
     _callback_tensor_inputs = [
         "latents",
         "prompt_embeds",
-        "negative_prompt_embeds",
         "add_text_embeds",
         "add_time_ids",
-        "negative_pooled_prompt_embeds",
-        "add_neg_time_ids",
         "mask",
         "masked_image_latents",
     ]
@@ -1671,13 +1668,8 @@ class StableDiffusionXLInpaintPipeline(
 
                     latents = callback_outputs.pop("latents", latents)
                     prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
-                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
                     add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
-                    negative_pooled_prompt_embeds = callback_outputs.pop(
-                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
-                    )
                     add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
-                    add_neg_time_ids = callback_outputs.pop("add_neg_time_ids", add_neg_time_ids)
                     mask = callback_outputs.pop("mask", mask)
                     masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents)
 

src/diffusers/pipelines/unidiffuser/modeling_uvit.py

@@ -104,8 +104,8 @@ class PatchEmbed(nn.Module):
 
         self.use_pos_embed = use_pos_embed
         if self.use_pos_embed:
-            pos_embed = get_2d_sincos_pos_embed(embed_dim, int(num_patches**0.5))
-            self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False)
+            pos_embed = get_2d_sincos_pos_embed(embed_dim, int(num_patches**0.5), output_type="pt")
+            self.register_buffer("pos_embed", pos_embed.float().unsqueeze(0), persistent=False)
 
     def forward(self, latent):
         latent = self.proj(latent)

src/diffusers/schedulers/scheduling_deis_multistep.py

@@ -149,6 +149,8 @@ class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
         use_karras_sigmas: Optional[bool] = False,
         use_exponential_sigmas: Optional[bool] = False,
         use_beta_sigmas: Optional[bool] = False,
+        use_flow_sigmas: Optional[bool] = False,
+        flow_shift: Optional[float] = 1.0,
         timestep_spacing: str = "linspace",
         steps_offset: int = 0,
     ):
@@ -282,6 +284,11 @@ class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
             sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
             timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
             sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
+        elif self.config.use_flow_sigmas:
+            alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)
+            sigmas = 1.0 - alphas
+            sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1]
+            timesteps = (sigmas * self.config.num_train_timesteps).copy()
         else:
             sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
             sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
@@ -362,8 +369,12 @@ class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
 
     # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
     def _sigma_to_alpha_sigma_t(self, sigma):
-        alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
-        sigma_t = sigma * alpha_t
+        if self.config.use_flow_sigmas:
+            alpha_t = 1 - sigma
+            sigma_t = sigma
+        else:
+            alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
+            sigma_t = sigma * alpha_t
 
         return alpha_t, sigma_t
 
@@ -490,10 +501,13 @@ class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
             x0_pred = model_output
         elif self.config.prediction_type == "v_prediction":
             x0_pred = alpha_t * sample - sigma_t * model_output
+        elif self.config.prediction_type == "flow_prediction":
+            sigma_t = self.sigmas[self.step_index]
+            x0_pred = sample - sigma_t * model_output
         else:
             raise ValueError(
-                f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
-                " `v_prediction` for the DEISMultistepScheduler."
+                f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, "
+                "`v_prediction`, or `flow_prediction` for the DEISMultistepScheduler."
             )
 
         if self.config.thresholding:

src/diffusers/schedulers/scheduling_dpmsolver_multistep.py

@@ -218,6 +218,8 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         use_exponential_sigmas: Optional[bool] = False,
         use_beta_sigmas: Optional[bool] = False,
         use_lu_lambdas: Optional[bool] = False,
+        use_flow_sigmas: Optional[bool] = False,
+        flow_shift: Optional[float] = 1.0,
         final_sigmas_type: Optional[str] = "zero",  # "zero", "sigma_min"
         lambda_min_clipped: float = -float("inf"),
         variance_type: Optional[str] = None,
@@ -407,6 +409,11 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
             sigmas = np.flip(sigmas).copy()
             sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
             timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
+        elif self.config.use_flow_sigmas:
+            alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)
+            sigmas = 1.0 - alphas
+            sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1]
+            timesteps = (sigmas * self.config.num_train_timesteps).copy()
         else:
             sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
 
@@ -495,8 +502,12 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
         return t
 
     def _sigma_to_alpha_sigma_t(self, sigma):
-        alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
-        sigma_t = sigma * alpha_t
+        if self.config.use_flow_sigmas:
+            alpha_t = 1 - sigma
+            sigma_t = sigma
+        else:
+            alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
+            sigma_t = sigma * alpha_t
 
         return alpha_t, sigma_t
 
@@ -650,10 +661,13 @@ class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
                 sigma = self.sigmas[self.step_index]
                 alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
                 x0_pred = alpha_t * sample - sigma_t * model_output
+            elif self.config.prediction_type == "flow_prediction":
+                sigma_t = self.sigmas[self.step_index]
+                x0_pred = sample - sigma_t * model_output
             else:
                 raise ValueError(
-                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
-                    " `v_prediction` for the DPMSolverMultistepScheduler."
+                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, "
+                    "`v_prediction`, or `flow_prediction` for the DPMSolverMultistepScheduler."
                 )
 
             if self.config.thresholding:

src/diffusers/schedulers/scheduling_dpmsolver_multistep_inverse.py

@@ -169,6 +169,8 @@ class DPMSolverMultistepInverseScheduler(SchedulerMixin, ConfigMixin):
         use_karras_sigmas: Optional[bool] = False,
         use_exponential_sigmas: Optional[bool] = False,
         use_beta_sigmas: Optional[bool] = False,
+        use_flow_sigmas: Optional[bool] = False,
+        flow_shift: Optional[float] = 1.0,
         lambda_min_clipped: float = -float("inf"),
         variance_type: Optional[str] = None,
         timestep_spacing: str = "linspace",
@@ -292,6 +294,11 @@ class DPMSolverMultistepInverseScheduler(SchedulerMixin, ConfigMixin):
         elif self.config.use_beta_sigmas:
             sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
             timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
+        elif self.config.use_flow_sigmas:
+            alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)
+            sigmas = 1.0 - alphas
+            sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1]
+            timesteps = (sigmas * self.config.num_train_timesteps).copy()
         else:
             sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
             sigma_max = (
@@ -379,8 +386,12 @@ class DPMSolverMultistepInverseScheduler(SchedulerMixin, ConfigMixin):
 
     # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
     def _sigma_to_alpha_sigma_t(self, sigma):
-        alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
-        sigma_t = sigma * alpha_t
+        if self.config.use_flow_sigmas:
+            alpha_t = 1 - sigma
+            sigma_t = sigma
+        else:
+            alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
+            sigma_t = sigma * alpha_t
 
         return alpha_t, sigma_t
 
@@ -522,10 +533,13 @@ class DPMSolverMultistepInverseScheduler(SchedulerMixin, ConfigMixin):
                 sigma = self.sigmas[self.step_index]
                 alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
                 x0_pred = alpha_t * sample - sigma_t * model_output
+            elif self.config.prediction_type == "flow_prediction":
+                sigma_t = self.sigmas[self.step_index]
+                x0_pred = sample - sigma_t * model_output
             else:
                 raise ValueError(
-                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
-                    " `v_prediction` for the DPMSolverMultistepScheduler."
+                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, "
+                    "`v_prediction`, or `flow_prediction` for the DPMSolverMultistepScheduler."
                 )
 
             if self.config.thresholding:

src/diffusers/schedulers/scheduling_dpmsolver_singlestep.py

@@ -164,6 +164,8 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
         use_karras_sigmas: Optional[bool] = False,
         use_exponential_sigmas: Optional[bool] = False,
         use_beta_sigmas: Optional[bool] = False,
+        use_flow_sigmas: Optional[bool] = False,
+        flow_shift: Optional[float] = 1.0,
         final_sigmas_type: Optional[str] = "zero",  # "zero", "sigma_min"
         lambda_min_clipped: float = -float("inf"),
         variance_type: Optional[str] = None,
@@ -356,6 +358,11 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
             sigmas = np.flip(sigmas).copy()
             sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
             timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
+        elif self.config.use_flow_sigmas:
+            alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)
+            sigmas = 1.0 - alphas
+            sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1]
+            timesteps = (sigmas * self.config.num_train_timesteps).copy()
         else:
             sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
 
@@ -454,8 +461,12 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
 
     # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
     def _sigma_to_alpha_sigma_t(self, sigma):
-        alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
-        sigma_t = sigma * alpha_t
+        if self.config.use_flow_sigmas:
+            alpha_t = 1 - sigma
+            sigma_t = sigma
+        else:
+            alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
+            sigma_t = sigma * alpha_t
 
         return alpha_t, sigma_t
 
@@ -595,10 +606,13 @@ class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
                 sigma = self.sigmas[self.step_index]
                 alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
                 x0_pred = alpha_t * sample - sigma_t * model_output
+            elif self.config.prediction_type == "flow_prediction":
+                sigma_t = self.sigmas[self.step_index]
+                x0_pred = sample - sigma_t * model_output
             else:
                 raise ValueError(
-                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
-                    " `v_prediction` for the DPMSolverSinglestepScheduler."
+                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, "
+                    "`v_prediction`, or `flow_prediction` for the DPMSolverSinglestepScheduler."
                 )
 
             if self.config.thresholding:

src/diffusers/schedulers/scheduling_flow_match_euler_discrete.py

@@ -75,6 +75,7 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         base_image_seq_len: Optional[int] = 256,
         max_image_seq_len: Optional[int] = 4096,
         invert_sigmas: bool = False,
+        shift_terminal: Optional[float] = None,
         use_karras_sigmas: Optional[bool] = False,
         use_exponential_sigmas: Optional[bool] = False,
         use_beta_sigmas: Optional[bool] = False,
@@ -181,6 +182,27 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
     def time_shift(self, mu: float, sigma: float, t: torch.Tensor):
         return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
 
+    def stretch_shift_to_terminal(self, t: torch.Tensor) -> torch.Tensor:
+        r"""
+        Stretches and shifts the timestep schedule to ensure it terminates at the configured `shift_terminal` config
+        value.
+
+        Reference:
+        https://github.com/Lightricks/LTX-Video/blob/a01a171f8fe3d99dce2728d60a73fecf4d4238ae/ltx_video/schedulers/rf.py#L51
+
+        Args:
+            t (`torch.Tensor`):
+                A tensor of timesteps to be stretched and shifted.
+
+        Returns:
+            `torch.Tensor`:
+                A tensor of adjusted timesteps such that the final value equals `self.config.shift_terminal`.
+        """
+        one_minus_z = 1 - t
+        scale_factor = one_minus_z[-1] / (1 - self.config.shift_terminal)
+        stretched_t = 1 - (one_minus_z / scale_factor)
+        return stretched_t
+
     def set_timesteps(
         self,
         num_inference_steps: int = None,
@@ -216,6 +238,9 @@ class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
         else:
             sigmas = self.config.shift * sigmas / (1 + (self.config.shift - 1) * sigmas)
 
+        if self.config.shift_terminal:
+            sigmas = self.stretch_shift_to_terminal(sigmas)
+
         if self.config.use_karras_sigmas:
             sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
 

src/diffusers/schedulers/scheduling_sasolver.py

@@ -167,6 +167,8 @@ class SASolverScheduler(SchedulerMixin, ConfigMixin):
         use_karras_sigmas: Optional[bool] = False,
         use_exponential_sigmas: Optional[bool] = False,
         use_beta_sigmas: Optional[bool] = False,
+        use_flow_sigmas: Optional[bool] = False,
+        flow_shift: Optional[float] = 1.0,
         lambda_min_clipped: float = -float("inf"),
         variance_type: Optional[str] = None,
         timestep_spacing: str = "linspace",
@@ -311,6 +313,11 @@ class SASolverScheduler(SchedulerMixin, ConfigMixin):
             sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
             timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
             sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
+        elif self.config.use_flow_sigmas:
+            alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)
+            sigmas = 1.0 - alphas
+            sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1]
+            timesteps = (sigmas * self.config.num_train_timesteps).copy()
         else:
             sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
             sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
@@ -391,8 +398,12 @@ class SASolverScheduler(SchedulerMixin, ConfigMixin):
 
     # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
     def _sigma_to_alpha_sigma_t(self, sigma):
-        alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
-        sigma_t = sigma * alpha_t
+        if self.config.use_flow_sigmas:
+            alpha_t = 1 - sigma
+            sigma_t = sigma
+        else:
+            alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
+            sigma_t = sigma * alpha_t
 
         return alpha_t, sigma_t
 
@@ -531,10 +542,13 @@ class SASolverScheduler(SchedulerMixin, ConfigMixin):
                 x0_pred = model_output
             elif self.config.prediction_type == "v_prediction":
                 x0_pred = alpha_t * sample - sigma_t * model_output
+            elif self.config.prediction_type == "flow_prediction":
+                sigma_t = self.sigmas[self.step_index]
+                x0_pred = sample - sigma_t * model_output
             else:
                 raise ValueError(
-                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
-                    " `v_prediction` for the SASolverScheduler."
+                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, "
+                    "`v_prediction`, or `flow_prediction` for the SASolverScheduler."
                 )
 
             if self.config.thresholding:

src/diffusers/schedulers/scheduling_unipc_multistep.py

@@ -206,6 +206,8 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
         use_karras_sigmas: Optional[bool] = False,
         use_exponential_sigmas: Optional[bool] = False,
         use_beta_sigmas: Optional[bool] = False,
+        use_flow_sigmas: Optional[bool] = False,
+        flow_shift: Optional[float] = 1.0,
         timestep_spacing: str = "linspace",
         steps_offset: int = 0,
         final_sigmas_type: Optional[str] = "zero",  # "zero", "sigma_min"
@@ -374,6 +376,11 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
                     f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}"
                 )
             sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32)
+        elif self.config.use_flow_sigmas:
+            alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)
+            sigmas = 1.0 - alphas
+            sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1]
+            timesteps = (sigmas * self.config.num_train_timesteps).copy()
         else:
             sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
             if self.config.final_sigmas_type == "sigma_min":
@@ -464,8 +471,12 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
 
     # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
     def _sigma_to_alpha_sigma_t(self, sigma):
-        alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
-        sigma_t = sigma * alpha_t
+        if self.config.use_flow_sigmas:
+            alpha_t = 1 - sigma
+            sigma_t = sigma
+        else:
+            alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
+            sigma_t = sigma * alpha_t
 
         return alpha_t, sigma_t
 
@@ -594,10 +605,13 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
                 x0_pred = model_output
             elif self.config.prediction_type == "v_prediction":
                 x0_pred = alpha_t * sample - sigma_t * model_output
+            elif self.config.prediction_type == "flow_prediction":
+                sigma_t = self.sigmas[self.step_index]
+                x0_pred = sample - sigma_t * model_output
             else:
                 raise ValueError(
-                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
-                    " `v_prediction` for the UniPCMultistepScheduler."
+                    f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, "
+                    "`v_prediction`, or `flow_prediction` for the UniPCMultistepScheduler."
                 )
 
             if self.config.thresholding:

src/diffusers/utils/dummy_pt_objects.py

@@ -107,6 +107,21 @@ class AutoencoderKLCogVideoX(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class AutoencoderKLLTXVideo(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class AutoencoderKLMochi(metaclass=DummyObject):
     _backends = ["torch"]
 
@@ -227,6 +242,21 @@ class ControlNetModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class ControlNetUnionModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class ControlNetXSAdapter(metaclass=DummyObject):
     _backends = ["torch"]
 
@@ -392,6 +422,21 @@ class LatteTransformer3DModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class LTXVideoTransformer3DModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class LuminaNextDiT2DModel(metaclass=DummyObject):
     _backends = ["torch"]
 
@@ -512,6 +557,21 @@ class PriorTransformer(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class SanaTransformer2DModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class SD3ControlNetModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -1067,6 +1067,36 @@ class LEditsPPPipelineStableDiffusionXL(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class LTXImageToVideoPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
+class LTXPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class LuminaText2ImgPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 
@@ -1232,6 +1262,36 @@ class ReduxImageEncoder(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class SanaPAGPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
+class SanaPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class SemanticStableDiffusionPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 
@@ -1982,6 +2042,51 @@ class StableDiffusionXLControlNetPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class StableDiffusionXLControlNetUnionImg2ImgPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionXLControlNetUnionInpaintPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionXLControlNetUnionPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class StableDiffusionXLControlNetXSPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/lora/test_lora_layers_flux.py

@@ -430,6 +430,122 @@ class FluxControlLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
         self.assertTrue(not np.allclose(original_output, lora_output_diff_alpha, atol=1e-3, rtol=1e-3))
         self.assertTrue(not np.allclose(lora_output_diff_alpha, lora_output_same_rank, atol=1e-3, rtol=1e-3))
 
+    def test_lora_expanding_shape_with_normal_lora_raises_error(self):
+        # TODO: This test checks if an error is raised when a lora expands shapes (like control loras) but
+        # another lora with correct shapes is loaded. This is not supported at the moment and should raise an error.
+        # When we do support it, this test should be removed. Context: https://github.com/huggingface/diffusers/issues/10180
+        components, _, _ = self.get_dummy_components(FlowMatchEulerDiscreteScheduler)
+
+        # Change the transformer config to mimic a real use case.
+        num_channels_without_control = 4
+        transformer = FluxTransformer2DModel.from_config(
+            components["transformer"].config, in_channels=num_channels_without_control
+        ).to(torch_device)
+        components["transformer"] = transformer
+
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        logger = logging.get_logger("diffusers.loaders.lora_pipeline")
+        logger.setLevel(logging.DEBUG)
+
+        out_features, in_features = pipe.transformer.x_embedder.weight.shape
+        rank = 4
+
+        shape_expander_lora_A = torch.nn.Linear(2 * in_features, rank, bias=False)
+        shape_expander_lora_B = torch.nn.Linear(rank, out_features, bias=False)
+        lora_state_dict = {
+            "transformer.x_embedder.lora_A.weight": shape_expander_lora_A.weight,
+            "transformer.x_embedder.lora_B.weight": shape_expander_lora_B.weight,
+        }
+        with CaptureLogger(logger) as cap_logger:
+            pipe.load_lora_weights(lora_state_dict, "adapter-1")
+
+        self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in denoiser")
+        self.assertTrue(pipe.get_active_adapters() == ["adapter-1"])
+        self.assertTrue(pipe.transformer.x_embedder.weight.data.shape[1] == 2 * in_features)
+        self.assertTrue(pipe.transformer.config.in_channels == 2 * in_features)
+        self.assertTrue(cap_logger.out.startswith("Expanding the nn.Linear input/output features for module"))
+
+        _, _, inputs = self.get_dummy_inputs(with_generator=False)
+        lora_output = pipe(**inputs, generator=torch.manual_seed(0))[0]
+
+        normal_lora_A = torch.nn.Linear(in_features, rank, bias=False)
+        normal_lora_B = torch.nn.Linear(rank, out_features, bias=False)
+        lora_state_dict = {
+            "transformer.x_embedder.lora_A.weight": normal_lora_A.weight,
+            "transformer.x_embedder.lora_B.weight": normal_lora_B.weight,
+        }
+
+        # The first lora expanded the input features of x_embedder. Here, we are trying to load a lora with the correct
+        # input features before expansion. This should raise an error about the weight shapes being incompatible.
+        self.assertRaisesRegex(
+            RuntimeError,
+            "size mismatch for x_embedder.lora_A.adapter-2.weight",
+            pipe.load_lora_weights,
+            lora_state_dict,
+            "adapter-2",
+        )
+        # We should have `adapter-1` as the only adapter.
+        self.assertTrue(pipe.get_active_adapters() == ["adapter-1"])
+
+        # Check if the output is the same after lora loading error
+        lora_output_after_error = pipe(**inputs, generator=torch.manual_seed(0))[0]
+        self.assertTrue(np.allclose(lora_output, lora_output_after_error, atol=1e-3, rtol=1e-3))
+
+        # Test the opposite case where the first lora has the correct input features and the second lora has expanded input features.
+        # This should raise a runtime error on input shapes being incompatible. But it doesn't. This is because PEFT renames the
+        # original layers as `base_layer` and the lora layers with the adapter names. This makes our logic to check if a lora
+        # weight is compatible with the current model inadequate. This should be addressed when attempting support for
+        # https://github.com/huggingface/diffusers/issues/10180 (TODO)
+        components, _, _ = self.get_dummy_components(FlowMatchEulerDiscreteScheduler)
+        # Change the transformer config to mimic a real use case.
+        num_channels_without_control = 4
+        transformer = FluxTransformer2DModel.from_config(
+            components["transformer"].config, in_channels=num_channels_without_control
+        ).to(torch_device)
+        components["transformer"] = transformer
+
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        logger = logging.get_logger("diffusers.loaders.lora_pipeline")
+        logger.setLevel(logging.DEBUG)
+
+        out_features, in_features = pipe.transformer.x_embedder.weight.shape
+        rank = 4
+
+        lora_state_dict = {
+            "transformer.x_embedder.lora_A.weight": normal_lora_A.weight,
+            "transformer.x_embedder.lora_B.weight": normal_lora_B.weight,
+        }
+
+        with CaptureLogger(logger) as cap_logger:
+            pipe.load_lora_weights(lora_state_dict, "adapter-1")
+            self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in denoiser")
+
+        self.assertTrue(pipe.transformer.x_embedder.weight.data.shape[1] == in_features)
+        self.assertTrue(pipe.transformer.config.in_channels == in_features)
+        self.assertFalse(cap_logger.out.startswith("Expanding the nn.Linear input/output features for module"))
+
+        lora_state_dict = {
+            "transformer.x_embedder.lora_A.weight": shape_expander_lora_A.weight,
+            "transformer.x_embedder.lora_B.weight": shape_expander_lora_B.weight,
+        }
+
+        # We should check for input shapes being incompatible here. But because above mentioned issue is
+        # not a supported use case, and because of the PEFT renaming, we will currently have a shape
+        # mismatch error.
+        self.assertRaisesRegex(
+            RuntimeError,
+            "size mismatch for x_embedder.lora_A.adapter-2.weight",
+            pipe.load_lora_weights,
+            lora_state_dict,
+            "adapter-2",
+        )
+
     @unittest.skip("Not supported in Flux.")
     def test_simple_inference_with_text_denoiser_block_scale_for_all_dict_options(self):
         pass

tests/lora/utils.py

@@ -76,6 +76,9 @@ def initialize_dummy_state_dict(state_dict):
     return {k: torch.randn(v.shape, device=torch_device, dtype=v.dtype) for k, v in state_dict.items()}
 
 
+POSSIBLE_ATTENTION_KWARGS_NAMES = ["cross_attention_kwargs", "joint_attention_kwargs", "attention_kwargs"]
+
+
 @require_peft_backend
 class PeftLoraLoaderMixinTests:
     pipeline_class = None
@@ -429,7 +432,7 @@ class PeftLoraLoaderMixinTests:
         call_signature_keys = inspect.signature(self.pipeline_class.__call__).parameters.keys()
 
         # TODO(diffusers): Discuss a common naming convention across library for 1.0.0 release
-        for possible_attention_kwargs in ["cross_attention_kwargs", "joint_attention_kwargs", "attention_kwargs"]:
+        for possible_attention_kwargs in POSSIBLE_ATTENTION_KWARGS_NAMES:
             if possible_attention_kwargs in call_signature_keys:
                 attention_kwargs_name = possible_attention_kwargs
                 break
@@ -790,7 +793,7 @@ class PeftLoraLoaderMixinTests:
         and makes sure it works as expected
         """
         call_signature_keys = inspect.signature(self.pipeline_class.__call__).parameters.keys()
-        for possible_attention_kwargs in ["cross_attention_kwargs", "joint_attention_kwargs", "attention_kwargs"]:
+        for possible_attention_kwargs in POSSIBLE_ATTENTION_KWARGS_NAMES:
             if possible_attention_kwargs in call_signature_keys:
                 attention_kwargs_name = possible_attention_kwargs
                 break
@@ -1885,3 +1888,88 @@ class PeftLoraLoaderMixinTests:
 
             _, _, inputs = self.get_dummy_inputs()
             _ = pipe(**inputs)[0]
+
+    def test_set_adapters_match_attention_kwargs(self):
+        """Test to check if outputs after `set_adapters()` and attention kwargs match."""
+        call_signature_keys = inspect.signature(self.pipeline_class.__call__).parameters.keys()
+        for possible_attention_kwargs in POSSIBLE_ATTENTION_KWARGS_NAMES:
+            if possible_attention_kwargs in call_signature_keys:
+                attention_kwargs_name = possible_attention_kwargs
+                break
+        assert attention_kwargs_name is not None
+
+        for scheduler_cls in self.scheduler_classes:
+            components, text_lora_config, denoiser_lora_config = self.get_dummy_components(scheduler_cls)
+            pipe = self.pipeline_class(**components)
+            pipe = pipe.to(torch_device)
+            pipe.set_progress_bar_config(disable=None)
+            _, _, inputs = self.get_dummy_inputs(with_generator=False)
+
+            output_no_lora = pipe(**inputs, generator=torch.manual_seed(0))[0]
+            self.assertTrue(output_no_lora.shape == self.output_shape)
+
+            if "text_encoder" in self.pipeline_class._lora_loadable_modules:
+                pipe.text_encoder.add_adapter(text_lora_config)
+                self.assertTrue(
+                    check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder"
+                )
+
+            denoiser = pipe.transformer if self.unet_kwargs is None else pipe.unet
+            denoiser.add_adapter(denoiser_lora_config)
+            self.assertTrue(check_if_lora_correctly_set(denoiser), "Lora not correctly set in denoiser.")
+
+            if self.has_two_text_encoders or self.has_three_text_encoders:
+                if "text_encoder_2" in self.pipeline_class._lora_loadable_modules:
+                    pipe.text_encoder_2.add_adapter(text_lora_config)
+                    self.assertTrue(
+                        check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
+                    )
+
+            lora_scale = 0.5
+            attention_kwargs = {attention_kwargs_name: {"scale": lora_scale}}
+            output_lora_scale = pipe(**inputs, generator=torch.manual_seed(0), **attention_kwargs)[0]
+            self.assertFalse(
+                np.allclose(output_no_lora, output_lora_scale, atol=1e-3, rtol=1e-3),
+                "Lora + scale should change the output",
+            )
+
+            pipe.set_adapters("default", lora_scale)
+            output_lora_scale_wo_kwargs = pipe(**inputs, generator=torch.manual_seed(0))[0]
+            self.assertTrue(
+                not np.allclose(output_no_lora, output_lora_scale_wo_kwargs, atol=1e-3, rtol=1e-3),
+                "Lora + scale should change the output",
+            )
+            self.assertTrue(
+                np.allclose(output_lora_scale, output_lora_scale_wo_kwargs, atol=1e-3, rtol=1e-3),
+                "Lora + scale should match the output of `set_adapters()`.",
+            )
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                modules_to_save = self._get_modules_to_save(pipe, has_denoiser=True)
+                lora_state_dicts = self._get_lora_state_dicts(modules_to_save)
+                self.pipeline_class.save_lora_weights(
+                    save_directory=tmpdirname, safe_serialization=True, **lora_state_dicts
+                )
+
+                self.assertTrue(os.path.isfile(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors")))
+                pipe = self.pipeline_class(**components)
+                pipe = pipe.to(torch_device)
+                pipe.set_progress_bar_config(disable=None)
+                pipe.load_lora_weights(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors"))
+
+                for module_name, module in modules_to_save.items():
+                    self.assertTrue(check_if_lora_correctly_set(module), f"Lora not correctly set in {module_name}")
+
+                output_lora_from_pretrained = pipe(**inputs, generator=torch.manual_seed(0), **attention_kwargs)[0]
+                self.assertTrue(
+                    not np.allclose(output_no_lora, output_lora_from_pretrained, atol=1e-3, rtol=1e-3),
+                    "Lora + scale should change the output",
+                )
+                self.assertTrue(
+                    np.allclose(output_lora_scale, output_lora_from_pretrained, atol=1e-3, rtol=1e-3),
+                    "Loading from saved checkpoints should give same results as attention_kwargs.",
+                )
+                self.assertTrue(
+                    np.allclose(output_lora_scale_wo_kwargs, output_lora_from_pretrained, atol=1e-3, rtol=1e-3),
+                    "Loading from saved checkpoints should give same results as set_adapters().",
+                )

tests/models/transformers/test_models_transformer_ltx.py

@@ -0,0 +1,83 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import torch
+
+from diffusers import LTXVideoTransformer3DModel
+from diffusers.utils.testing_utils import enable_full_determinism, torch_device
+
+from ..test_modeling_common import ModelTesterMixin
+
+
+enable_full_determinism()
+
+
+class LTXTransformerTests(ModelTesterMixin, unittest.TestCase):
+    model_class = LTXVideoTransformer3DModel
+    main_input_name = "hidden_states"
+    uses_custom_attn_processor = True
+
+    @property
+    def dummy_input(self):
+        batch_size = 2
+        num_channels = 4
+        num_frames = 2
+        height = 16
+        width = 16
+        embedding_dim = 16
+        sequence_length = 16
+
+        hidden_states = torch.randn((batch_size, num_frames * height * width, num_channels)).to(torch_device)
+        encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
+        encoder_attention_mask = torch.ones((batch_size, sequence_length)).bool().to(torch_device)
+        timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device)
+
+        return {
+            "hidden_states": hidden_states,
+            "encoder_hidden_states": encoder_hidden_states,
+            "timestep": timestep,
+            "encoder_attention_mask": encoder_attention_mask,
+            "num_frames": num_frames,
+            "height": height,
+            "width": width,
+        }
+
+    @property
+    def input_shape(self):
+        return (512, 4)
+
+    @property
+    def output_shape(self):
+        return (512, 4)
+
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
+            "in_channels": 4,
+            "out_channels": 4,
+            "num_attention_heads": 2,
+            "attention_head_dim": 8,
+            "cross_attention_dim": 16,
+            "num_layers": 1,
+            "qk_norm": "rms_norm_across_heads",
+            "caption_channels": 16,
+        }
+        inputs_dict = self.dummy_input
+        return init_dict, inputs_dict
+
+    def test_gradient_checkpointing_is_applied(self):
+        expected_set = {"LTXVideoTransformer3DModel"}
+        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)

tests/models/transformers/test_models_transformer_sana.py

@@ -0,0 +1,82 @@
+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import torch
+
+from diffusers import SanaTransformer2DModel
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    torch_device,
+)
+
+from ..test_modeling_common import ModelTesterMixin
+
+
+enable_full_determinism()
+
+
+class SanaTransformerTests(ModelTesterMixin, unittest.TestCase):
+    model_class = SanaTransformer2DModel
+    main_input_name = "hidden_states"
+    uses_custom_attn_processor = True
+
+    @property
+    def dummy_input(self):
+        batch_size = 2
+        num_channels = 4
+        height = 32
+        width = 32
+        embedding_dim = 8
+        sequence_length = 8
+
+        hidden_states = torch.randn((batch_size, num_channels, height, width)).to(torch_device)
+        encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
+        timestep = torch.randint(0, 1000, size=(batch_size,)).to(torch_device)
+
+        return {
+            "hidden_states": hidden_states,
+            "encoder_hidden_states": encoder_hidden_states,
+            "timestep": timestep,
+        }
+
+    @property
+    def input_shape(self):
+        return (4, 32, 32)
+
+    @property
+    def output_shape(self):
+        return (4, 32, 32)
+
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
+            "patch_size": 1,
+            "in_channels": 4,
+            "out_channels": 4,
+            "num_layers": 1,
+            "attention_head_dim": 4,
+            "num_attention_heads": 2,
+            "num_cross_attention_heads": 2,
+            "cross_attention_head_dim": 4,
+            "cross_attention_dim": 8,
+            "caption_channels": 8,
+            "sample_size": 32,
+        }
+        inputs_dict = self.dummy_input
+        return init_dict, inputs_dict
+
+    def test_gradient_checkpointing_is_applied(self):
+        expected_set = {"SanaTransformer2DModel"}
+        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)

tests/models/transformers/test_models_transformer_sd3.py

@@ -18,6 +18,7 @@ import unittest
 import torch
 
 from diffusers import SD3Transformer2DModel
+from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.testing_utils import (
     enable_full_determinism,
     torch_device,
@@ -80,6 +81,20 @@ class SD3TransformerTests(ModelTesterMixin, unittest.TestCase):
         inputs_dict = self.dummy_input
         return init_dict, inputs_dict
 
+    @unittest.skipIf(
+        torch_device != "cuda" or not is_xformers_available(),
+        reason="XFormers attention is only available with CUDA and `xformers` installed",
+    )
+    def test_xformers_enable_works(self):
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**init_dict)
+
+        model.enable_xformers_memory_efficient_attention()
+
+        assert (
+            model.transformer_blocks[0].attn.processor.__class__.__name__ == "XFormersJointAttnProcessor"
+        ), "xformers is not enabled"
+
     @unittest.skip("SD3Transformer2DModel uses a dedicated attention processor. This test doesn't apply")
     def test_set_attn_processor_for_determinism(self):
         pass
@@ -140,6 +155,20 @@ class SD35TransformerTests(ModelTesterMixin, unittest.TestCase):
         inputs_dict = self.dummy_input
         return init_dict, inputs_dict
 
+    @unittest.skipIf(
+        torch_device != "cuda" or not is_xformers_available(),
+        reason="XFormers attention is only available with CUDA and `xformers` installed",
+    )
+    def test_xformers_enable_works(self):
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**init_dict)
+
+        model.enable_xformers_memory_efficient_attention()
+
+        assert (
+            model.transformer_blocks[0].attn.processor.__class__.__name__ == "XFormersJointAttnProcessor"
+        ), "xformers is not enabled"
+
     @unittest.skip("SD3Transformer2DModel uses a dedicated attention processor. This test doesn't apply")
     def test_set_attn_processor_for_determinism(self):
         pass

tests/pipelines/ltx/test_ltx.py

@@ -0,0 +1,256 @@
+# Copyright 2024 The HuggingFace Team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import unittest
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer, T5EncoderModel
+
+from diffusers import AutoencoderKLLTXVideo, FlowMatchEulerDiscreteScheduler, LTXPipeline, LTXVideoTransformer3DModel
+from diffusers.utils.testing_utils import enable_full_determinism, torch_device
+
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import PipelineTesterMixin, to_np
+
+
+enable_full_determinism()
+
+
+class LTXPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = LTXPipeline
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+    test_xformers_attention = False
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = LTXVideoTransformer3DModel(
+            in_channels=8,
+            out_channels=8,
+            patch_size=1,
+            patch_size_t=1,
+            num_attention_heads=4,
+            attention_head_dim=8,
+            cross_attention_dim=32,
+            num_layers=1,
+            caption_channels=32,
+        )
+
+        torch.manual_seed(0)
+        vae = AutoencoderKLLTXVideo(
+            latent_channels=8,
+            block_out_channels=(8, 8, 8, 8),
+            spatio_temporal_scaling=(True, True, False, False),
+            layers_per_block=(1, 1, 1, 1, 1),
+            patch_size=1,
+            patch_size_t=1,
+            encoder_causal=True,
+            decoder_causal=False,
+        )
+        vae.use_framewise_encoding = False
+        vae.use_framewise_decoding = False
+
+        torch.manual_seed(0)
+        scheduler = FlowMatchEulerDiscreteScheduler()
+        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        components = {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+
+        inputs = {
+            "prompt": "dance monkey",
+            "negative_prompt": "",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 3.0,
+            "height": 32,
+            "width": 32,
+            # 8 * k + 1 is the recommendation
+            "num_frames": 9,
+            "max_sequence_length": 16,
+            "output_type": "pt",
+        }
+
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        video = pipe(**inputs).frames
+        generated_video = video[0]
+
+        self.assertEqual(generated_video.shape, (9, 3, 32, 32))
+        expected_video = torch.randn(9, 3, 32, 32)
+        max_diff = np.abs(generated_video - expected_video).max()
+        self.assertLessEqual(max_diff, 1e10)
+
+    def test_callback_inputs(self):
+        sig = inspect.signature(self.pipeline_class.__call__)
+        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
+        has_callback_step_end = "callback_on_step_end" in sig.parameters
+
+        if not (has_callback_tensor_inputs and has_callback_step_end):
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        self.assertTrue(
+            hasattr(pipe, "_callback_tensor_inputs"),
+            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
+        )
+
+        def callback_inputs_subset(pipe, i, t, callback_kwargs):
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        def callback_inputs_all(pipe, i, t, callback_kwargs):
+            for tensor_name in pipe._callback_tensor_inputs:
+                assert tensor_name in callback_kwargs
+
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        # Test passing in a subset
+        inputs["callback_on_step_end"] = callback_inputs_subset
+        inputs["callback_on_step_end_tensor_inputs"] = ["latents"]
+        output = pipe(**inputs)[0]
+
+        # Test passing in a everything
+        inputs["callback_on_step_end"] = callback_inputs_all
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+
+        def callback_inputs_change_tensor(pipe, i, t, callback_kwargs):
+            is_last = i == (pipe.num_timesteps - 1)
+            if is_last:
+                callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"])
+            return callback_kwargs
+
+        inputs["callback_on_step_end"] = callback_inputs_change_tensor
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+        assert output.abs().sum() < 1e10
+
+    def test_inference_batch_single_identical(self):
+        self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-3)
+
+    def test_attention_slicing_forward_pass(
+        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
+    ):
+        if not self.test_attention_slicing:
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        generator_device = "cpu"
+        inputs = self.get_dummy_inputs(generator_device)
+        output_without_slicing = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=1)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing1 = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=2)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing2 = pipe(**inputs)[0]
+
+        if test_max_difference:
+            max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max()
+            max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max()
+            self.assertLess(
+                max(max_diff1, max_diff2),
+                expected_max_diff,
+                "Attention slicing should not affect the inference results",
+            )
+
+    def test_vae_tiling(self, expected_diff_max: float = 0.2):
+        generator_device = "cpu"
+        components = self.get_dummy_components()
+
+        pipe = self.pipeline_class(**components)
+        pipe.to("cpu")
+        pipe.set_progress_bar_config(disable=None)
+
+        # Without tiling
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_without_tiling = pipe(**inputs)[0]
+
+        # With tiling
+        pipe.vae.enable_tiling(
+            tile_sample_min_height=96,
+            tile_sample_min_width=96,
+            tile_sample_stride_height=64,
+            tile_sample_stride_width=64,
+        )
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_with_tiling = pipe(**inputs)[0]
+
+        self.assertLess(
+            (to_np(output_without_tiling) - to_np(output_with_tiling)).max(),
+            expected_diff_max,
+            "VAE tiling should not affect the inference results",
+        )

tests/pipelines/ltx/test_ltx_image2video.py

@@ -0,0 +1,264 @@
+# Copyright 2024 The HuggingFace Team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import unittest
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer, T5EncoderModel
+
+from diffusers import (
+    AutoencoderKLLTXVideo,
+    FlowMatchEulerDiscreteScheduler,
+    LTXImageToVideoPipeline,
+    LTXVideoTransformer3DModel,
+)
+from diffusers.utils.testing_utils import enable_full_determinism, torch_device
+
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import PipelineTesterMixin, to_np
+
+
+enable_full_determinism()
+
+
+class LTXImageToVideoPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = LTXImageToVideoPipeline
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS.union({"image"})
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+    test_xformers_attention = False
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = LTXVideoTransformer3DModel(
+            in_channels=8,
+            out_channels=8,
+            patch_size=1,
+            patch_size_t=1,
+            num_attention_heads=4,
+            attention_head_dim=8,
+            cross_attention_dim=32,
+            num_layers=1,
+            caption_channels=32,
+        )
+
+        torch.manual_seed(0)
+        vae = AutoencoderKLLTXVideo(
+            latent_channels=8,
+            block_out_channels=(8, 8, 8, 8),
+            spatio_temporal_scaling=(True, True, False, False),
+            layers_per_block=(1, 1, 1, 1, 1),
+            patch_size=1,
+            patch_size_t=1,
+            encoder_causal=True,
+            decoder_causal=False,
+        )
+        vae.use_framewise_encoding = False
+        vae.use_framewise_decoding = False
+
+        torch.manual_seed(0)
+        scheduler = FlowMatchEulerDiscreteScheduler()
+        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        components = {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+
+        image = torch.randn((1, 3, 32, 32), generator=generator, device=device)
+
+        inputs = {
+            "image": image,
+            "prompt": "dance monkey",
+            "negative_prompt": "",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 3.0,
+            "height": 32,
+            "width": 32,
+            # 8 * k + 1 is the recommendation
+            "num_frames": 9,
+            "max_sequence_length": 16,
+            "output_type": "pt",
+        }
+
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        video = pipe(**inputs).frames
+        generated_video = video[0]
+
+        self.assertEqual(generated_video.shape, (9, 3, 32, 32))
+        expected_video = torch.randn(9, 3, 32, 32)
+        max_diff = np.abs(generated_video - expected_video).max()
+        self.assertLessEqual(max_diff, 1e10)
+
+    def test_callback_inputs(self):
+        sig = inspect.signature(self.pipeline_class.__call__)
+        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
+        has_callback_step_end = "callback_on_step_end" in sig.parameters
+
+        if not (has_callback_tensor_inputs and has_callback_step_end):
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        self.assertTrue(
+            hasattr(pipe, "_callback_tensor_inputs"),
+            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
+        )
+
+        def callback_inputs_subset(pipe, i, t, callback_kwargs):
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        def callback_inputs_all(pipe, i, t, callback_kwargs):
+            for tensor_name in pipe._callback_tensor_inputs:
+                assert tensor_name in callback_kwargs
+
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        # Test passing in a subset
+        inputs["callback_on_step_end"] = callback_inputs_subset
+        inputs["callback_on_step_end_tensor_inputs"] = ["latents"]
+        output = pipe(**inputs)[0]
+
+        # Test passing in a everything
+        inputs["callback_on_step_end"] = callback_inputs_all
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+
+        def callback_inputs_change_tensor(pipe, i, t, callback_kwargs):
+            is_last = i == (pipe.num_timesteps - 1)
+            if is_last:
+                callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"])
+            return callback_kwargs
+
+        inputs["callback_on_step_end"] = callback_inputs_change_tensor
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+        assert output.abs().sum() < 1e10
+
+    def test_inference_batch_single_identical(self):
+        self._test_inference_batch_single_identical(batch_size=3, expected_max_diff=1e-3)
+
+    def test_attention_slicing_forward_pass(
+        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
+    ):
+        if not self.test_attention_slicing:
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        generator_device = "cpu"
+        inputs = self.get_dummy_inputs(generator_device)
+        output_without_slicing = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=1)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing1 = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=2)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing2 = pipe(**inputs)[0]
+
+        if test_max_difference:
+            max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max()
+            max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max()
+            self.assertLess(
+                max(max_diff1, max_diff2),
+                expected_max_diff,
+                "Attention slicing should not affect the inference results",
+            )
+
+    def test_vae_tiling(self, expected_diff_max: float = 0.2):
+        generator_device = "cpu"
+        components = self.get_dummy_components()
+
+        pipe = self.pipeline_class(**components)
+        pipe.to("cpu")
+        pipe.set_progress_bar_config(disable=None)
+
+        # Without tiling
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_without_tiling = pipe(**inputs)[0]
+
+        # With tiling
+        pipe.vae.enable_tiling(
+            tile_sample_min_height=96,
+            tile_sample_min_width=96,
+            tile_sample_stride_height=64,
+            tile_sample_stride_width=64,
+        )
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_with_tiling = pipe(**inputs)[0]
+
+        self.assertLess(
+            (to_np(output_without_tiling) - to_np(output_with_tiling)).max(),
+            expected_diff_max,
+            "VAE tiling should not affect the inference results",
+        )

tests/pipelines/pag/test_pag_sana.py

@@ -0,0 +1,339 @@
+# Copyright 2024 The HuggingFace Team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import unittest
+
+import numpy as np
+import torch
+from transformers import Gemma2Config, Gemma2ForCausalLM, GemmaTokenizer
+
+from diffusers import (
+    AutoencoderDC,
+    FlowMatchEulerDiscreteScheduler,
+    SanaPAGPipeline,
+    SanaPipeline,
+    SanaTransformer2DModel,
+)
+from diffusers.utils.testing_utils import enable_full_determinism, torch_device
+
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import PipelineTesterMixin, to_np
+
+
+enable_full_determinism()
+
+
+class SanaPAGPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = SanaPAGPipeline
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+    test_xformers_attention = False
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = SanaTransformer2DModel(
+            patch_size=1,
+            in_channels=4,
+            out_channels=4,
+            num_layers=2,
+            num_attention_heads=2,
+            attention_head_dim=4,
+            num_cross_attention_heads=2,
+            cross_attention_head_dim=4,
+            cross_attention_dim=8,
+            caption_channels=8,
+            sample_size=32,
+        )
+
+        torch.manual_seed(0)
+        vae = AutoencoderDC(
+            in_channels=3,
+            latent_channels=4,
+            attention_head_dim=2,
+            encoder_block_types=(
+                "ResBlock",
+                "EfficientViTBlock",
+            ),
+            decoder_block_types=(
+                "ResBlock",
+                "EfficientViTBlock",
+            ),
+            encoder_block_out_channels=(8, 8),
+            decoder_block_out_channels=(8, 8),
+            encoder_qkv_multiscales=((), (5,)),
+            decoder_qkv_multiscales=((), (5,)),
+            encoder_layers_per_block=(1, 1),
+            decoder_layers_per_block=[1, 1],
+            downsample_block_type="conv",
+            upsample_block_type="interpolate",
+            decoder_norm_types="rms_norm",
+            decoder_act_fns="silu",
+            scaling_factor=0.41407,
+        )
+
+        torch.manual_seed(0)
+        scheduler = FlowMatchEulerDiscreteScheduler(shift=7.0)
+
+        torch.manual_seed(0)
+        text_encoder_config = Gemma2Config(
+            head_dim=16,
+            hidden_size=32,
+            initializer_range=0.02,
+            intermediate_size=64,
+            max_position_embeddings=8192,
+            model_type="gemma2",
+            num_attention_heads=2,
+            num_hidden_layers=1,
+            num_key_value_heads=2,
+            vocab_size=8,
+            attn_implementation="eager",
+        )
+        text_encoder = Gemma2ForCausalLM(text_encoder_config)
+        tokenizer = GemmaTokenizer.from_pretrained("hf-internal-testing/dummy-gemma")
+
+        components = {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+        inputs = {
+            "prompt": "",
+            "negative_prompt": "",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+            "pag_scale": 3.0,
+            "height": 32,
+            "width": 32,
+            "max_sequence_length": 16,
+            "output_type": "pt",
+            "complex_human_instruction": None,
+        }
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs)[0]
+        generated_image = image[0]
+
+        self.assertEqual(generated_image.shape, (3, 32, 32))
+        expected_image = torch.randn(3, 32, 32)
+        max_diff = np.abs(generated_image - expected_image).max()
+        self.assertLessEqual(max_diff, 1e10)
+
+    def test_callback_inputs(self):
+        sig = inspect.signature(self.pipeline_class.__call__)
+        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
+        has_callback_step_end = "callback_on_step_end" in sig.parameters
+
+        if not (has_callback_tensor_inputs and has_callback_step_end):
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        self.assertTrue(
+            hasattr(pipe, "_callback_tensor_inputs"),
+            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
+        )
+
+        def callback_inputs_subset(pipe, i, t, callback_kwargs):
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        def callback_inputs_all(pipe, i, t, callback_kwargs):
+            for tensor_name in pipe._callback_tensor_inputs:
+                assert tensor_name in callback_kwargs
+
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        # Test passing in a subset
+        inputs["callback_on_step_end"] = callback_inputs_subset
+        inputs["callback_on_step_end_tensor_inputs"] = ["latents"]
+        output = pipe(**inputs)[0]
+
+        # Test passing in a everything
+        inputs["callback_on_step_end"] = callback_inputs_all
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+
+        def callback_inputs_change_tensor(pipe, i, t, callback_kwargs):
+            is_last = i == (pipe.num_timesteps - 1)
+            if is_last:
+                callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"])
+            return callback_kwargs
+
+        inputs["callback_on_step_end"] = callback_inputs_change_tensor
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+        assert output.abs().sum() < 1e10
+
+    def test_attention_slicing_forward_pass(
+        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
+    ):
+        if not self.test_attention_slicing:
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        generator_device = "cpu"
+        inputs = self.get_dummy_inputs(generator_device)
+        output_without_slicing = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=1)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing1 = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=2)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing2 = pipe(**inputs)[0]
+
+        if test_max_difference:
+            max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max()
+            max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max()
+            self.assertLess(
+                max(max_diff1, max_diff2),
+                expected_max_diff,
+                "Attention slicing should not affect the inference results",
+            )
+
+    def test_pag_disable_enable(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+
+        # base pipeline (expect same output when pag is disabled)
+        pipe_sd = SanaPipeline(**components)
+        pipe_sd = pipe_sd.to(device)
+        pipe_sd.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        del inputs["pag_scale"]
+        assert (
+            "pag_scale" not in inspect.signature(pipe_sd.__call__).parameters
+        ), f"`pag_scale` should not be a call parameter of the base pipeline {pipe_sd.__class__.__name__}."
+        out = pipe_sd(**inputs).images[0, -3:, -3:, -1]
+
+        components = self.get_dummy_components()
+
+        # pag disabled with pag_scale=0.0
+        pipe_pag = self.pipeline_class(**components)
+        pipe_pag = pipe_pag.to(device)
+        pipe_pag.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["pag_scale"] = 0.0
+        out_pag_disabled = pipe_pag(**inputs).images[0, -3:, -3:, -1]
+
+        assert np.abs(out.flatten() - out_pag_disabled.flatten()).max() < 1e-3
+
+    def test_pag_applied_layers(self):
+        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        components = self.get_dummy_components()
+
+        # base pipeline
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        all_self_attn_layers = [k for k in pipe.transformer.attn_processors.keys() if "attn1" in k]
+        original_attn_procs = pipe.transformer.attn_processors
+        pag_layers = ["blocks.0", "blocks.1"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert set(pipe.pag_attn_processors) == set(all_self_attn_layers)
+
+        # blocks.0
+        block_0_self_attn = ["transformer_blocks.0.attn1.processor"]
+        pipe.transformer.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["blocks.0"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert set(pipe.pag_attn_processors) == set(block_0_self_attn)
+
+        pipe.transformer.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["blocks.0.attn1"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert set(pipe.pag_attn_processors) == set(block_0_self_attn)
+
+        pipe.transformer.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["blocks.(0|1)"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert (len(pipe.pag_attn_processors)) == 2
+
+        pipe.transformer.set_attn_processor(original_attn_procs.copy())
+        pag_layers = ["blocks.0", r"blocks\.1"]
+        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
+        assert len(pipe.pag_attn_processors) == 2
+
+    # TODO(aryan): Create a dummy gemma model with smol vocab size
+    @unittest.skip(
+        "A very small vocab size is used for fast tests. So, any kind of prompt other than the empty default used in other tests will lead to a embedding lookup error. This test uses a long prompt that causes the error."
+    )
+    def test_inference_batch_consistent(self):
+        pass
+
+    @unittest.skip(
+        "A very small vocab size is used for fast tests. So, any kind of prompt other than the empty default used in other tests will lead to a embedding lookup error. This test uses a long prompt that causes the error."
+    )
+    def test_inference_batch_single_identical(self):
+        pass
+
+    def test_float16_inference(self):
+        # Requires higher tolerance as model seems very sensitive to dtype
+        super().test_float16_inference(expected_max_diff=0.08)

tests/pipelines/sana/test_sana.py

@@ -0,0 +1,340 @@
+# Copyright 2024 The HuggingFace Team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import gc
+import inspect
+import unittest
+
+import numpy as np
+import torch
+from transformers import Gemma2Config, Gemma2ForCausalLM, GemmaTokenizer
+
+from diffusers import AutoencoderDC, FlowMatchEulerDiscreteScheduler, SanaPipeline, SanaTransformer2DModel
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import PipelineTesterMixin, to_np
+
+
+enable_full_determinism()
+
+
+class SanaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = SanaPipeline
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "latents",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+    test_xformers_attention = False
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = SanaTransformer2DModel(
+            patch_size=1,
+            in_channels=4,
+            out_channels=4,
+            num_layers=1,
+            num_attention_heads=2,
+            attention_head_dim=4,
+            num_cross_attention_heads=2,
+            cross_attention_head_dim=4,
+            cross_attention_dim=8,
+            caption_channels=8,
+            sample_size=32,
+        )
+
+        torch.manual_seed(0)
+        vae = AutoencoderDC(
+            in_channels=3,
+            latent_channels=4,
+            attention_head_dim=2,
+            encoder_block_types=(
+                "ResBlock",
+                "EfficientViTBlock",
+            ),
+            decoder_block_types=(
+                "ResBlock",
+                "EfficientViTBlock",
+            ),
+            encoder_block_out_channels=(8, 8),
+            decoder_block_out_channels=(8, 8),
+            encoder_qkv_multiscales=((), (5,)),
+            decoder_qkv_multiscales=((), (5,)),
+            encoder_layers_per_block=(1, 1),
+            decoder_layers_per_block=[1, 1],
+            downsample_block_type="conv",
+            upsample_block_type="interpolate",
+            decoder_norm_types="rms_norm",
+            decoder_act_fns="silu",
+            scaling_factor=0.41407,
+        )
+
+        torch.manual_seed(0)
+        scheduler = FlowMatchEulerDiscreteScheduler(shift=7.0)
+
+        torch.manual_seed(0)
+        text_encoder_config = Gemma2Config(
+            head_dim=16,
+            hidden_size=32,
+            initializer_range=0.02,
+            intermediate_size=64,
+            max_position_embeddings=8192,
+            model_type="gemma2",
+            num_attention_heads=2,
+            num_hidden_layers=1,
+            num_key_value_heads=2,
+            vocab_size=8,
+            attn_implementation="eager",
+        )
+        text_encoder = Gemma2ForCausalLM(text_encoder_config)
+        tokenizer = GemmaTokenizer.from_pretrained("hf-internal-testing/dummy-gemma")
+
+        components = {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+        inputs = {
+            "prompt": "",
+            "negative_prompt": "",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 6.0,
+            "height": 32,
+            "width": 32,
+            "max_sequence_length": 16,
+            "output_type": "pt",
+            "complex_human_instruction": None,
+        }
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs)[0]
+        generated_image = image[0]
+
+        self.assertEqual(generated_image.shape, (3, 32, 32))
+        expected_image = torch.randn(3, 32, 32)
+        max_diff = np.abs(generated_image - expected_image).max()
+        self.assertLessEqual(max_diff, 1e10)
+
+    def test_callback_inputs(self):
+        sig = inspect.signature(self.pipeline_class.__call__)
+        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
+        has_callback_step_end = "callback_on_step_end" in sig.parameters
+
+        if not (has_callback_tensor_inputs and has_callback_step_end):
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        self.assertTrue(
+            hasattr(pipe, "_callback_tensor_inputs"),
+            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
+        )
+
+        def callback_inputs_subset(pipe, i, t, callback_kwargs):
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        def callback_inputs_all(pipe, i, t, callback_kwargs):
+            for tensor_name in pipe._callback_tensor_inputs:
+                assert tensor_name in callback_kwargs
+
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        # Test passing in a subset
+        inputs["callback_on_step_end"] = callback_inputs_subset
+        inputs["callback_on_step_end_tensor_inputs"] = ["latents"]
+        output = pipe(**inputs)[0]
+
+        # Test passing in a everything
+        inputs["callback_on_step_end"] = callback_inputs_all
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+
+        def callback_inputs_change_tensor(pipe, i, t, callback_kwargs):
+            is_last = i == (pipe.num_timesteps - 1)
+            if is_last:
+                callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"])
+            return callback_kwargs
+
+        inputs["callback_on_step_end"] = callback_inputs_change_tensor
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+        assert output.abs().sum() < 1e10
+
+    def test_attention_slicing_forward_pass(
+        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
+    ):
+        if not self.test_attention_slicing:
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        generator_device = "cpu"
+        inputs = self.get_dummy_inputs(generator_device)
+        output_without_slicing = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=1)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing1 = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=2)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing2 = pipe(**inputs)[0]
+
+        if test_max_difference:
+            max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max()
+            max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max()
+            self.assertLess(
+                max(max_diff1, max_diff2),
+                expected_max_diff,
+                "Attention slicing should not affect the inference results",
+            )
+
+    # TODO(aryan): Create a dummy gemma model with smol vocab size
+    @unittest.skip(
+        "A very small vocab size is used for fast tests. So, any kind of prompt other than the empty default used in other tests will lead to a embedding lookup error. This test uses a long prompt that causes the error."
+    )
+    def test_inference_batch_consistent(self):
+        pass
+
+    @unittest.skip(
+        "A very small vocab size is used for fast tests. So, any kind of prompt other than the empty default used in other tests will lead to a embedding lookup error. This test uses a long prompt that causes the error."
+    )
+    def test_inference_batch_single_identical(self):
+        pass
+
+    def test_float16_inference(self):
+        # Requires higher tolerance as model seems very sensitive to dtype
+        super().test_float16_inference(expected_max_diff=0.08)
+
+
+@slow
+@require_torch_gpu
+class SanaPipelineIntegrationTests(unittest.TestCase):
+    prompt = "A painting of a squirrel eating a burger."
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_sana_1024(self):
+        generator = torch.Generator("cpu").manual_seed(0)
+
+        pipe = SanaPipeline.from_pretrained(
+            "Efficient-Large-Model/Sana_1600M_1024px_diffusers", torch_dtype=torch.float16
+        )
+        pipe.enable_model_cpu_offload()
+
+        image = pipe(
+            prompt=self.prompt,
+            height=1024,
+            width=1024,
+            generator=generator,
+            num_inference_steps=20,
+            output_type="np",
+        ).images[0]
+
+        image = image.flatten()
+        output_slice = np.concatenate((image[:16], image[-16:]))
+
+        # fmt: off
+        expected_slice = np.array([0.0427, 0.0789, 0.0662, 0.0464, 0.082, 0.0574, 0.0535, 0.0886, 0.0647, 0.0549, 0.0872, 0.0605, 0.0593, 0.0942, 0.0674, 0.0581, 0.0076, 0.0168, 0.0027, 0.0063, 0.0159, 0.0, 0.0071, 0.0198, 0.0034, 0.0105, 0.0212, 0.0, 0.0, 0.0166, 0.0042, 0.0125])
+        # fmt: on
+
+        self.assertTrue(np.allclose(output_slice, expected_slice, atol=1e-4))
+
+    def test_sana_512(self):
+        generator = torch.Generator("cpu").manual_seed(0)
+
+        pipe = SanaPipeline.from_pretrained(
+            "Efficient-Large-Model/Sana_1600M_512px_diffusers", torch_dtype=torch.float16
+        )
+        pipe.enable_model_cpu_offload()
+
+        image = pipe(
+            prompt=self.prompt,
+            height=512,
+            width=512,
+            generator=generator,
+            num_inference_steps=20,
+            output_type="np",
+        ).images[0]
+
+        image = image.flatten()
+        output_slice = np.concatenate((image[:16], image[-16:]))
+
+        # fmt: off
+        expected_slice = np.array([0.0803, 0.0774, 0.1108, 0.0872, 0.093, 0.1118, 0.0952, 0.0898, 0.1038, 0.0818, 0.0754, 0.0894, 0.074, 0.0691, 0.0906, 0.0671, 0.0154, 0.0254, 0.0203, 0.0178, 0.0283, 0.0193, 0.0215, 0.0273, 0.0188, 0.0212, 0.0273, 0.0151, 0.0061, 0.0244, 0.0212, 0.0259])
+        # fmt: on
+
+        self.assertTrue(np.allclose(output_slice, expected_slice, atol=1e-4))

tests/single_file/test_model_autoencoder_dc_single_file.py

@@ -0,0 +1,126 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import gc
+import unittest
+
+import torch
+
+from diffusers import (
+    AutoencoderDC,
+)
+from diffusers.utils.testing_utils import (
+    backend_empty_cache,
+    enable_full_determinism,
+    load_hf_numpy,
+    numpy_cosine_similarity_distance,
+    require_torch_accelerator,
+    slow,
+    torch_device,
+)
+
+
+enable_full_determinism()
+
+
+@slow
+@require_torch_accelerator
+class AutoencoderDCSingleFileTests(unittest.TestCase):
+    model_class = AutoencoderDC
+    ckpt_path = "https://huggingface.co/mit-han-lab/dc-ae-f32c32-sana-1.0/blob/main/model.safetensors"
+    repo_id = "mit-han-lab/dc-ae-f32c32-sana-1.0-diffusers"
+    main_input_name = "sample"
+    base_precision = 1e-2
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    def get_file_format(self, seed, shape):
+        return f"gaussian_noise_s={seed}_shape={'_'.join([str(s) for s in shape])}.npy"
+
+    def get_sd_image(self, seed=0, shape=(4, 3, 512, 512), fp16=False):
+        dtype = torch.float16 if fp16 else torch.float32
+        image = torch.from_numpy(load_hf_numpy(self.get_file_format(seed, shape))).to(torch_device).to(dtype)
+        return image
+
+    def test_single_file_inference_same_as_pretrained(self):
+        model_1 = self.model_class.from_pretrained(self.repo_id).to(torch_device)
+        model_2 = self.model_class.from_single_file(self.ckpt_path, config=self.repo_id).to(torch_device)
+
+        image = self.get_sd_image(33)
+
+        with torch.no_grad():
+            sample_1 = model_1(image).sample
+            sample_2 = model_2(image).sample
+
+        assert sample_1.shape == sample_2.shape
+
+        output_slice_1 = sample_1.flatten().float().cpu()
+        output_slice_2 = sample_2.flatten().float().cpu()
+
+        assert numpy_cosine_similarity_distance(output_slice_1, output_slice_2) < 1e-4
+
+    def test_single_file_components(self):
+        model = self.model_class.from_pretrained(self.repo_id)
+        model_single_file = self.model_class.from_single_file(self.ckpt_path)
+
+        PARAMS_TO_IGNORE = ["torch_dtype", "_name_or_path", "_use_default_values", "_diffusers_version"]
+        for param_name, param_value in model_single_file.config.items():
+            if param_name in PARAMS_TO_IGNORE:
+                continue
+            assert (
+                model.config[param_name] == param_value
+            ), f"{param_name} differs between pretrained loading and single file loading"
+
+    def test_single_file_in_type_variant_components(self):
+        # `in` variant checkpoints require passing in a `config` parameter
+        # in order to set the scaling factor correctly.
+        # `in` and `mix` variants have the same keys and we cannot automatically infer a scaling factor.
+        # We default to using teh `mix` config
+        repo_id = "mit-han-lab/dc-ae-f128c512-in-1.0-diffusers"
+        ckpt_path = "https://huggingface.co/mit-han-lab/dc-ae-f128c512-in-1.0/blob/main/model.safetensors"
+
+        model = self.model_class.from_pretrained(repo_id)
+        model_single_file = self.model_class.from_single_file(ckpt_path, config=repo_id)
+
+        PARAMS_TO_IGNORE = ["torch_dtype", "_name_or_path", "_use_default_values", "_diffusers_version"]
+        for param_name, param_value in model_single_file.config.items():
+            if param_name in PARAMS_TO_IGNORE:
+                continue
+            assert (
+                model.config[param_name] == param_value
+            ), f"{param_name} differs between pretrained loading and single file loading"
+
+    def test_single_file_mix_type_variant_components(self):
+        repo_id = "mit-han-lab/dc-ae-f128c512-mix-1.0-diffusers"
+        ckpt_path = "https://huggingface.co/mit-han-lab/dc-ae-f128c512-mix-1.0/blob/main/model.safetensors"
+
+        model = self.model_class.from_pretrained(repo_id)
+        model_single_file = self.model_class.from_single_file(ckpt_path, config=repo_id)
+
+        PARAMS_TO_IGNORE = ["torch_dtype", "_name_or_path", "_use_default_values", "_diffusers_version"]
+        for param_name, param_value in model_single_file.config.items():
+            if param_name in PARAMS_TO_IGNORE:
+                continue
+            assert (
+                model.config[param_name] == param_value
+            ), f"{param_name} differs between pretrained loading and single file loading"