frame wise guidance

* add HD-Painter pipeline

* style fixing

* refactor, change doc, fix ruff

* fix docs

* used correct ruff version

---------

Co-authored-by: Hayk Manukyan <youremail@yourdomain.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

docs/source/en/tutorials/using_peft_for_inference.md

@@ -133,6 +133,62 @@ image
 
 ![no-lora](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_20_1.png)
 
+### Customize adapters strength
+For even more customization, you can control how strongly the adapter affects each part of the pipeline. For this, pass a dictionary with the control strengths (called "scales") to [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`].
+
+For example, here's how you can turn on the adapter for the `down` parts, but turn it off for the `mid` and `up` parts:
+```python
+pipe.enable_lora()  # enable lora again, after we disabled it above
+prompt = "toy_face of a hacker with a hoodie, pixel art"
+adapter_weight_scales = { "unet": { "down": 1, "mid": 0, "up": 0} }
+pipe.set_adapters("pixel", adapter_weight_scales)
+image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
+image
+```
+
+![block-lora-text-and-down](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_block_down.png)
+
+Let's see how turning off the `down` part and turning on the `mid` and `up` part respectively changes the image.
+```python
+adapter_weight_scales = { "unet": { "down": 0, "mid": 1, "up": 0} }
+pipe.set_adapters("pixel", adapter_weight_scales)
+image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
+image
+```
+
+![block-lora-text-and-mid](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_block_mid.png)
+
+```python
+adapter_weight_scales = { "unet": { "down": 0, "mid": 0, "up": 1} }
+pipe.set_adapters("pixel", adapter_weight_scales)
+image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
+image
+```
+
+![block-lora-text-and-up](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_block_up.png)
+
+Looks cool!
+
+This is a really powerful feature. You can use it to control the adapter strengths down to per-transformer level. And you can even use it for multiple adapters.
+```python
+adapter_weight_scales_toy = 0.5
+adapter_weight_scales_pixel = {
+    "unet": {
+        "down": 0.9,  # all transformers in the down-part will use scale 0.9
+        # "mid"  # because, in this example, "mid" is not given, all transformers in the mid part will use the default scale 1.0
+        "up": {
+            "block_0": 0.6,  # all 3 transformers in the 0th block in the up-part will use scale 0.6
+            "block_1": [0.4, 0.8, 1.0],  # the 3 transformers in the 1st block in the up-part will use scales 0.4, 0.8 and 1.0 respectively
+        }
+    }
+}
+pipe.set_adapters(["toy", "pixel"], [adapter_weight_scales_toy, adapter_weight_scales_pixel])
+image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
+image
+```
+
+![block-lora-mixed](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_block_mixed.png)
+
 ## Manage active adapters
 
 You have attached multiple adapters in this tutorial, and if you're feeling a bit lost on what adapters have been attached to the pipeline's components, use the [`~diffusers.loaders.LoraLoaderMixin.get_active_adapters`] method to check the list of active adapters:

docs/source/en/using-diffusers/loading_adapters.md

@@ -153,18 +153,43 @@ image
     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_attn_proc.png" />
 </div>
 
-<Tip>
-
-For both [`~loaders.LoraLoaderMixin.load_lora_weights`] and [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`], you can pass the `cross_attention_kwargs={"scale": 0.5}` parameter to adjust how much of the LoRA weights to use. A value of `0` is the same as only using the base model weights, and a value of `1` is equivalent to using the fully finetuned LoRA.
-
-</Tip>
-
 To unload the LoRA weights, use the [`~loaders.LoraLoaderMixin.unload_lora_weights`] method to discard the LoRA weights and restore the model to its original weights:
 
 ```py
 pipeline.unload_lora_weights()
 ```
 
+### Adjust LoRA weight scale
+
+For both [`~loaders.LoraLoaderMixin.load_lora_weights`] and [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`], you can pass the `cross_attention_kwargs={"scale": 0.5}` parameter to adjust how much of the LoRA weights to use. A value of `0` is the same as only using the base model weights, and a value of `1` is equivalent to using the fully finetuned LoRA.
+
+For more granular control on the amount of LoRA weights used per layer, you can use [`~loaders.LoraLoaderMixin.set_adapters`] and pass a dictionary specifying by how much to scale the weights in each layer by.
+```python
+pipe = ... # create pipeline
+pipe.load_lora_weights(..., adapter_name="my_adapter") 
+scales = {
+    "text_encoder": 0.5,
+    "text_encoder_2": 0.5,  # only usable if pipe has a 2nd text encoder
+    "unet": {
+        "down": 0.9,  # all transformers in the down-part will use scale 0.9
+        # "mid"  # in this example "mid" is not given, therefore all transformers in the mid part will use the default scale 1.0
+        "up": {
+            "block_0": 0.6,  # all 3 transformers in the 0th block in the up-part will use scale 0.6
+            "block_1": [0.4, 0.8, 1.0],  # the 3 transformers in the 1st block in the up-part will use scales 0.4, 0.8 and 1.0 respectively
+        }
+    }
+}
+pipe.set_adapters("my_adapter", scales)
+```
+
+This also works with multiple adapters - see [this guide](https://huggingface.co/docs/diffusers/tutorials/using_peft_for_inference#customize-adapters-strength) for how to do it.
+
+<Tip warning={true}>
+
+Currently, [`~loaders.LoraLoaderMixin.set_adapters`] only supports scaling attention weights. If a LoRA has other parts (e.g., resnets or down-/upsamplers), they will keep a scale of 1.0.
+
+</Tip>
+
 ### Kohya and TheLastBen
 
 Other popular LoRA trainers from the community include those by [Kohya](https://github.com/kohya-ss/sd-scripts/) and [TheLastBen](https://github.com/TheLastBen/fast-stable-diffusion). These trainers create different LoRA checkpoints than those trained by 🤗 Diffusers, but they can still be loaded in the same way.

examples/community/README.md

@@ -10,6 +10,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 
 | Example                                                                                                                               | Description                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Code Example                                                                              | Colab                                                                                                                                                                                                              |                                                        Author |
 |:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
+| HD-Painter                                                                                                                            | [HD-Painter](https://github.com/Picsart-AI-Research/HD-Painter) enables prompt-faithfull and high resolution (up to 2k) image inpainting upon any diffusion-based image inpainting method.                                                                                                                                                                                                                                                                                                               | [HD-Painter](#hd-painter)                                                                 | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/PAIR/HD-Painter)                                                                              | [Manukyan Hayk](https://github.com/haikmanukyan) and [Sargsyan Andranik](https://github.com/AndranikSargsyan) |
 | Marigold Monocular Depth Estimation                                                                                                   | A universal monocular depth estimator, utilizing Stable Diffusion, delivering sharp predictions in the wild. (See the [project page](https://marigoldmonodepth.github.io) and [full codebase](https://github.com/prs-eth/marigold) for more details.)                                                                                                                                                                                                                                                        | [Marigold Depth Estimation](#marigold-depth-estimation)                                   | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/toshas/marigold) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/12G8reD13DdpMie5ZQlaFNo2WCGeNUH-u?usp=sharing) | [Bingxin Ke](https://github.com/markkua) and [Anton Obukhov](https://github.com/toshas) |
 | LLM-grounded Diffusion (LMD+)                                                                                                         | LMD greatly improves the prompt following ability of text-to-image generation models by introducing an LLM as a front-end prompt parser and layout planner. [Project page.](https://llm-grounded-diffusion.github.io/) [See our full codebase (also with diffusers).](https://github.com/TonyLianLong/LLM-groundedDiffusion)                                                                                                                                                                                                                                                                                                                                                                                                                                   | [LLM-grounded Diffusion (LMD+)](#llm-grounded-diffusion)                             | [Huggingface Demo](https://huggingface.co/spaces/longlian/llm-grounded-diffusion) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1SXzMSeAB-LJYISb2yrUOdypLz4OYWUKj) |                [Long (Tony) Lian](https://tonylian.com/) |
 | CLIP Guided Stable Diffusion                                                                                                          | Doing CLIP guidance for text to image generation with Stable Diffusion                                                                                                                                                                                                                                                                                                                                                                                                                                   | [CLIP Guided Stable Diffusion](#clip-guided-stable-diffusion)                             | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/CLIP_Guided_Stable_diffusion_with_diffusers.ipynb) |                [Suraj Patil](https://github.com/patil-suraj/) |
@@ -75,6 +76,48 @@ pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custo
 
 ## Example usages
 
+### HD-Painter
+
+Implementation of [HD-Painter: High-Resolution and Prompt-Faithful Text-Guided Image Inpainting with Diffusion Models](https://arxiv.org/abs/2312.14091).
+
+![teaser-img](https://raw.githubusercontent.com/Picsart-AI-Research/HD-Painter/main/__assets__/github/teaser.jpg)
+
+The abstract from the paper is:
+
+Recent progress in text-guided image inpainting, based on the unprecedented success of text-to-image diffusion models, has led to exceptionally realistic and visually plausible results.
+However, there is still significant potential for improvement in current text-to-image inpainting models, particularly in better aligning the inpainted area with user prompts and performing high-resolution inpainting.
+Therefore, in this paper we introduce _HD-Painter_, a completely **training-free** approach that **accurately follows to prompts** and coherently **scales to high-resolution** image inpainting.
+To this end, we design the _Prompt-Aware Introverted Attention (PAIntA)_ layer enhancing self-attention scores by prompt information and resulting in better text alignment generations.
+To further improve the prompt coherence we introduce the _Reweighting Attention Score Guidance (RASG)_ mechanism seamlessly integrating a post-hoc sampling strategy into general form of DDIM to prevent out-of-distribution latent shifts.
+Moreover, HD-Painter allows extension to larger scales by introducing a specialized super-resolution technique customized for inpainting, enabling the completion of missing regions in images of up to 2K resolution. 
+Our experiments demonstrate that HD-Painter surpasses existing state-of-the-art approaches qualitatively and quantitatively, achieving an impressive generation accuracy improvement of **61.4** vs **51.9**. 
+We will make the codes publicly available.
+
+You can find additional information about Text2Video-Zero in the [paper](https://arxiv.org/abs/2312.14091) or the [original codebase](https://github.com/Picsart-AI-Research/HD-Painter).
+
+#### Usage example
+
+```python
+import torch
+from diffusers import DiffusionPipeline, DDIMScheduler
+from diffusers.utils import load_image, make_image_grid
+
+pipe = DiffusionPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-2-inpainting",
+    custom_pipeline="hd_painter"
+)
+pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+
+prompt = "wooden boat"
+init_image = load_image("https://raw.githubusercontent.com/Picsart-AI-Research/HD-Painter/main/__assets__/samples/images/2.jpg")
+mask_image = load_image("https://raw.githubusercontent.com/Picsart-AI-Research/HD-Painter/main/__assets__/samples/masks/2.png")
+
+image = pipe (prompt, init_image, mask_image, use_rasg = True, use_painta = True, generator=torch.manual_seed(12345)).images[0]
+
+make_image_grid([init_image, mask_image, image], rows=1, cols=3)
+
+```
+
 ### Marigold Depth Estimation
 
 Marigold is a universal monocular depth estimator that delivers accurate and sharp predictions in the wild. Based on Stable Diffusion, it is trained exclusively with synthetic depth data and excels in zero-shot adaptation to real-world imagery. This pipeline is an official implementation of the inference process. More details can be found on our [project page](https://marigoldmonodepth.github.io) and [full codebase](https://github.com/prs-eth/marigold) (also implemented with diffusers).
@@ -85,14 +128,25 @@ This depth estimation pipeline processes a single input image through multiple d
 
 ```python
 import numpy as np
+import torch
 from PIL import Image
 from diffusers import DiffusionPipeline
 from diffusers.utils import load_image
 
+# Original DDIM version (higher quality)
 pipe = DiffusionPipeline.from_pretrained(
-    "Bingxin/Marigold",
+    "prs-eth/marigold-v1-0",
     custom_pipeline="marigold_depth_estimation"
     # torch_dtype=torch.float16,                # (optional) Run with half-precision (16-bit float).
+    # variant="fp16",                           # (optional) Use with `torch_dtype=torch.float16`, to directly load fp16 checkpoint
+)
+
+# (New) LCM version (faster speed)
+pipe = DiffusionPipeline.from_pretrained(
+    "prs-eth/marigold-lcm-v1-0",
+    custom_pipeline="marigold_depth_estimation"
+    # torch_dtype=torch.float16,                # (optional) Run with half-precision (16-bit float).
+    # variant="fp16",                           # (optional) Use with `torch_dtype=torch.float16`, to directly load fp16 checkpoint
 )
 
 pipe.to("cuda")
@@ -101,12 +155,21 @@ img_path_or_url = "https://share.phys.ethz.ch/~pf/bingkedata/marigold/pipeline_e
 image: Image.Image = load_image(img_path_or_url)
 
 pipeline_output = pipe(
-    image,                  # Input image.
+    image,                    # Input image.
+    # ----- recommended setting for DDIM version -----
     # denoising_steps=10,     # (optional) Number of denoising steps of each inference pass. Default: 10.
     # ensemble_size=10,       # (optional) Number of inference passes in the ensemble. Default: 10.
+    # ------------------------------------------------
+    
+    # ----- recommended setting for LCM version ------
+    # denoising_steps=4,
+    # ensemble_size=5,
+    # -------------------------------------------------
+    
     # processing_res=768,     # (optional) Maximum resolution of processing. If set to 0: will not resize at all. Defaults to 768.
     # match_input_res=True,   # (optional) Resize depth prediction to match input resolution.
     # batch_size=0,           # (optional) Inference batch size, no bigger than `num_ensemble`. If set to 0, the script will automatically decide the proper batch size. Defaults to 0.
+    # seed=2024,              # (optional) Random seed can be set to ensure additional reproducibility. Default: None (unseeded). Note: forcing --batch_size 1 helps to increase reproducibility. To ensure full reproducibility, deterministic mode needs to be used.
     # color_map="Spectral",   # (optional) Colormap used to colorize the depth map. Defaults to "Spectral". Set to `None` to skip colormap generation.
     # show_progress_bar=True, # (optional) If true, will show progress bars of the inference progress.
 )
@@ -3743,3 +3806,80 @@ onestep_image = pipe(prompt, num_inference_steps=1).images[0]
 # Multistep sampling
 multistep_image = pipe(prompt, num_inference_steps=4).images[0]
 ```
+
+# Perturbed-Attention Guidance
+
+[Project](https://ku-cvlab.github.io/Perturbed-Attention-Guidance/) / [arXiv](https://arxiv.org/abs/2403.17377) / [GitHub](https://github.com/KU-CVLAB/Perturbed-Attention-Guidance)
+
+This implementation is based on [Diffusers](https://huggingface.co/docs/diffusers/index). StableDiffusionPAGPipeline is a modification of StableDiffusionPipeline to support Perturbed-Attention Guidance (PAG).
+
+## Example Usage
+
+```
+import os
+import torch
+
+from accelerate.utils import set_seed
+
+from diffusers import StableDiffusionPipeline
+from diffusers.utils import load_image, make_image_grid
+from diffusers.utils.torch_utils import randn_tensor
+
+pipe = StableDiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    custom_pipeline="hyoungwoncho/sd_perturbed_attention_guidance",
+    torch_dtype=torch.float16
+)
+
+device="cuda"
+pipe = pipe.to(device)
+
+pag_scale = 5.0
+pag_applied_layers_index = ['m0']
+
+batch_size = 4
+seed=10
+
+base_dir = "./results/"
+grid_dir = base_dir + "/pag" + str(pag_scale) + "/"
+
+if not os.path.exists(grid_dir):
+    os.makedirs(grid_dir)
+
+set_seed(seed)
+
+latent_input = randn_tensor(shape=(batch_size,4,64,64),generator=None, device=device, dtype=torch.float16)
+
+output_baseline = pipe(
+    "",
+    width=512,
+    height=512,
+    num_inference_steps=50,
+    guidance_scale=0.0,
+    pag_scale=0.0,
+    pag_applied_layers_index=pag_applied_layers_index,
+    num_images_per_prompt=batch_size,
+    latents=latent_input
+).images
+
+output_pag = pipe(
+    "",
+    width=512,
+    height=512,
+    num_inference_steps=50,
+    guidance_scale=0.0,
+    pag_scale=5.0,
+    pag_applied_layers_index=pag_applied_layers_index,
+    num_images_per_prompt=batch_size,
+    latents=latent_input
+).images
+
+grid_image = make_image_grid(output_baseline + output_pag, rows=2, cols=batch_size)
+grid_image.save(grid_dir + "sample.png")
+```
+
+## PAG Parameters
+
+pag_scale : gudiance scale of PAG (ex: 5.0)
+
+pag_applied_layers_index : index of the layer to apply perturbation (ex: ['m0'])

examples/community/hd_painter.py

@@ -0,0 +1,994 @@
+import math
+import numbers
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.image_processor import PipelineImageInput
+from diffusers.models import AsymmetricAutoencoderKL, ImageProjection
+from diffusers.models.attention_processor import Attention, AttnProcessor
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint import (
+    StableDiffusionInpaintPipeline,
+    retrieve_timesteps,
+)
+from diffusers.utils import deprecate
+
+
+class RASGAttnProcessor:
+    def __init__(self, mask, token_idx, scale_factor):
+        self.attention_scores = None  # Stores the last output of the similarity matrix here. Each layer will get its own RASGAttnProcessor assigned
+        self.mask = mask
+        self.token_idx = token_idx
+        self.scale_factor = scale_factor
+        self.mask_resoltuion = mask.shape[-1] * mask.shape[-2]  # 64 x 64 if the image is 512x512
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        temb: Optional[torch.FloatTensor] = None,
+        scale: float = 1.0,
+    ) -> torch.Tensor:
+        # Same as the default AttnProcessor up untill the part where similarity matrix gets saved
+        downscale_factor = self.mask_resoltuion // hidden_states.shape[1]
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        # Automatically recognize the resolution and save the attention similarity values
+        # We need to use the values before the softmax function, hence the rewritten get_attention_scores function.
+        if downscale_factor == self.scale_factor**2:
+            self.attention_scores = get_attention_scores(attn, query, key, attention_mask)
+            attention_probs = self.attention_scores.softmax(dim=-1)
+            attention_probs = attention_probs.to(query.dtype)
+        else:
+            attention_probs = attn.get_attention_scores(query, key, attention_mask)  # Original code
+
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class PAIntAAttnProcessor:
+    def __init__(self, transformer_block, mask, token_idx, do_classifier_free_guidance, scale_factors):
+        self.transformer_block = transformer_block  # Stores the parent transformer block.
+        self.mask = mask
+        self.scale_factors = scale_factors
+        self.do_classifier_free_guidance = do_classifier_free_guidance
+        self.token_idx = token_idx
+        self.shape = mask.shape[2:]
+        self.mask_resoltuion = mask.shape[-1] * mask.shape[-2]  # 64 x 64
+        self.default_processor = AttnProcessor()
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        temb: Optional[torch.FloatTensor] = None,
+        scale: float = 1.0,
+    ) -> torch.Tensor:
+        # Automatically recognize the resolution of the current attention layer and resize the masks accordingly
+        downscale_factor = self.mask_resoltuion // hidden_states.shape[1]
+
+        mask = None
+        for factor in self.scale_factors:
+            if downscale_factor == factor**2:
+                shape = (self.shape[0] // factor, self.shape[1] // factor)
+                mask = F.interpolate(self.mask, shape, mode="bicubic")  # B, 1, H, W
+                break
+        if mask is None:
+            return self.default_processor(attn, hidden_states, encoder_hidden_states, attention_mask, temb, scale)
+
+        # STARTS HERE
+        residual = hidden_states
+        # Save the input hidden_states for later use
+        input_hidden_states = hidden_states
+
+        # ================================================== #
+        # =============== SELF ATTENTION 1 ================= #
+        # ================================================== #
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        # self_attention_probs = attn.get_attention_scores(query, key, attention_mask) # We can't use post-softmax attention scores in this case
+        self_attention_scores = get_attention_scores(
+            attn, query, key, attention_mask
+        )  # The custom function returns pre-softmax probabilities
+        self_attention_probs = self_attention_scores.softmax(
+            dim=-1
+        )  # Manually compute the probabilities here, the scores will be reused in the second part of PAIntA
+        self_attention_probs = self_attention_probs.to(query.dtype)
+
+        hidden_states = torch.bmm(self_attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        # x = x + self.attn1(self.norm1(x))
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:  # So many residuals everywhere
+            hidden_states = hidden_states + residual
+
+        self_attention_output_hidden_states = hidden_states / attn.rescale_output_factor
+
+        # ================================================== #
+        # ============ BasicTransformerBlock =============== #
+        # ================================================== #
+        # We use a hack by running the code from the BasicTransformerBlock that is between Self and Cross attentions here
+        # The other option would've been modifying the BasicTransformerBlock and adding this functionality here.
+        # I assumed that changing the BasicTransformerBlock would have been a bigger deal and decided to use this hack isntead.
+
+        # The SelfAttention block recieves the normalized latents from the BasicTransformerBlock,
+        # But the residual of the output is the non-normalized version.
+        # Therefore we unnormalize the input hidden state here
+        unnormalized_input_hidden_states = (
+            input_hidden_states + self.transformer_block.norm1.bias
+        ) * self.transformer_block.norm1.weight
+
+        # TODO: return if neccessary
+        # if self.use_ada_layer_norm_zero:
+        #     attn_output = gate_msa.unsqueeze(1) * attn_output
+        # elif self.use_ada_layer_norm_single:
+        #     attn_output = gate_msa * attn_output
+
+        transformer_hidden_states = self_attention_output_hidden_states + unnormalized_input_hidden_states
+        if transformer_hidden_states.ndim == 4:
+            transformer_hidden_states = transformer_hidden_states.squeeze(1)
+
+        # TODO: return if neccessary
+        # 2.5 GLIGEN Control
+        # if gligen_kwargs is not None:
+        #     transformer_hidden_states = self.fuser(transformer_hidden_states, gligen_kwargs["objs"])
+        # NOTE: we experimented with using GLIGEN and HDPainter together, the results were not that great
+
+        # 3. Cross-Attention
+        if self.transformer_block.use_ada_layer_norm:
+            # transformer_norm_hidden_states = self.transformer_block.norm2(transformer_hidden_states, timestep)
+            raise NotImplementedError()
+        elif self.transformer_block.use_ada_layer_norm_zero or self.transformer_block.use_layer_norm:
+            transformer_norm_hidden_states = self.transformer_block.norm2(transformer_hidden_states)
+        elif self.transformer_block.use_ada_layer_norm_single:
+            # For PixArt norm2 isn't applied here:
+            # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+            transformer_norm_hidden_states = transformer_hidden_states
+        elif self.transformer_block.use_ada_layer_norm_continuous:
+            # transformer_norm_hidden_states = self.transformer_block.norm2(transformer_hidden_states, added_cond_kwargs["pooled_text_emb"])
+            raise NotImplementedError()
+        else:
+            raise ValueError("Incorrect norm")
+
+        if self.transformer_block.pos_embed is not None and self.transformer_block.use_ada_layer_norm_single is False:
+            transformer_norm_hidden_states = self.transformer_block.pos_embed(transformer_norm_hidden_states)
+
+        # ================================================== #
+        # ================= CROSS ATTENTION ================ #
+        # ================================================== #
+
+        # We do an initial pass of the CrossAttention up to obtaining the similarity matrix here.
+        # The similarity matrix is used to obtain scaling coefficients for the attention matrix of the self attention
+        # We reuse the previously computed self-attention matrix, and only repeat the steps after the softmax
+
+        cross_attention_input_hidden_states = (
+            transformer_norm_hidden_states  # Renaming the variable for the sake of readability
+        )
+
+        # TODO: check if classifier_free_guidance is being used before splitting here
+        if self.do_classifier_free_guidance:
+            # Our scaling coefficients depend only on the conditional part, so we split the inputs
+            (
+                _cross_attention_input_hidden_states_unconditional,
+                cross_attention_input_hidden_states_conditional,
+            ) = cross_attention_input_hidden_states.chunk(2)
+
+            # Same split for the encoder_hidden_states i.e. the tokens
+            # Since the SelfAttention processors don't get the encoder states as input, we inject them into the processor in the begining.
+            _encoder_hidden_states_unconditional, encoder_hidden_states_conditional = self.encoder_hidden_states.chunk(
+                2
+            )
+        else:
+            cross_attention_input_hidden_states_conditional = cross_attention_input_hidden_states
+            encoder_hidden_states_conditional = self.encoder_hidden_states.chunk(2)
+
+        # Rename the variables for the sake of readability
+        # The part below is the beginning of the __call__ function of the following CrossAttention layer
+        cross_attention_hidden_states = cross_attention_input_hidden_states_conditional
+        cross_attention_encoder_hidden_states = encoder_hidden_states_conditional
+
+        attn2 = self.transformer_block.attn2
+
+        if attn2.spatial_norm is not None:
+            cross_attention_hidden_states = attn2.spatial_norm(cross_attention_hidden_states, temb)
+
+        input_ndim = cross_attention_hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = cross_attention_hidden_states.shape
+            cross_attention_hidden_states = cross_attention_hidden_states.view(
+                batch_size, channel, height * width
+            ).transpose(1, 2)
+
+        (
+            batch_size,
+            sequence_length,
+            _,
+        ) = cross_attention_hidden_states.shape  # It is definitely a cross attention, so no need for an if block
+        # TODO: change the attention_mask here
+        attention_mask = attn2.prepare_attention_mask(
+            None, sequence_length, batch_size
+        )  # I assume the attention mask is the same...
+
+        if attn2.group_norm is not None:
+            cross_attention_hidden_states = attn2.group_norm(cross_attention_hidden_states.transpose(1, 2)).transpose(
+                1, 2
+            )
+
+        query2 = attn2.to_q(cross_attention_hidden_states)
+
+        if attn2.norm_cross:
+            cross_attention_encoder_hidden_states = attn2.norm_encoder_hidden_states(
+                cross_attention_encoder_hidden_states
+            )
+
+        key2 = attn2.to_k(cross_attention_encoder_hidden_states)
+        query2 = attn2.head_to_batch_dim(query2)
+        key2 = attn2.head_to_batch_dim(key2)
+
+        cross_attention_probs = attn2.get_attention_scores(query2, key2, attention_mask)
+
+        # CrossAttention ends here, the remaining part is not used
+
+        # ================================================== #
+        # ================ SELF ATTENTION 2 ================ #
+        # ================================================== #
+        # DEJA VU!
+
+        mask = (mask > 0.5).to(self_attention_output_hidden_states.dtype)
+        m = mask.to(self_attention_output_hidden_states.device)
+        # m = rearrange(m, 'b c h w -> b (h w) c').contiguous()
+        m = m.permute(0, 2, 3, 1).reshape((m.shape[0], -1, m.shape[1])).contiguous()  # B HW 1
+        m = torch.matmul(m, m.permute(0, 2, 1)) + (1 - m)
+
+        # # Compute scaling coefficients for the similarity matrix
+        # # Select the cross attention values for the correct tokens only!
+        # cross_attention_probs = cross_attention_probs.mean(dim = 0)
+        # cross_attention_probs = cross_attention_probs[:, self.token_idx].sum(dim=1)
+
+        # cross_attention_probs = cross_attention_probs.reshape(shape)
+        # gaussian_smoothing = GaussianSmoothing(channels=1, kernel_size=3, sigma=0.5, dim=2).to(self_attention_output_hidden_states.device)
+        # cross_attention_probs = gaussian_smoothing(cross_attention_probs.unsqueeze(0))[0] # optional smoothing
+        # cross_attention_probs = cross_attention_probs.reshape(-1)
+        # cross_attention_probs = ((cross_attention_probs - torch.median(cross_attention_probs.ravel())) / torch.max(cross_attention_probs.ravel())).clip(0, 1)
+
+        # c = (1 - m) * cross_attention_probs.reshape(1, 1, -1) + m # PAIntA scaling coefficients
+
+        # Compute scaling coefficients for the similarity matrix
+        # Select the cross attention values for the correct tokens only!
+
+        batch_size, dims, channels = cross_attention_probs.shape
+        batch_size = batch_size // attn.heads
+        cross_attention_probs = cross_attention_probs.reshape((batch_size, attn.heads, dims, channels))  # B, D, HW, T
+
+        cross_attention_probs = cross_attention_probs.mean(dim=1)  # B, HW, T
+        cross_attention_probs = cross_attention_probs[..., self.token_idx].sum(dim=-1)  # B, HW
+        cross_attention_probs = cross_attention_probs.reshape((batch_size,) + shape)  # , B, H, W
+
+        gaussian_smoothing = GaussianSmoothing(channels=1, kernel_size=3, sigma=0.5, dim=2).to(
+            self_attention_output_hidden_states.device
+        )
+        cross_attention_probs = gaussian_smoothing(cross_attention_probs[:, None])[:, 0]  # optional smoothing B, H, W
+
+        # Median normalization
+        cross_attention_probs = cross_attention_probs.reshape(batch_size, -1)  # B, HW
+        cross_attention_probs = (
+            cross_attention_probs - cross_attention_probs.median(dim=-1, keepdim=True).values
+        ) / cross_attention_probs.max(dim=-1, keepdim=True).values
+        cross_attention_probs = cross_attention_probs.clip(0, 1)
+
+        c = (1 - m) * cross_attention_probs.reshape(batch_size, 1, -1) + m
+        c = c.repeat_interleave(attn.heads, 0)  # BD, HW
+        if self.do_classifier_free_guidance:
+            c = torch.cat([c, c])  # 2BD, HW
+
+        # Rescaling the original self-attention matrix
+        self_attention_scores_rescaled = self_attention_scores * c
+        self_attention_probs_rescaled = self_attention_scores_rescaled.softmax(dim=-1)
+
+        # Continuing the self attention normally using the new matrix
+        hidden_states = torch.bmm(self_attention_probs_rescaled, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + input_hidden_states
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class StableDiffusionHDPainterPipeline(StableDiffusionInpaintPipeline):
+    def get_tokenized_prompt(self, prompt):
+        out = self.tokenizer(prompt)
+        return [self.tokenizer.decode(x) for x in out["input_ids"]]
+
+    def init_attn_processors(
+        self,
+        mask,
+        token_idx,
+        use_painta=True,
+        use_rasg=True,
+        painta_scale_factors=[2, 4],  # 64x64 -> [16x16, 32x32]
+        rasg_scale_factor=4,  # 64x64 -> 16x16
+        self_attention_layer_name="attn1",
+        cross_attention_layer_name="attn2",
+        list_of_painta_layer_names=None,
+        list_of_rasg_layer_names=None,
+    ):
+        default_processor = AttnProcessor()
+        width, height = mask.shape[-2:]
+        width, height = width // self.vae_scale_factor, height // self.vae_scale_factor
+
+        painta_scale_factors = [x * self.vae_scale_factor for x in painta_scale_factors]
+        rasg_scale_factor = self.vae_scale_factor * rasg_scale_factor
+
+        attn_processors = {}
+        for x in self.unet.attn_processors:
+            if (list_of_painta_layer_names is None and self_attention_layer_name in x) or (
+                list_of_painta_layer_names is not None and x in list_of_painta_layer_names
+            ):
+                if use_painta:
+                    transformer_block = self.unet.get_submodule(x.replace(".attn1.processor", ""))
+                    attn_processors[x] = PAIntAAttnProcessor(
+                        transformer_block, mask, token_idx, self.do_classifier_free_guidance, painta_scale_factors
+                    )
+                else:
+                    attn_processors[x] = default_processor
+            elif (list_of_rasg_layer_names is None and cross_attention_layer_name in x) or (
+                list_of_rasg_layer_names is not None and x in list_of_rasg_layer_names
+            ):
+                if use_rasg:
+                    attn_processors[x] = RASGAttnProcessor(mask, token_idx, rasg_scale_factor)
+                else:
+                    attn_processors[x] = default_processor
+
+        self.unet.set_attn_processor(attn_processors)
+        # import json
+        # with open('/home/hayk.manukyan/repos/diffusers/debug.txt', 'a')  as f:
+        #     json.dump({x:str(y) for x,y in self.unet.attn_processors.items()}, f, indent=4)
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        image: PipelineImageInput = None,
+        mask_image: PipelineImageInput = None,
+        masked_image_latents: torch.FloatTensor = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        padding_mask_crop: Optional[int] = None,
+        strength: float = 1.0,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        guidance_scale: float = 7.5,
+        positive_prompt: Optional[str] = "",
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.01,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        clip_skip: int = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        use_painta=True,
+        use_rasg=True,
+        self_attention_layer_name=".attn1",
+        cross_attention_layer_name=".attn2",
+        painta_scale_factors=[2, 4],  # 16 x 16 and 32 x 32
+        rasg_scale_factor=4,  # 16x16 by default
+        list_of_painta_layer_names=None,
+        list_of_rasg_layer_names=None,
+        **kwargs,
+    ):
+        callback = kwargs.pop("callback", None)
+        callback_steps = kwargs.pop("callback_steps", None)
+
+        if callback is not None:
+            deprecate(
+                "callback",
+                "1.0.0",
+                "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+        if callback_steps is not None:
+            deprecate(
+                "callback_steps",
+                "1.0.0",
+                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        #
+        prompt_no_positives = prompt
+        if isinstance(prompt, list):
+            prompt = [x + positive_prompt for x in prompt]
+        else:
+            prompt = prompt + positive_prompt
+
+        # 1. Check inputs
+        self.check_inputs(
+            prompt,
+            image,
+            mask_image,
+            height,
+            width,
+            strength,
+            callback_steps,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+            padding_mask_crop,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._clip_skip = clip_skip
+        self._cross_attention_kwargs = cross_attention_kwargs
+        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]
+
+        # assert batch_size == 1, "Does not work with batch size > 1 currently"
+
+        device = self._execution_device
+
+        # 3. Encode input prompt
+        text_encoder_lora_scale = (
+            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+        )
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            device,
+            num_images_per_prompt,
+            self.do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+            clip_skip=self.clip_skip,
+        )
+        # For classifier free guidance, we need to do two forward passes.
+        # Here we concatenate the unconditional and text embeddings into a single batch
+        # to avoid doing two forward passes
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+        if ip_adapter_image is not None:
+            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+            image_embeds, negative_image_embeds = self.encode_image(
+                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+            )
+            if self.do_classifier_free_guidance:
+                image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
+        # 4. set timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+        timesteps, num_inference_steps = self.get_timesteps(
+            num_inference_steps=num_inference_steps, strength=strength, device=device
+        )
+        # check that number of inference steps is not < 1 - as this doesn't make sense
+        if num_inference_steps < 1:
+            raise ValueError(
+                f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline"
+                f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
+            )
+        # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
+        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+        # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
+        is_strength_max = strength == 1.0
+
+        # 5. Preprocess mask and image
+
+        if padding_mask_crop is not None:
+            crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop)
+            resize_mode = "fill"
+        else:
+            crops_coords = None
+            resize_mode = "default"
+
+        original_image = image
+        init_image = self.image_processor.preprocess(
+            image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode
+        )
+        init_image = init_image.to(dtype=torch.float32)
+
+        # 6. Prepare latent variables
+        num_channels_latents = self.vae.config.latent_channels
+        num_channels_unet = self.unet.config.in_channels
+        return_image_latents = num_channels_unet == 4
+
+        latents_outputs = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+            image=init_image,
+            timestep=latent_timestep,
+            is_strength_max=is_strength_max,
+            return_noise=True,
+            return_image_latents=return_image_latents,
+        )
+
+        if return_image_latents:
+            latents, noise, image_latents = latents_outputs
+        else:
+            latents, noise = latents_outputs
+
+        # 7. Prepare mask latent variables
+        mask_condition = self.mask_processor.preprocess(
+            mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords
+        )
+
+        if masked_image_latents is None:
+            masked_image = init_image * (mask_condition < 0.5)
+        else:
+            masked_image = masked_image_latents
+
+        mask, masked_image_latents = self.prepare_mask_latents(
+            mask_condition,
+            masked_image,
+            batch_size * num_images_per_prompt,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            self.do_classifier_free_guidance,
+        )
+
+        # 7.5 Setting up HD-Painter
+
+        # Get the indices of the tokens to be modified by both RASG and PAIntA
+        token_idx = list(range(1, self.get_tokenized_prompt(prompt_no_positives).index("<|endoftext|>"))) + [
+            self.get_tokenized_prompt(prompt).index("<|endoftext|>")
+        ]
+
+        # Setting up the attention processors
+        self.init_attn_processors(
+            mask_condition,
+            token_idx,
+            use_painta,
+            use_rasg,
+            painta_scale_factors=painta_scale_factors,
+            rasg_scale_factor=rasg_scale_factor,
+            self_attention_layer_name=self_attention_layer_name,
+            cross_attention_layer_name=cross_attention_layer_name,
+            list_of_painta_layer_names=list_of_painta_layer_names,
+            list_of_rasg_layer_names=list_of_rasg_layer_names,
+        )
+
+        # 8. Check that sizes of mask, masked image and latents match
+        if num_channels_unet == 9:
+            # default case for runwayml/stable-diffusion-inpainting
+            num_channels_mask = mask.shape[1]
+            num_channels_masked_image = masked_image_latents.shape[1]
+            if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
+                raise ValueError(
+                    f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+                    f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+                    f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
+                    f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+                    " `pipeline.unet` or your `mask_image` or `image` input."
+                )
+        elif num_channels_unet != 4:
+            raise ValueError(
+                f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}."
+            )
+
+        # 9. 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)
+
+        if use_rasg:
+            extra_step_kwargs["generator"] = None
+
+        # 9.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
+        # 9.2 Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
+        # 10. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+        painta_active = True
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                if t < 500 and painta_active:
+                    self.init_attn_processors(
+                        mask_condition,
+                        token_idx,
+                        False,
+                        use_rasg,
+                        painta_scale_factors=painta_scale_factors,
+                        rasg_scale_factor=rasg_scale_factor,
+                        self_attention_layer_name=self_attention_layer_name,
+                        cross_attention_layer_name=cross_attention_layer_name,
+                        list_of_painta_layer_names=list_of_painta_layer_names,
+                        list_of_rasg_layer_names=list_of_rasg_layer_names,
+                    )
+                    painta_active = False
+
+                with torch.enable_grad():
+                    self.unet.zero_grad()
+                    latents = latents.detach()
+                    latents.requires_grad = True
+
+                    # expand the latents if we are doing classifier free guidance
+                    latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+                    # concat latents, mask, masked_image_latents in the channel dimension
+                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                    if num_channels_unet == 9:
+                        latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
+
+                    self.scheduler.latents = latents
+                    self.encoder_hidden_states = prompt_embeds
+                    for attn_processor in self.unet.attn_processors.values():
+                        attn_processor.encoder_hidden_states = prompt_embeds
+
+                    # predict the noise residual
+                    noise_pred = self.unet(
+                        latent_model_input,
+                        t,
+                        encoder_hidden_states=prompt_embeds,
+                        timestep_cond=timestep_cond,
+                        cross_attention_kwargs=self.cross_attention_kwargs,
+                        added_cond_kwargs=added_cond_kwargs,
+                        return_dict=False,
+                    )[0]
+
+                    # perform guidance
+                    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)
+
+                    if use_rasg:
+                        # Perform RASG
+                        _, _, height, width = mask_condition.shape  # 512 x 512
+                        scale_factor = self.vae_scale_factor * rasg_scale_factor  # 8 * 4 = 32
+
+                        # TODO: Fix for > 1 batch_size
+                        rasg_mask = F.interpolate(
+                            mask_condition, (height // scale_factor, width // scale_factor), mode="bicubic"
+                        )[0, 0]  # mode is nearest by default, B, H, W
+
+                        # Aggregate the saved attention maps
+                        attn_map = []
+                        for processor in self.unet.attn_processors.values():
+                            if hasattr(processor, "attention_scores") and processor.attention_scores is not None:
+                                if self.do_classifier_free_guidance:
+                                    attn_map.append(processor.attention_scores.chunk(2)[1])  # (B/2) x H, 256, 77
+                                else:
+                                    attn_map.append(processor.attention_scores)  # B x H, 256, 77 ?
+
+                        attn_map = (
+                            torch.cat(attn_map)
+                            .mean(0)
+                            .permute(1, 0)
+                            .reshape((-1, height // scale_factor, width // scale_factor))
+                        )  # 77, 16, 16
+
+                        # Compute the attention score
+                        attn_score = -sum(
+                            [
+                                F.binary_cross_entropy_with_logits(x - 1.0, rasg_mask.to(device))
+                                for x in attn_map[token_idx]
+                            ]
+                        )
+
+                        # Backward the score and compute the gradients
+                        attn_score.backward()
+
+                        # Normalzie the gradients and compute the noise component
+                        variance_noise = latents.grad.detach()
+                        # print("VARIANCE SHAPE", variance_noise.shape)
+                        variance_noise -= torch.mean(variance_noise, [1, 2, 3], keepdim=True)
+                        variance_noise /= torch.std(variance_noise, [1, 2, 3], keepdim=True)
+                    else:
+                        variance_noise = None
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(
+                    noise_pred, t, latents, **extra_step_kwargs, return_dict=False, variance_noise=variance_noise
+                )[0]
+
+                if num_channels_unet == 4:
+                    init_latents_proper = image_latents
+                    if self.do_classifier_free_guidance:
+                        init_mask, _ = mask.chunk(2)
+                    else:
+                        init_mask = mask
+
+                    if i < len(timesteps) - 1:
+                        noise_timestep = timesteps[i + 1]
+                        init_latents_proper = self.scheduler.add_noise(
+                            init_latents_proper, noise, torch.tensor([noise_timestep])
+                        )
+
+                    latents = (1 - init_mask) * init_latents_proper + init_mask * latents
+
+                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)
+                    mask = callback_outputs.pop("mask", mask)
+                    masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents)
+
+                # 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 callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+        if not output_type == "latent":
+            condition_kwargs = {}
+            if isinstance(self.vae, AsymmetricAutoencoderKL):
+                init_image = init_image.to(device=device, dtype=masked_image_latents.dtype)
+                init_image_condition = init_image.clone()
+                init_image = self._encode_vae_image(init_image, generator=generator)
+                mask_condition = mask_condition.to(device=device, dtype=masked_image_latents.dtype)
+                condition_kwargs = {"image": init_image_condition, "mask": mask_condition}
+            image = self.vae.decode(
+                latents / self.vae.config.scaling_factor, return_dict=False, generator=generator, **condition_kwargs
+            )[0]
+            image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+        else:
+            image = latents
+            has_nsfw_concept = None
+
+        if has_nsfw_concept is None:
+            do_denormalize = [True] * image.shape[0]
+        else:
+            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+        if padding_mask_crop is not None:
+            image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image]
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+
+
+# ============= Utility Functions ============== #
+
+
+class GaussianSmoothing(nn.Module):
+    """
+    Apply gaussian smoothing on a
+    1d, 2d or 3d tensor. Filtering is performed seperately for each channel
+    in the input using a depthwise convolution.
+    Arguments:
+        channels (int, sequence): Number of channels of the input tensors. Output will
+            have this number of channels as well.
+        kernel_size (int, sequence): Size of the gaussian kernel.
+        sigma (float, sequence): Standard deviation of the gaussian kernel.
+        dim (int, optional): The number of dimensions of the data.
+            Default value is 2 (spatial).
+    """
+
+    def __init__(self, channels, kernel_size, sigma, dim=2):
+        super(GaussianSmoothing, self).__init__()
+        if isinstance(kernel_size, numbers.Number):
+            kernel_size = [kernel_size] * dim
+        if isinstance(sigma, numbers.Number):
+            sigma = [sigma] * dim
+
+        # The gaussian kernel is the product of the
+        # gaussian function of each dimension.
+        kernel = 1
+        meshgrids = torch.meshgrid([torch.arange(size, dtype=torch.float32) for size in kernel_size])
+        for size, std, mgrid in zip(kernel_size, sigma, meshgrids):
+            mean = (size - 1) / 2
+            kernel *= 1 / (std * math.sqrt(2 * math.pi)) * torch.exp(-(((mgrid - mean) / (2 * std)) ** 2))
+
+        # Make sure sum of values in gaussian kernel equals 1.
+        kernel = kernel / torch.sum(kernel)
+
+        # Reshape to depthwise convolutional weight
+        kernel = kernel.view(1, 1, *kernel.size())
+        kernel = kernel.repeat(channels, *[1] * (kernel.dim() - 1))
+
+        self.register_buffer("weight", kernel)
+        self.groups = channels
+
+        if dim == 1:
+            self.conv = F.conv1d
+        elif dim == 2:
+            self.conv = F.conv2d
+        elif dim == 3:
+            self.conv = F.conv3d
+        else:
+            raise RuntimeError("Only 1, 2 and 3 dimensions are supported. Received {}.".format(dim))
+
+    def forward(self, input):
+        """
+        Apply gaussian filter to input.
+        Arguments:
+            input (torch.Tensor): Input to apply gaussian filter on.
+        Returns:
+            filtered (torch.Tensor): Filtered output.
+        """
+        return self.conv(input, weight=self.weight.to(input.dtype), groups=self.groups, padding="same")
+
+
+def get_attention_scores(
+    self, query: torch.Tensor, key: torch.Tensor, attention_mask: torch.Tensor = None
+) -> torch.Tensor:
+    r"""
+    Compute the attention scores.
+
+    Args:
+        query (`torch.Tensor`): The query tensor.
+        key (`torch.Tensor`): The key tensor.
+        attention_mask (`torch.Tensor`, *optional*): The attention mask to use. If `None`, no mask is applied.
+
+    Returns:
+        `torch.Tensor`: The attention probabilities/scores.
+    """
+    if self.upcast_attention:
+        query = query.float()
+        key = key.float()
+
+    if attention_mask is None:
+        baddbmm_input = torch.empty(
+            query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device
+        )
+        beta = 0
+    else:
+        baddbmm_input = attention_mask
+        beta = 1
+
+    attention_scores = torch.baddbmm(
+        baddbmm_input,
+        query,
+        key.transpose(-1, -2),
+        beta=beta,
+        alpha=self.scale,
+    )
+    del baddbmm_input
+
+    if self.upcast_softmax:
+        attention_scores = attention_scores.float()
+
+    return attention_scores

examples/community/marigold_depth_estimation.py

@@ -18,6 +18,7 @@
 # --------------------------------------------------------------------------
 
 
+import logging
 import math
 from typing import Dict, Union
 
@@ -25,6 +26,7 @@ import matplotlib
 import numpy as np
 import torch
 from PIL import Image
+from PIL.Image import Resampling
 from scipy.optimize import minimize
 from torch.utils.data import DataLoader, TensorDataset
 from tqdm.auto import tqdm
@@ -34,13 +36,14 @@ from diffusers import (
     AutoencoderKL,
     DDIMScheduler,
     DiffusionPipeline,
+    LCMScheduler,
     UNet2DConditionModel,
 )
 from diffusers.utils import BaseOutput, check_min_version
 
 
 # Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.28.0.dev0")
+check_min_version("0.25.0")
 
 
 class MarigoldDepthOutput(BaseOutput):
@@ -61,6 +64,19 @@ class MarigoldDepthOutput(BaseOutput):
     uncertainty: Union[None, np.ndarray]
 
 
+def get_pil_resample_method(method_str: str) -> Resampling:
+    resample_method_dic = {
+        "bilinear": Resampling.BILINEAR,
+        "bicubic": Resampling.BICUBIC,
+        "nearest": Resampling.NEAREST,
+    }
+    resample_method = resample_method_dic.get(method_str, None)
+    if resample_method is None:
+        raise ValueError(f"Unknown resampling method: {resample_method}")
+    else:
+        return resample_method
+
+
 class MarigoldPipeline(DiffusionPipeline):
     """
     Pipeline for monocular depth estimation using Marigold: https://marigoldmonodepth.github.io.
@@ -113,7 +129,9 @@ class MarigoldPipeline(DiffusionPipeline):
         ensemble_size: int = 10,
         processing_res: int = 768,
         match_input_res: bool = True,
+        resample_method: str = "bilinear",
         batch_size: int = 0,
+        seed: Union[int, None] = None,
         color_map: str = "Spectral",
         show_progress_bar: bool = True,
         ensemble_kwargs: Dict = None,
@@ -129,7 +147,9 @@ class MarigoldPipeline(DiffusionPipeline):
                 If set to 0: will not resize at all.
             match_input_res (`bool`, *optional*, defaults to `True`):
                 Resize depth prediction to match input resolution.
-                Only valid if `limit_input_res` is not None.
+                Only valid if `processing_res` > 0.
+            resample_method: (`str`, *optional*, defaults to `bilinear`):
+                Resampling method used to resize images and depth predictions. This can be one of `bilinear`, `bicubic` or `nearest`, defaults to: `bilinear`.
             denoising_steps (`int`, *optional*, defaults to `10`):
                 Number of diffusion denoising steps (DDIM) during inference.
             ensemble_size (`int`, *optional*, defaults to `10`):
@@ -137,6 +157,8 @@ class MarigoldPipeline(DiffusionPipeline):
             batch_size (`int`, *optional*, defaults to `0`):
                 Inference batch size, no bigger than `num_ensemble`.
                 If set to 0, the script will automatically decide the proper batch size.
+            seed (`int`, *optional*, defaults to `None`)
+                Reproducibility seed.
             show_progress_bar (`bool`, *optional*, defaults to `True`):
                 Display a progress bar of diffusion denoising.
             color_map (`str`, *optional*, defaults to `"Spectral"`, pass `None` to skip colorized depth map generation):
@@ -146,8 +168,7 @@ class MarigoldPipeline(DiffusionPipeline):
         Returns:
             `MarigoldDepthOutput`: Output class for Marigold monocular depth prediction pipeline, including:
             - **depth_np** (`np.ndarray`) Predicted depth map, with depth values in the range of [0, 1]
-            - **depth_colored** (`None` or `PIL.Image.Image`) Colorized depth map, with the shape of [3, H, W] and
-                    values in [0, 1]. None if `color_map` is `None`
+            - **depth_colored** (`PIL.Image.Image`) Colorized depth map, with the shape of [3, H, W] and values in [0, 1], None if `color_map` is `None`
             - **uncertainty** (`None` or `np.ndarray`) Uncalibrated uncertainty(MAD, median absolute deviation)
                     coming from ensembling. None if `ensemble_size = 1`
         """
@@ -158,13 +179,21 @@ class MarigoldPipeline(DiffusionPipeline):
         if not match_input_res:
             assert processing_res is not None, "Value error: `resize_output_back` is only valid with "
         assert processing_res >= 0
-        assert denoising_steps >= 1
         assert ensemble_size >= 1
 
+        # Check if denoising step is reasonable
+        self._check_inference_step(denoising_steps)
+
+        resample_method: Resampling = get_pil_resample_method(resample_method)
+
         # ----------------- Image Preprocess -----------------
         # Resize image
         if processing_res > 0:
-            input_image = self.resize_max_res(input_image, max_edge_resolution=processing_res)
+            input_image = self.resize_max_res(
+                input_image,
+                max_edge_resolution=processing_res,
+                resample_method=resample_method,
+            )
         # Convert the image to RGB, to 1.remove the alpha channel 2.convert B&W to 3-channel
         input_image = input_image.convert("RGB")
         image = np.asarray(input_image)
@@ -203,9 +232,10 @@ class MarigoldPipeline(DiffusionPipeline):
                 rgb_in=batched_img,
                 num_inference_steps=denoising_steps,
                 show_pbar=show_progress_bar,
+                seed=seed,
             )
-            depth_pred_ls.append(depth_pred_raw.detach().clone())
-        depth_preds = torch.concat(depth_pred_ls, axis=0).squeeze()
+            depth_pred_ls.append(depth_pred_raw.detach())
+        depth_preds = torch.concat(depth_pred_ls, dim=0).squeeze()
         torch.cuda.empty_cache()  # clear vram cache for ensembling
 
         # ----------------- Test-time ensembling -----------------
@@ -227,7 +257,7 @@ class MarigoldPipeline(DiffusionPipeline):
         # Resize back to original resolution
         if match_input_res:
             pred_img = Image.fromarray(depth_pred)
-            pred_img = pred_img.resize(input_size)
+            pred_img = pred_img.resize(input_size, resample=resample_method)
             depth_pred = np.asarray(pred_img)
 
         # Clip output range
@@ -243,12 +273,32 @@ class MarigoldPipeline(DiffusionPipeline):
             depth_colored_img = Image.fromarray(depth_colored_hwc)
         else:
             depth_colored_img = None
+
         return MarigoldDepthOutput(
             depth_np=depth_pred,
             depth_colored=depth_colored_img,
             uncertainty=pred_uncert,
         )
 
+    def _check_inference_step(self, n_step: int):
+        """
+        Check if denoising step is reasonable
+        Args:
+            n_step (`int`): denoising steps
+        """
+        assert n_step >= 1
+
+        if isinstance(self.scheduler, DDIMScheduler):
+            if n_step < 10:
+                logging.warning(
+                    f"Too few denoising steps: {n_step}. Recommended to use the LCM checkpoint for few-step inference."
+                )
+        elif isinstance(self.scheduler, LCMScheduler):
+            if not 1 <= n_step <= 4:
+                logging.warning(f"Non-optimal setting of denoising steps: {n_step}. Recommended setting is 1-4 steps.")
+        else:
+            raise RuntimeError(f"Unsupported scheduler type: {type(self.scheduler)}")
+
     def _encode_empty_text(self):
         """
         Encode text embedding for empty prompt.
@@ -265,7 +315,13 @@ class MarigoldPipeline(DiffusionPipeline):
         self.empty_text_embed = self.text_encoder(text_input_ids)[0].to(self.dtype)
 
     @torch.no_grad()
-    def single_infer(self, rgb_in: torch.Tensor, num_inference_steps: int, show_pbar: bool) -> torch.Tensor:
+    def single_infer(
+        self,
+        rgb_in: torch.Tensor,
+        num_inference_steps: int,
+        seed: Union[int, None],
+        show_pbar: bool,
+    ) -> torch.Tensor:
         """
         Perform an individual depth prediction without ensembling.
 
@@ -286,10 +342,20 @@ class MarigoldPipeline(DiffusionPipeline):
         timesteps = self.scheduler.timesteps  # [T]
 
         # Encode image
-        rgb_latent = self._encode_rgb(rgb_in)
+        rgb_latent = self.encode_rgb(rgb_in)
 
         # Initial depth map (noise)
-        depth_latent = torch.randn(rgb_latent.shape, device=device, dtype=self.dtype)  # [B, 4, h, w]
+        if seed is None:
+            rand_num_generator = None
+        else:
+            rand_num_generator = torch.Generator(device=device)
+            rand_num_generator.manual_seed(seed)
+        depth_latent = torch.randn(
+            rgb_latent.shape,
+            device=device,
+            dtype=self.dtype,
+            generator=rand_num_generator,
+        )  # [B, 4, h, w]
 
         # Batched empty text embedding
         if self.empty_text_embed is None:
@@ -314,9 +380,9 @@ class MarigoldPipeline(DiffusionPipeline):
             noise_pred = self.unet(unet_input, t, encoder_hidden_states=batch_empty_text_embed).sample  # [B, 4, h, w]
 
             # compute the previous noisy sample x_t -> x_t-1
-            depth_latent = self.scheduler.step(noise_pred, t, depth_latent).prev_sample
-        torch.cuda.empty_cache()
-        depth = self._decode_depth(depth_latent)
+            depth_latent = self.scheduler.step(noise_pred, t, depth_latent, generator=rand_num_generator).prev_sample
+
+        depth = self.decode_depth(depth_latent)
 
         # clip prediction
         depth = torch.clip(depth, -1.0, 1.0)
@@ -325,7 +391,7 @@ class MarigoldPipeline(DiffusionPipeline):
 
         return depth
 
-    def _encode_rgb(self, rgb_in: torch.Tensor) -> torch.Tensor:
+    def encode_rgb(self, rgb_in: torch.Tensor) -> torch.Tensor:
         """
         Encode RGB image into latent.
 
@@ -344,7 +410,7 @@ class MarigoldPipeline(DiffusionPipeline):
         rgb_latent = mean * self.rgb_latent_scale_factor
         return rgb_latent
 
-    def _decode_depth(self, depth_latent: torch.Tensor) -> torch.Tensor:
+    def decode_depth(self, depth_latent: torch.Tensor) -> torch.Tensor:
         """
         Decode depth latent into depth map.
 
@@ -365,7 +431,7 @@ class MarigoldPipeline(DiffusionPipeline):
         return depth_mean
 
     @staticmethod
-    def resize_max_res(img: Image.Image, max_edge_resolution: int) -> Image.Image:
+    def resize_max_res(img: Image.Image, max_edge_resolution: int, resample_method=Resampling.BILINEAR) -> Image.Image:
         """
         Resize image to limit maximum edge length while keeping aspect ratio.
 
@@ -374,6 +440,8 @@ class MarigoldPipeline(DiffusionPipeline):
                 Image to be resized.
             max_edge_resolution (`int`):
                 Maximum edge length (pixel).
+            resample_method (`PIL.Image.Resampling`):
+                Resampling method used to resize images.
 
         Returns:
             `Image.Image`: Resized image.
@@ -384,7 +452,7 @@ class MarigoldPipeline(DiffusionPipeline):
         new_width = int(original_width * downscale_factor)
         new_height = int(original_height * downscale_factor)
 
-        resized_img = img.resize((new_width, new_height))
+        resized_img = img.resize((new_width, new_height), resample=resample_method)
         return resized_img
 
     @staticmethod

examples/community/pipeline_stable_diffusion_xl_instantid.py

@@ -46,6 +46,11 @@ except Exception:
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
+logger.warning(
+    "To use instant id pipelines, please make sure you have the `insightface` library installed: `pip install insightface`."
+    "Please refer to: https://huggingface.co/InstantX/InstantID for further instructions regarding inference"
+)
+
 
 def FeedForward(dim, mult=4):
     inner_dim = int(dim * mult)

scripts/convert_open_sora_stdit_to_diffusers.py

@@ -0,0 +1,172 @@
+import argparse
+import os
+
+import torch
+from safetensors import safe_open
+
+from diffusers import Transformer3DModel
+
+
+ckpt_id = "PixArt-alpha/PixArt-alpha"
+# https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/scripts/inference.py#L125
+interpolation_scale = {256: 0.5, 512: 1}
+
+
+def main(args):
+    state_dict = {}
+    with safe_open(args.orig_ckpt_path, framework="pt", device="cpu") as f:
+        for k in f.keys():
+            state_dict[k] = f.get_tensor(k)
+    converted_state_dict = {}
+
+    # Patch embeddings.
+    converted_state_dict["pos_embed.proj.weight"] = state_dict.pop("x_embedder.proj.weight")
+    converted_state_dict["pos_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["adaln_single.emb.timestep_embedder.linear_1.weight"] = state_dict.pop(
+        "t_embedder.mlp.0.weight"
+    )
+    converted_state_dict["adaln_single.emb.timestep_embedder.linear_1.bias"] = state_dict.pop("t_embedder.mlp.0.bias")
+    converted_state_dict["adaln_single.emb.timestep_embedder.linear_2.weight"] = state_dict.pop(
+        "t_embedder.mlp.2.weight"
+    )
+    converted_state_dict["adaln_single.emb.timestep_embedder.linear_2.bias"] = state_dict.pop("t_embedder.mlp.2.bias")
+
+    # Shared norm.
+    converted_state_dict["adaln_single.linear.weight"] = state_dict.pop("t_block.1.weight")
+    converted_state_dict["adaln_single.linear.bias"] = state_dict.pop("t_block.1.bias")
+
+    for depth in range(28):
+        # Transformer blocks.
+        converted_state_dict[f"transformer_blocks.{depth}.scale_shift_table"] = state_dict.pop(
+            f"blocks.{depth}.scale_shift_table"
+        )
+
+        # Attention is all you need 🤘
+
+        # Self attention.
+        q, k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.attn.qkv.weight"), 3, dim=0)
+        q_bias, k_bias, v_bias = torch.chunk(state_dict.pop(f"blocks.{depth}.attn.qkv.bias"), 3, dim=0)
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.bias"] = q_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.bias"] = k_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.weight"] = v
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.bias"] = v_bias
+        # 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"
+        )
+
+        # Temporal attention.
+        q, k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.attn_temp.qkv.weight"), 3, dim=0)
+        q_bias, k_bias, v_bias = torch.chunk(state_dict.pop(f"blocks.{depth}.attn_temp.qkv.bias"), 3, dim=0)
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_q.bias"] = q_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_k.bias"] = k_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_v.weight"] = v
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_v.bias"] = v_bias
+        # Projection.
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_out.0.weight"] = state_dict.pop(
+            f"blocks.{depth}.attn_temp.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn_temporal.to_out.0.bias"] = state_dict.pop(
+            f"blocks.{depth}.attn_temp.proj.bias"
+        )
+
+        # Feed-forward.
+        converted_state_dict[f"transformer_blocks.{depth}.ff.net.0.proj.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.fc1.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.net.0.proj.bias"] = state_dict.pop(
+            f"blocks.{depth}.mlp.fc1.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.net.2.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.fc2.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.net.2.bias"] = state_dict.pop(
+            f"blocks.{depth}.mlp.fc2.bias"
+        )
+
+        # 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")
+    converted_state_dict["pos_embed_temporal"] = state_dict.pop("pos_embed_temporal")
+
+    # DiT XL/2
+    transformer = Transformer3DModel(
+        sample_size=(16, args.image_size // 8, args.image_size // 8),
+        patch_size=(1, 2, 2),
+        num_layers=28,
+        attention_head_dim=72,
+        num_attention_heads=16,
+        in_channels=4,
+        out_channels=8,
+        cross_attention_dim=1152,
+        num_embeds_ada_norm=1000,
+        norm_eps=1e-6,
+        caption_channels=4096,
+    )
+    transformer.load_state_dict(converted_state_dict, strict=True)
+
+    assert transformer.pos_embed.pos_embed is not None
+    state_dict.pop("pos_embed")
+    state_dict.pop("y_embedder.y_embedding")
+    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.save_pretrained(os.path.join(args.dump_path, "transformer"))
+
+
+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=256,
+        type=int,
+        choices=[256, 512],
+        required=False,
+        help="Image size of pretrained model, either 256 or 512.",
+    )
+    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
+
+    args = parser.parse_args()
+    main(args)

src/diffusers/__init__.py

@@ -90,6 +90,7 @@ else:
             "T2IAdapter",
             "T5FilmDecoder",
             "Transformer2DModel",
+            "Transformer3DModel",
             "UNet1DModel",
             "UNet2DConditionModel",
             "UNet2DModel",
@@ -256,6 +257,7 @@ else:
             "LEditsPPPipelineStableDiffusion",
             "LEditsPPPipelineStableDiffusionXL",
             "MusicLDMPipeline",
+            "OpenSoraPipeline",
             "PaintByExamplePipeline",
             "PIAPipeline",
             "PixArtAlphaPipeline",
@@ -483,6 +485,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             T2IAdapter,
             T5FilmDecoder,
             Transformer2DModel,
+            Transformer3DModel,
             UNet1DModel,
             UNet2DConditionModel,
             UNet2DModel,
@@ -628,6 +631,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LEditsPPPipelineStableDiffusion,
             LEditsPPPipelineStableDiffusionXL,
             MusicLDMPipeline,
+            OpenSoraPipeline,
             PaintByExamplePipeline,
             PIAPipeline,
             PixArtAlphaPipeline,

src/diffusers/loaders/lora.py

@@ -11,6 +11,7 @@
 # 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 copy
 import inspect
 import os
 from pathlib import Path
@@ -985,7 +986,7 @@ class LoraLoaderMixin:
         self,
         adapter_names: Union[List[str], str],
         text_encoder: Optional["PreTrainedModel"] = None,  # noqa: F821
-        text_encoder_weights: List[float] = None,
+        text_encoder_weights: Optional[Union[float, List[float], List[None]]] = None,
     ):
         """
         Sets the adapter layers for the text encoder.
@@ -1003,15 +1004,20 @@ class LoraLoaderMixin:
             raise ValueError("PEFT backend is required for this method.")
 
         def process_weights(adapter_names, weights):
-            if weights is None:
-                weights = [1.0] * len(adapter_names)
-            elif isinstance(weights, float):
-                weights = [weights]
+            # Expand weights into a list, one entry per adapter
+            # e.g. for 2 adapters:  7 -> [7,7] ; [3, None] -> [3, None]
+            if not isinstance(weights, list):
+                weights = [weights] * len(adapter_names)
 
             if len(adapter_names) != len(weights):
                 raise ValueError(
                     f"Length of adapter names {len(adapter_names)} is not equal to the length of the weights {len(weights)}"
                 )
+
+            # Set None values to default of 1.0
+            # e.g. [7,7] -> [7,7] ; [3, None] -> [3,1]
+            weights = [w if w is not None else 1.0 for w in weights]
+
             return weights
 
         adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
@@ -1059,17 +1065,77 @@ class LoraLoaderMixin:
     def set_adapters(
         self,
         adapter_names: Union[List[str], str],
-        adapter_weights: Optional[List[float]] = None,
+        adapter_weights: Optional[Union[float, Dict, List[float], List[Dict]]] = None,
     ):
+        adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
+
+        adapter_weights = copy.deepcopy(adapter_weights)
+
+        # Expand weights into a list, one entry per adapter
+        if not isinstance(adapter_weights, list):
+            adapter_weights = [adapter_weights] * len(adapter_names)
+
+        if len(adapter_names) != len(adapter_weights):
+            raise ValueError(
+                f"Length of adapter names {len(adapter_names)} is not equal to the length of the weights {len(adapter_weights)}"
+            )
+
+        # Decompose weights into weights for unet, text_encoder and text_encoder_2
+        unet_lora_weights, text_encoder_lora_weights, text_encoder_2_lora_weights = [], [], []
+
+        list_adapters = self.get_list_adapters()  # eg {"unet": ["adapter1", "adapter2"], "text_encoder": ["adapter2"]}
+        all_adapters = {
+            adapter for adapters in list_adapters.values() for adapter in adapters
+        }  # eg ["adapter1", "adapter2"]
+        invert_list_adapters = {
+            adapter: [part for part, adapters in list_adapters.items() if adapter in adapters]
+            for adapter in all_adapters
+        }  # eg {"adapter1": ["unet"], "adapter2": ["unet", "text_encoder"]}
+
+        for adapter_name, weights in zip(adapter_names, adapter_weights):
+            if isinstance(weights, dict):
+                unet_lora_weight = weights.pop("unet", None)
+                text_encoder_lora_weight = weights.pop("text_encoder", None)
+                text_encoder_2_lora_weight = weights.pop("text_encoder_2", None)
+
+                if len(weights) > 0:
+                    raise ValueError(
+                        f"Got invalid key '{weights.keys()}' in lora weight dict for adapter {adapter_name}."
+                    )
+
+                if text_encoder_2_lora_weight is not None and not hasattr(self, "text_encoder_2"):
+                    logger.warning(
+                        "Lora weight dict contains text_encoder_2 weights but will be ignored because pipeline does not have text_encoder_2."
+                    )
+
+                # warn if adapter doesn't have parts specified by adapter_weights
+                for part_weight, part_name in zip(
+                    [unet_lora_weight, text_encoder_lora_weight, text_encoder_2_lora_weight],
+                    ["unet", "text_encoder", "text_encoder_2"],
+                ):
+                    if part_weight is not None and part_name not in invert_list_adapters[adapter_name]:
+                        logger.warning(
+                            f"Lora weight dict for adapter '{adapter_name}' contains {part_name}, but this will be ignored because {adapter_name} does not contain weights for {part_name}. Valid parts for {adapter_name} are: {invert_list_adapters[adapter_name]}."
+                        )
+
+            else:
+                unet_lora_weight = weights
+                text_encoder_lora_weight = weights
+                text_encoder_2_lora_weight = weights
+
+            unet_lora_weights.append(unet_lora_weight)
+            text_encoder_lora_weights.append(text_encoder_lora_weight)
+            text_encoder_2_lora_weights.append(text_encoder_2_lora_weight)
+
         unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
         # Handle the UNET
-        unet.set_adapters(adapter_names, adapter_weights)
+        unet.set_adapters(adapter_names, unet_lora_weights)
 
         # Handle the Text Encoder
         if hasattr(self, "text_encoder"):
-            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder, adapter_weights)
+            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder, text_encoder_lora_weights)
         if hasattr(self, "text_encoder_2"):
-            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder_2, adapter_weights)
+            self.set_adapters_for_text_encoder(adapter_names, self.text_encoder_2, text_encoder_2_lora_weights)
 
     def disable_lora(self):
         if not USE_PEFT_BACKEND:

src/diffusers/loaders/unet.py

@@ -47,6 +47,7 @@ from .single_file_utils import (
     infer_stable_cascade_single_file_config,
     load_single_file_model_checkpoint,
 )
+from .unet_loader_utils import _maybe_expand_lora_scales
 from .utils import AttnProcsLayers
 
 
@@ -564,7 +565,7 @@ class UNet2DConditionLoadersMixin:
     def set_adapters(
         self,
         adapter_names: Union[List[str], str],
-        weights: Optional[Union[List[float], float]] = None,
+        weights: Optional[Union[float, Dict, List[float], List[Dict], List[None]]] = None,
     ):
         """
         Set the currently active adapters for use in the UNet.
@@ -597,9 +598,9 @@ class UNet2DConditionLoadersMixin:
 
         adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
 
-        if weights is None:
-            weights = [1.0] * len(adapter_names)
-        elif isinstance(weights, float):
+        # Expand weights into a list, one entry per adapter
+        # examples for e.g. 2 adapters:  [{...}, 7] -> [7,7] ; None -> [None, None]
+        if not isinstance(weights, list):
             weights = [weights] * len(adapter_names)
 
         if len(adapter_names) != len(weights):
@@ -607,6 +608,13 @@ class UNet2DConditionLoadersMixin:
                 f"Length of adapter names {len(adapter_names)} is not equal to the length of their weights {len(weights)}."
             )
 
+        # Set None values to default of 1.0
+        # e.g. [{...}, 7] -> [{...}, 7] ; [None, None] -> [1.0, 1.0]
+        weights = [w if w is not None else 1.0 for w in weights]
+
+        # e.g. [{...}, 7] -> [{expanded dict...}, 7]
+        weights = _maybe_expand_lora_scales(self, weights)
+
         set_weights_and_activate_adapters(self, adapter_names, weights)
 
     def disable_lora(self):

src/diffusers/loaders/unet_loader_utils.py

@@ -0,0 +1,154 @@
+# 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.
+import copy
+from typing import TYPE_CHECKING, Dict, List, Union
+
+from ..utils import logging
+
+
+if TYPE_CHECKING:
+    # import here to avoid circular imports
+    from ..models import UNet2DConditionModel
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def _translate_into_actual_layer_name(name):
+    """Translate user-friendly name (e.g. 'mid') into actual layer name (e.g. 'mid_block.attentions.0')"""
+    if name == "mid":
+        return "mid_block.attentions.0"
+
+    updown, block, attn = name.split(".")
+
+    updown = updown.replace("down", "down_blocks").replace("up", "up_blocks")
+    block = block.replace("block_", "")
+    attn = "attentions." + attn
+
+    return ".".join((updown, block, attn))
+
+
+def _maybe_expand_lora_scales(unet: "UNet2DConditionModel", weight_scales: List[Union[float, Dict]]):
+    blocks_with_transformer = {
+        "down": [i for i, block in enumerate(unet.down_blocks) if hasattr(block, "attentions")],
+        "up": [i for i, block in enumerate(unet.up_blocks) if hasattr(block, "attentions")],
+    }
+    transformer_per_block = {"down": unet.config.layers_per_block, "up": unet.config.layers_per_block + 1}
+
+    expanded_weight_scales = [
+        _maybe_expand_lora_scales_for_one_adapter(
+            weight_for_adapter, blocks_with_transformer, transformer_per_block, unet.state_dict()
+        )
+        for weight_for_adapter in weight_scales
+    ]
+
+    return expanded_weight_scales
+
+
+def _maybe_expand_lora_scales_for_one_adapter(
+    scales: Union[float, Dict],
+    blocks_with_transformer: Dict[str, int],
+    transformer_per_block: Dict[str, int],
+    state_dict: None,
+):
+    """
+    Expands the inputs into a more granular dictionary. See the example below for more details.
+
+    Parameters:
+        scales (`Union[float, Dict]`):
+            Scales dict to expand.
+        blocks_with_transformer (`Dict[str, int]`):
+            Dict with keys 'up' and 'down', showing which blocks have transformer layers
+        transformer_per_block (`Dict[str, int]`):
+            Dict with keys 'up' and 'down', showing how many transformer layers each block has
+
+    E.g. turns
+    ```python
+    scales = {
+        'down': 2,
+        'mid': 3,
+        'up': {
+            'block_0': 4,
+            'block_1': [5, 6, 7]
+        }
+    }
+    blocks_with_transformer = {
+        'down': [1,2],
+        'up': [0,1]
+    }
+    transformer_per_block = {
+        'down': 2,
+        'up': 3
+    }
+    ```
+    into
+    ```python
+    {
+        'down.block_1.0': 2,
+        'down.block_1.1': 2,
+        'down.block_2.0': 2,
+        'down.block_2.1': 2,
+        'mid': 3,
+        'up.block_0.0': 4,
+        'up.block_0.1': 4,
+        'up.block_0.2': 4,
+        'up.block_1.0': 5,
+        'up.block_1.1': 6,
+        'up.block_1.2': 7,
+    }
+    ```
+    """
+    if sorted(blocks_with_transformer.keys()) != ["down", "up"]:
+        raise ValueError("blocks_with_transformer needs to be a dict with keys `'down' and `'up'`")
+
+    if sorted(transformer_per_block.keys()) != ["down", "up"]:
+        raise ValueError("transformer_per_block needs to be a dict with keys `'down' and `'up'`")
+
+    if not isinstance(scales, dict):
+        # don't expand if scales is a single number
+        return scales
+
+    scales = copy.deepcopy(scales)
+
+    if "mid" not in scales:
+        scales["mid"] = 1
+
+    for updown in ["up", "down"]:
+        if updown not in scales:
+            scales[updown] = 1
+
+        # eg {"down": 1} to {"down": {"block_1": 1, "block_2": 1}}}
+        if not isinstance(scales[updown], dict):
+            scales[updown] = {f"block_{i}": scales[updown] for i in blocks_with_transformer[updown]}
+
+        # eg {"down": "block_1": 1}} to {"down": "block_1": [1, 1]}}
+        for i in blocks_with_transformer[updown]:
+            block = f"block_{i}"
+            if not isinstance(scales[updown][block], list):
+                scales[updown][block] = [scales[updown][block] for _ in range(transformer_per_block[updown])]
+
+        # eg {"down": "block_1": [1, 1]}}  to {"down.block_1.0": 1, "down.block_1.1": 1}
+        for i in blocks_with_transformer[updown]:
+            block = f"block_{i}"
+            for tf_idx, value in enumerate(scales[updown][block]):
+                scales[f"{updown}.{block}.{tf_idx}"] = value
+
+        del scales[updown]
+
+    for layer in scales.keys():
+        if not any(_translate_into_actual_layer_name(layer) in module for module in state_dict.keys()):
+            raise ValueError(
+                f"Can't set lora scale for layer {layer}. It either doesn't exist in this unet or it has no attentions."
+            )
+
+    return {_translate_into_actual_layer_name(name): weight for name, weight in scales.items()}

src/diffusers/models/__init__.py

@@ -38,6 +38,7 @@ if is_torch_available():
     _import_structure["transformers.prior_transformer"] = ["PriorTransformer"]
     _import_structure["transformers.t5_film_transformer"] = ["T5FilmDecoder"]
     _import_structure["transformers.transformer_2d"] = ["Transformer2DModel"]
+    _import_structure["transformers.transformer_3d"] = ["Transformer3DModel"]
     _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
     _import_structure["unets.unet_1d"] = ["UNet1DModel"]
     _import_structure["unets.unet_2d"] = ["UNet2DModel"]
@@ -75,6 +76,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             PriorTransformer,
             T5FilmDecoder,
             Transformer2DModel,
+            Transformer3DModel,
             TransformerTemporalModel,
         )
         from .unets import (

src/diffusers/models/embeddings.py

@@ -187,6 +187,75 @@ class PatchEmbed(nn.Module):
         return (latent + pos_embed).to(latent.dtype)
 
 
+class PatchEmbed3D(nn.Module):
+    """Video to Patch Embedding"""
+
+    def __init__(
+        self,
+        height=224,
+        width=224,
+        patch_size=(1, 2, 2),
+        in_channels=3,
+        embed_dim=768,
+        layer_norm=False,
+        bias=True,
+        interpolation_scale=1,
+    ):
+        super().__init__()
+
+        num_patches = (height // patch_size[1]) * (width // patch_size[2])
+        self.layer_norm = layer_norm
+        self.emed_dim = embed_dim
+
+        self.proj = nn.Conv3d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size, bias=bias)
+        if layer_norm:
+            self.norm = nn.LayerNorm(embed_dim, elementwise_affine=False, eps=1e-6)
+        else:
+            self.norm = None
+
+        self.patch_size = patch_size
+        # See:
+        # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L161
+        self.height, self.width = height // patch_size[1], width // patch_size[2]
+        self.base_size = height // patch_size[1]
+        self.interpolation_scale = interpolation_scale
+        pos_embed = get_2d_sincos_pos_embed(
+            embed_dim, int(num_patches**0.5), base_size=self.base_size, interpolation_scale=self.interpolation_scale
+        )
+        self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False)
+
+    def forward(self, latent):
+        height, width = latent.shape[-2] // self.patch_size[1], latent.shape[-1] // self.patch_size[2]
+
+        latent = self.proj(latent)  # (B C T H W)
+
+        if self.layer_norm:
+            batch_size, _, num_frames, height, width = latent.size()
+            latent = latent.flatten(2).transpose(1, 2)
+            latent = self.norm(latent)
+            latent = latent.transpose(1, 2).view(batch_size, self.emed_dim, num_frames, height, width)
+
+        latent = latent.flatten(3).permute(0, 2, 3, 1)  # BCTHW -> BT(HW)C
+
+        # Interpolate positional embeddings if needed.
+        # (For PixArt-Alpha: https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L162C151-L162C160)
+        if self.height != height or self.width != width:
+            pos_embed = get_2d_sincos_pos_embed(
+                embed_dim=self.pos_embed.shape[-1],
+                grid_size=(height, width),
+                base_size=self.base_size,
+                interpolation_scale=self.interpolation_scale,
+            )
+            pos_embed = torch.from_numpy(pos_embed)
+            pos_embed = pos_embed.float().unsqueeze(0).to(latent.device)
+        else:
+            pos_embed = self.pos_embed
+
+        latent = (latent + pos_embed).to(latent.dtype)
+        latent = latent.flatten(1, 2)  # BT(H*W)C -> B(T*H*W)C
+        return latent
+
+
 class TimestepEmbedding(nn.Module):
     def __init__(
         self,

src/diffusers/models/transformers/__init__.py

@@ -6,4 +6,5 @@ if is_torch_available():
     from .prior_transformer import PriorTransformer
     from .t5_film_transformer import T5FilmDecoder
     from .transformer_2d import Transformer2DModel
+    from .transformer_3d import Transformer3DModel
     from .transformer_temporal import TransformerTemporalModel

src/diffusers/models/transformers/transformer_3d.py

@@ -0,0 +1,434 @@
+# 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 dataclasses import dataclass
+from typing import Optional, Tuple
+
+import numpy as np
+import torch
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import BaseOutput, logging
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention import FeedForward, _chunked_feed_forward
+from ..attention_processor import Attention
+from ..embeddings import PatchEmbed3D, PixArtAlphaTextProjection, get_1d_sincos_pos_embed_from_grid
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormSingle
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class Transformer3DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer3DModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: torch.FloatTensor
+
+
+def get_1d_sincos_pos_embed(embed_dim, length, scale=1.0):
+    pos = np.arange(0, length)[..., None] / scale
+    return get_1d_sincos_pos_embed_from_grid(embed_dim, pos)
+
+
+@maybe_allow_in_graph
+class Transformer3DBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        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.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        cross_attention_dim: int,
+        num_embeds_ada_norm: Optional[int] = None,
+        norm_eps: float = 1e-6,
+        num_temporal_patches: int = 16,
+        num_spatial_patches: int = 256,
+    ):
+        super().__init__()
+        # We keep these boolean flags for backward-compatibility.
+
+        self.use_ada_layer_norm_single = True
+        self.num_embeds_ada_norm = num_embeds_ada_norm
+        self.num_temporal_patches = num_temporal_patches
+        self.num_spatial_patches = num_spatial_patches
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Spatial Self-Attn
+        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,
+            bias=True,
+        )
+
+        # 2. Temporal Self-Attn
+        self.attn_temporal = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            bias=True,
+        )
+
+        # 2. Cross-Attn
+        # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+        # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+        # the second cross attention block.
+        self.norm2 = nn.LayerNorm(dim, norm_eps, elementwise_affine=False)
+        self.attn2 = Attention(
+            query_dim=dim,
+            cross_attention_dim=cross_attention_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            bias=True,
+        )
+
+        # 3. Feed-forward
+        self.ff = FeedForward(dim, activation_fn="gelu-approximate")
+
+        # 4. Scale-shift for PixArt-Alpha.
+        self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor,
+        timestep: torch.LongTensor,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        temporal_pos_embed: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        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)
+        norm_hidden_states = self.norm1(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+        norm_hidden_states = norm_hidden_states.squeeze(1)
+
+        # 1. Spatial Self-Attention
+        # reshape (batch, num_temporal_patches*num_spatial_patches, dim) -> (batch * num_temporal_patches, num_spatial_patches, dim)
+        norm_hidden_states = norm_hidden_states.view(
+            batch_size, self.num_temporal_patches, self.num_spatial_patches, -1
+        )
+        norm_hidden_states = norm_hidden_states.view(
+            batch_size * self.num_temporal_patches, self.num_spatial_patches, -1
+        )
+
+        attn_output = self.attn1(norm_hidden_states)
+
+        # reshape (batch * num_temporal_patches, num_spatial_patches, dim) -> (batch, num_temporal_patches*num_spatial_patches, dim)
+        attn_output = attn_output.view(batch_size, self.num_temporal_patches, self.num_spatial_patches, -1)
+        attn_output = attn_output.view(batch_size, self.num_temporal_patches * self.num_spatial_patches, -1)
+
+        attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2. Temporal Self-Attention
+        # reshape (batch, num_temporal_patches*num_spatial_patches, dim) -> (batch * num_spatial_patches, num_temporal_patches, dim)
+        temporal_hidden_states = (
+            hidden_states.view(batch_size, self.num_temporal_patches, self.num_spatial_patches, -1)
+            .transpose(1, 2)
+            .contiguous()
+        )
+        temporal_hidden_states = temporal_hidden_states.view(
+            batch_size * self.num_spatial_patches, self.num_temporal_patches, -1
+        )
+
+        if temporal_pos_embed is not None:
+            temporal_hidden_states = temporal_hidden_states + temporal_pos_embed
+
+        attn_output = self.attn_temporal(temporal_hidden_states)
+
+        # reshape (batch * num_spatial_patches, num_temporal_patches, dim) -> (batch, num_temporal_patches*num_spatial_patches, dim)
+        attn_output = (
+            attn_output.view(batch_size, self.num_spatial_patches, self.num_temporal_patches, -1)
+            .transpose(1, 2)
+            .contiguous()
+        )
+        attn_output = attn_output.view(batch_size, self.num_temporal_patches * self.num_spatial_patches, -1)
+
+        attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 3. Cross-Attention
+        # For PixArt norm2 isn't applied here:
+        # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+        norm_hidden_states = hidden_states
+        attn_output = self.attn2(
+            norm_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
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class Transformer3DModel(ModelMixin, ConfigMixin):
+    """
+    A 3D Transformer model for image-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states.
+
+            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        sample_size: Tuple[int] = (2, 4, 4),
+        patch_size: Tuple[int] = (1, 2, 2),
+        in_channels: int = 4,
+        out_channels: int = 8,
+        num_layers: int = 1,
+        cross_attention_dim: int = 256,
+        num_embeds_ada_norm: int = 1000,
+        norm_eps: float = 1e-6,
+        caption_channels: int = 256,
+        interpolation_scale: float = None,
+    ):
+        super().__init__()
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # 2. Define input layers
+        self.height = sample_size[1]
+        self.width = sample_size[2]
+        self.num_patches = np.prod([sample_size[i] // patch_size[i] for i in range(3)])
+        self.num_temporal_patches = sample_size[0] // patch_size[0]
+        self.num_spatial_patches = self.num_patches // self.num_temporal_patches
+
+        self.patch_size = patch_size
+        interpolation_scale = (
+            interpolation_scale if interpolation_scale is not None else max(self.config.sample_size[1] // 64, 1)
+        )
+        self.pos_embed = PatchEmbed3D(
+            height=self.height,
+            width=self.width,
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=inner_dim,
+            interpolation_scale=interpolation_scale,
+        )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                Transformer3DBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    cross_attention_dim=cross_attention_dim,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    norm_eps=norm_eps,
+                    num_temporal_patches=self.num_temporal_patches,
+                    num_spatial_patches=self.num_spatial_patches,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+        self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+        self.proj_out = nn.Linear(inner_dim, np.prod(patch_size) * self.out_channels)
+
+        # 5. PixArt-Alpha blocks.
+        self.adaln_single = AdaLayerNormSingle(inner_dim)
+
+        self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+
+        interpolation_scale = max(self.config.sample_size[0] // 16, 1)
+        temporal_pos_embed = get_1d_sincos_pos_embed(inner_dim, self.num_temporal_patches, scale=interpolation_scale)
+        temporal_pos_embed = torch.from_numpy(temporal_pos_embed).float().unsqueeze(0).requires_grad_(False)
+        self.register_buffer("pos_embed_temporal", temporal_pos_embed)
+
+        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: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # 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
+        height, width = hidden_states.shape[-2] // self.patch_size[1], hidden_states.shape[-1] // self.patch_size[2]
+        # import ipdb; ipdb.set_trace()
+        hidden_states = self.pos_embed(hidden_states)
+
+        batch_size = hidden_states.shape[0]
+        added_cond_kwargs = {"resolution": None, "aspect_ratio": None}
+        timestep, embedded_timestep = self.adaln_single(
+            timestep, added_cond_kwargs=added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+        )
+
+        # 2. Blocks
+        encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+        encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        for i, block in enumerate(self.transformer_blocks):
+            hidden_states = block(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                timestep=timestep,
+                temporal_pos_embed=self.pos_embed_temporal if i == 0 else None,
+            )
+
+        # 3. Output
+        shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
+        hidden_states = self.norm_out(hidden_states)
+        # Modulation
+        hidden_states = hidden_states * (1 + scale) + shift
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = hidden_states.squeeze(1)
+
+        # unpatchify
+        hidden_states = hidden_states.reshape(
+            shape=(
+                -1,
+                self.num_temporal_patches,
+                height,
+                width,
+                self.patch_size[1],
+                self.patch_size[2],
+                self.out_channels,
+            )
+        )
+        hidden_states = torch.einsum("nthwpqc->ncthpwq", hidden_states)
+        output = hidden_states.reshape(
+            shape=(
+                -1,
+                self.out_channels,
+                self.num_temporal_patches,
+                height * self.patch_size[1],
+                width * self.patch_size[2],
+            )
+        )
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer3DModelOutput(sample=output)

src/diffusers/models/transformers/transformer_temporal.py

@@ -311,10 +311,10 @@ class TransformerSpatioTemporalModel(nn.Module):
         time_context_first_timestep = time_context[None, :].reshape(
             batch_size, num_frames, -1, time_context.shape[-1]
         )[:, 0]
-        time_context = time_context_first_timestep[None, :].broadcast_to(
-            height * width, batch_size, 1, time_context.shape[-1]
+        time_context = time_context_first_timestep[:, None].broadcast_to(
+            batch_size, height * width, time_context.shape[-2], time_context.shape[-1]
         )
-        time_context = time_context.reshape(height * width * batch_size, 1, time_context.shape[-1])
+        time_context = time_context.reshape(batch_size * height * width, -1, time_context.shape[-1])
 
         residual = hidden_states
 

src/diffusers/pipelines/__init__.py

@@ -247,6 +247,7 @@ else:
         "WuerstchenDecoderPipeline",
         "WuerstchenPriorPipeline",
     ]
+    _import_structure["open_sora"] = ["OpenSoraPipeline"]
 try:
     if not is_onnx_available():
         raise OptionalDependencyNotAvailable()
@@ -438,6 +439,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LEditsPPPipelineStableDiffusionXL,
         )
         from .musicldm import MusicLDMPipeline
+        from .open_sora import OpenSoraPipeline
         from .paint_by_example import PaintByExamplePipeline
         from .pia import PIAPipeline
         from .pixart_alpha import PixArtAlphaPipeline

src/diffusers/pipelines/open_sora/__init__.py

@@ -0,0 +1,48 @@
+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_open_sora"] = ["OpenSoraPipeline"]
+
+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_open_sora import OpenSoraPipeline
+
+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/open_sora/pipeline_open_sora.py

@@ -0,0 +1,912 @@
+# Copyright 2024 Open-Sora 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 dataclasses import dataclass
+from typing import Callable, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...models import AutoencoderKL, Transformer3DModel
+from ...schedulers import DPMSolverMultistepScheduler
+from ...utils import (
+    BACKENDS_MAPPING,
+    BaseOutput,
+    deprecate,
+    is_bs4_available,
+    is_ftfy_available,
+    logging,
+    replace_example_docstring,
+)
+from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+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 OpenSoraPipeline
+
+        >>> # You can replace the checkpoint id with "PixArt-alpha/PixArt-XL-2-512x512" too.
+        >>> pipe = OpenSoraPipeline.from_pretrained("PixArt-alpha/PixArt-XL-2-1024-MS", torch_dtype=torch.float16)
+        >>> # Enable memory optimizations.
+        >>> pipe.enable_model_cpu_offload()
+
+        >>> prompt = "A small cactus with a happy face in the Sahara desert."
+        >>> image = pipe(prompt).images[0]
+        ```
+"""
+
+
+@dataclass
+class OpenSoraPipelineOutput(BaseOutput):
+    r"""
+    Output class for Stable Video Diffusion pipeline.
+
+    Args:
+        frames (`[List[List[PIL.Image.Image]]`, `np.ndarray`, `torch.FloatTensor`]):
+            List of denoised PIL images of length `batch_size` or numpy array or torch tensor
+            of shape `(batch_size, num_frames, height, width, num_channels)`.
+    """
+
+    frames: Union[List[List[PIL.Image.Image]], np.ndarray, torch.FloatTensor]
+
+
+def _append_dims(x, target_dims):
+    """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
+    dims_to_append = target_dims - x.ndim
+    if dims_to_append < 0:
+        raise ValueError(f"input has {x.ndim} dims but target_dims is {target_dims}, which is less")
+    return x[(...,) + (None,) * dims_to_append]
+
+
+# Copied from diffusers.pipelines.animatediff.pipeline_animatediff.tensor2vid
+def tensor2vid(video: torch.Tensor, processor: VaeImageProcessor, output_type: str = "np"):
+    batch_size, channels, num_frames, height, width = video.shape
+    outputs = []
+    for batch_idx in range(batch_size):
+        batch_vid = video[batch_idx].permute(1, 0, 2, 3)
+        batch_output = processor.postprocess(batch_vid, output_type)
+
+        outputs.append(batch_output)
+
+    if output_type == "np":
+        outputs = np.stack(outputs)
+
+    elif output_type == "pt":
+        outputs = torch.stack(outputs)
+
+    elif not output_type == "pil":
+        raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil']")
+
+    return outputs
+
+
+# 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,
+    **kwargs,
+):
+    """
+    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 support arbitrary spacing between timesteps. If `None`, then the default
+                timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+                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:
+        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)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class OpenSoraPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using Open-Sora.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            Frozen text-encoder. Open-Sora uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the
+            [t5-v1_1-xxl](https://huggingface.co/Open-Sora/Open-Sora/tree/main/t5-v1_1-xxl) variant.
+        tokenizer (`T5Tokenizer`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        transformer ([`Transformer2DModel`]):
+            A text conditioned `Transformer2DModel` to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+    """
+
+    bad_punct_regex = re.compile(
+        r"["
+        + "#®•©™&@·º½¾¿¡§~"
+        + r"\)"
+        + r"\("
+        + r"\]"
+        + r"\["
+        + r"\}"
+        + r"\{"
+        + r"\|"
+        + "\\"
+        + r"\/"
+        + r"\*"
+        + r"]{1,}"
+    )  # noqa
+
+    _optional_components = ["tokenizer", "text_encoder"]
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+
+    def __init__(
+        self,
+        tokenizer: T5Tokenizer,
+        text_encoder: T5EncoderModel,
+        vae: AutoencoderKL,
+        transformer: Transformer3DModel,
+        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.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+    # Adapted from https://github.com/Open-Sora/Open-Sora/blob/master/diffusion/model/utils.py
+    def mask_text_embeddings(self, emb, mask):
+        if emb.shape[0] == 1:
+            keep_index = mask.sum().item()
+            return emb[:, :, :keep_index, :], keep_index
+        else:
+            masked_feature = emb * mask[:, None, :, None]
+            return masked_feature, emb.shape[2]
+
+    # Adapted from diffusers.pipelines.deepfloyd_if.pipeline_if.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.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        prompt_attention_mask: Optional[torch.FloatTensor] = None,
+        negative_prompt_attention_mask: Optional[torch.FloatTensor] = None,
+        clean_caption: bool = False,
+        max_sequence_length: int = 120,
+        **kwargs,
+    ):
+        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
+                Open-Sora, 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.FloatTensor`, *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.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For Open-Sora, 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 120): Maximum sequence length to use for the prompt.
+        """
+
+        if "mask_feature" in kwargs:
+            deprecation_message = "The use of `mask_feature` is deprecated. It is no longer used in any computation and that doesn't affect the end results. It will be removed in a future version."
+            deprecate("mask_feature", "1.0.0", deprecation_message, standard_warn=False)
+
+        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]
+
+        # See Section 3.1. of the paper.
+        max_length = max_sequence_length
+
+        if prompt_embeds is None:
+            prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
+            text_inputs = self.tokenizer(
+                prompt,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                add_special_tokens=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids
+            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_length - 1 : -1])
+                logger.warning(
+                    "The following part of your input was truncated because CLIP can only handle sequences up to"
+                    f" {max_length} tokens: {removed_text}"
+                )
+
+            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]
+
+        if self.text_encoder is not None:
+            dtype = self.text_encoder.dtype
+        elif self.transformer is not None:
+            dtype = self.transformer.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
+            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,
+        negative_prompt,
+        callback_steps,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_attention_mask=None,
+        negative_prompt_attention_mask=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        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: int,
+        num_frames: int,
+        num_channels_latents: int,
+        height: int,
+        width: int,
+        dtype: torch.dtype,
+        device: Union[str, torch.device],
+        generator: torch.Generator,
+        latents: Optional[torch.FloatTensor] = None,
+    ):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            num_frames,
+            height // self.vae_scale_factor,
+            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."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def decode_latents(self, latents: torch.FloatTensor, num_frames: int, decode_chunk_size: int = 14):
+        # [batch, channels, frames, height, width] -> [batch, frames, channels, height, width]
+        latents = latents.permute(0, 2, 1, 3, 4)
+        # [batch, frames, channels, height, width] -> [batch*frames, channels, height, width]
+        latents = latents.flatten(0, 1)
+
+        latents = 1 / self.vae.config.scaling_factor * latents
+
+        forward_vae_fn = self.vae._orig_mod.forward if is_compiled_module(self.vae) else self.vae.forward
+        accepts_num_frames = "num_frames" in set(inspect.signature(forward_vae_fn).parameters.keys())
+
+        # decode decode_chunk_size frames at a time to avoid OOM
+        frames = []
+        for i in range(0, latents.shape[0], decode_chunk_size):
+            num_frames_in = latents[i : i + decode_chunk_size].shape[0]
+            decode_kwargs = {}
+            if accepts_num_frames:
+                # we only pass num_frames_in if it's expected
+                decode_kwargs["num_frames"] = num_frames_in
+
+            frame = self.vae.decode(latents[i : i + decode_chunk_size], **decode_kwargs).sample
+            frames.append(frame)
+        frames = torch.cat(frames, dim=0)
+
+        # [batch*frames, channels, height, width] -> [batch, channels, frames, height, width]
+        frames = frames.reshape(-1, num_frames, *frames.shape[1:]).permute(0, 2, 1, 3, 4)
+
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        frames = frames.float()
+        return frames
+
+    @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,
+        guidance_scale: float = 4.5,
+        min_guidance_scale: Optional[float] = None,
+        max_guidance_scale: Optional[float] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        prompt_attention_mask: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_attention_mask: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        clean_caption: bool = True,
+        max_sequence_length: int = 120,
+        decode_chunk_size: int = 4,
+        **kwargs,
+    ) -> 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 100):
+                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. If not defined, equal spaced `num_inference_steps`
+                timesteps are used. Must be in descending order.
+            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.FloatTensor`, *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.FloatTensor`, *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.FloatTensor`, *optional*): Pre-generated attention mask for text embeddings.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Alpha 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.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            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.
+            max_sequence_length (`int` defaults to 120): Maximum sequence length to use with the `prompt`.
+
+        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
+        """
+        # 1. Check inputs. Raise error if not correct
+        num_frames = self.transformer.config.sample_size[0]
+        height = height or self.transformer.config.sample_size[1] * self.vae_scale_factor
+        width = width or self.transformer.config.sample_size[2] * self.vae_scale_factor
+
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt,
+            callback_steps,
+            prompt_embeds,
+            negative_prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_attention_mask,
+        )
+
+        # 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
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt,
+            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,
+        )
+        if 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)
+
+        # 5. Prepare latents.
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_frames,
+            latent_channels,
+            height,
+            width,
+            prompt_embeds.dtype,
+            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)
+
+        # 8. Prepare guidance scale
+        # TODO: Hacky for testing, make this cleaner
+        if min_guidance_scale is not None and max_guidance_scale is not None:
+            guidance_scale = torch.linspace(min_guidance_scale, max_guidance_scale, num_frames).unsqueeze(0)
+            guidance_scale = guidance_scale.to(device, latents.dtype)
+            guidance_scale = guidance_scale.repeat(batch_size * num_images_per_prompt, 1)
+            guidance_scale = _append_dims(guidance_scale, latents.ndim).transpose(1, 2)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                current_timestep = t
+                if not torch.is_tensor(current_timestep):
+                    # 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 = latent_model_input.device.type == "mps"
+                    if isinstance(current_timestep, float):
+                        dtype = torch.float32 if is_mps else torch.float64
+                    else:
+                        dtype = torch.int32 if is_mps else torch.int64
+                    current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device)
+                elif len(current_timestep.shape) == 0:
+                    current_timestep = current_timestep[None].to(latent_model_input.device)
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                current_timestep = current_timestep.expand(latent_model_input.shape[0])
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    encoder_attention_mask=prompt_attention_mask,
+                    timestep=current_timestep,
+                    return_dict=False,
+                )[0]
+
+                # 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
+
+                # perform guidance
+                if 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 = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                # 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 callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+        if not output_type == "latent":
+            frames = self.decode_latents(latents, num_frames, decode_chunk_size)
+            frames = tensor2vid(frames, self.image_processor, output_type=output_type)
+        else:
+            frames = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (frames,)
+
+        return OpenSoraPipelineOutput(frames=frames)

src/diffusers/pipelines/stable_diffusion_sag/pipeline_stable_diffusion_sag.py

@@ -401,6 +401,40 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, StableDiffusionMixin, Textua
 
             return image_embeds, uncond_image_embeds
 
+    def prepare_ip_adapter_image_embeds(
+        self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+    ):
+        if ip_adapter_image_embeds is None:
+            if not isinstance(ip_adapter_image, list):
+                ip_adapter_image = [ip_adapter_image]
+
+            if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+                raise ValueError(
+                    f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+                )
+
+            image_embeds = []
+            for single_ip_adapter_image, image_proj_layer in zip(
+                ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+            ):
+                output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+                single_image_embeds, single_negative_image_embeds = self.encode_image(
+                    single_ip_adapter_image, device, 1, output_hidden_state
+                )
+                single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+                single_negative_image_embeds = torch.stack(
+                    [single_negative_image_embeds] * num_images_per_prompt, dim=0
+                )
+
+                if do_classifier_free_guidance:
+                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+                    single_image_embeds = single_image_embeds.to(device)
+
+                image_embeds.append(single_image_embeds)
+        else:
+            image_embeds = ip_adapter_image_embeds
+        return image_embeds
+
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
     def run_safety_checker(self, image, device, dtype):
         if self.safety_checker is None:
@@ -535,6 +569,7 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, StableDiffusionMixin, Textua
         prompt_embeds: Optional[torch.FloatTensor] = None,
         negative_prompt_embeds: Optional[torch.FloatTensor] = None,
         ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
         output_type: Optional[str] = "pil",
         return_dict: bool = True,
         callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
@@ -583,6 +618,9 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, StableDiffusionMixin, Textua
                 not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
             ip_adapter_image: (`PipelineImageInput`, *optional*):
                 Optional image input to work with IP Adapters.
+            ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. If not
+                provided, embeddings are computed from the `ip_adapter_image` input argument.
             output_type (`str`, *optional*, defaults to `"pil"`):
                 The output format of the generated image. Choose between `PIL.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
@@ -636,13 +674,24 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, StableDiffusionMixin, Textua
         # `sag_scale = 0` means no self-attention guidance
         do_self_attention_guidance = sag_scale > 0.0
 
-        if ip_adapter_image is not None:
-            output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
-            image_embeds, negative_image_embeds = self.encode_image(
-                ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+            ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds(
+                ip_adapter_image,
+                ip_adapter_image_embeds,
+                device,
+                batch_size * num_images_per_prompt,
+                do_classifier_free_guidance,
             )
+
             if do_classifier_free_guidance:
-                image_embeds = torch.cat([negative_image_embeds, image_embeds])
+                image_embeds = []
+                negative_image_embeds = []
+                for tmp_image_embeds in ip_adapter_image_embeds:
+                    single_negative_image_embeds, single_image_embeds = tmp_image_embeds.chunk(2)
+                    image_embeds.append(single_image_embeds)
+                    negative_image_embeds.append(single_negative_image_embeds)
+            else:
+                image_embeds = ip_adapter_image_embeds
 
         # 3. Encode input prompt
         prompt_embeds, negative_prompt_embeds = self.encode_prompt(
@@ -687,8 +736,18 @@ class StableDiffusionSAGPipeline(DiffusionPipeline, StableDiffusionMixin, Textua
         extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
 
         # 6.1 Add image embeds for IP-Adapter
-        added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
-        added_uncond_kwargs = {"image_embeds": negative_image_embeds} if ip_adapter_image is not None else None
+        added_cond_kwargs = (
+            {"image_embeds": image_embeds}
+            if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+            else None
+        )
+
+        if do_classifier_free_guidance:
+            added_uncond_kwargs = (
+                {"image_embeds": negative_image_embeds}
+                if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+                else None
+            )
 
         # 7. Denoising loop
         store_processor = CrossAttnStoreProcessor()

src/diffusers/schedulers/scheduling_unipc_multistep.py

@@ -127,6 +127,9 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
             Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
         steps_offset (`int`, defaults to 0):
             An offset added to the inference steps, as required by some model families.
+        final_sigmas_type (`str`, defaults to `"zero"`):
+            The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final sigma
+            is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0.
     """
 
     _compatibles = [e.name for e in KarrasDiffusionSchedulers]
@@ -153,6 +156,7 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
         use_karras_sigmas: Optional[bool] = False,
         timestep_spacing: str = "linspace",
         steps_offset: int = 0,
+        final_sigmas_type: Optional[str] = "zero",  # "zero", "sigma_min"
     ):
         if trained_betas is not None:
             self.betas = torch.tensor(trained_betas, dtype=torch.float32)
@@ -265,10 +269,25 @@ class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
             sigmas = np.flip(sigmas).copy()
             sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
             timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
-            sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
+            if self.config.final_sigmas_type == "sigma_min":
+                sigma_last = sigmas[-1]
+            elif self.config.final_sigmas_type == "zero":
+                sigma_last = 0
+            else:
+                raise ValueError(
+                    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)
         else:
             sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
-            sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
+            if self.config.final_sigmas_type == "sigma_min":
+                sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
+            elif self.config.final_sigmas_type == "zero":
+                sigma_last = 0
+            else:
+                raise ValueError(
+                    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)
 
         self.sigmas = torch.from_numpy(sigmas)

src/diffusers/utils/dummy_pt_objects.py

@@ -227,6 +227,21 @@ class Transformer2DModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class Transformer3DModel(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 UNet1DModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/peft_utils.py

@@ -230,16 +230,26 @@ def delete_adapter_layers(model, adapter_name):
 def set_weights_and_activate_adapters(model, adapter_names, weights):
     from peft.tuners.tuners_utils import BaseTunerLayer
 
+    def get_module_weight(weight_for_adapter, module_name):
+        if not isinstance(weight_for_adapter, dict):
+            # If weight_for_adapter is a single number, always return it.
+            return weight_for_adapter
+
+        for layer_name, weight_ in weight_for_adapter.items():
+            if layer_name in module_name:
+                return weight_
+        raise RuntimeError(f"No LoRA weight found for module {module_name}.")
+
     # iterate over each adapter, make it active and set the corresponding scaling weight
     for adapter_name, weight in zip(adapter_names, weights):
-        for module in model.modules():
+        for module_name, module in model.named_modules():
             if isinstance(module, BaseTunerLayer):
                 # For backward compatbility with previous PEFT versions
                 if hasattr(module, "set_adapter"):
                     module.set_adapter(adapter_name)
                 else:
                     module.active_adapter = adapter_name
-                module.set_scale(adapter_name, weight)
+                module.set_scale(adapter_name, get_module_weight(weight, module_name))
 
     # set multiple active adapters
     for module in model.modules():

src/diffusers/utils/testing_utils.py

@@ -105,10 +105,21 @@ def numpy_cosine_similarity_distance(a, b):
     return distance
 
 
-def print_tensor_test(tensor, filename="test_corrections.txt", expected_tensor_name="expected_slice"):
+def print_tensor_test(
+    tensor,
+    limit_to_slices=None,
+    max_torch_print=None,
+    filename="test_corrections.txt",
+    expected_tensor_name="expected_slice",
+):
+    if max_torch_print:
+        torch.set_printoptions(threshold=10_000)
+
     test_name = os.environ.get("PYTEST_CURRENT_TEST")
     if not torch.is_tensor(tensor):
         tensor = torch.from_numpy(tensor)
+    if limit_to_slices:
+        tensor = tensor[0, -3:, -3:, -1]
 
     tensor_str = str(tensor.detach().cpu().flatten().to(torch.float32)).replace("\n", "")
     # format is usually:
@@ -117,7 +128,7 @@ def print_tensor_test(tensor, filename="test_corrections.txt", expected_tensor_n
     test_file, test_class, test_fn = test_name.split("::")
     test_fn = test_fn.split()[0]
     with open(filename, "a") as f:
-        print(";".join([test_file, test_class, test_fn, output_str]), file=f)
+        print("::".join([test_file, test_class, test_fn, output_str]), file=f)
 
 
 def get_tests_dir(append_path=None):

test.py

@@ -0,0 +1,39 @@
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+from diffusers import Transformer3DModel, AutoencoderTiny, DPMSolverMultistepScheduler, OpenSoraPipeline
+
+channels, num_frames, height, width, text_dim = 4, 2, 4, 4, 32
+
+model = Transformer3DModel(
+    in_channels=channels,
+    out_channels=channels*2,
+    cross_attention_dim=1408,
+    caption_channels=text_dim,
+    num_embeds_ada_norm=1000,
+    sample_size=(num_frames, height, width),
+)
+
+x = torch.randn(1, channels, num_frames, height, width)
+y = torch.randn(1, 77, 32)
+t = torch.ones(1)
+
+# with torch.no_grad():
+#     out = model(x, y, t)
+#     print(out.sample.shape)  # torch.Size([1, 8, 2, 4, 4])
+
+
+text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+tokenizer = T5Tokenizer.from_pretrained("google-t5/t5-small")
+vae = AutoencoderTiny.from_pretrained("madebyollin/taesd")
+scheduler = DPMSolverMultistepScheduler.from_pretrained("PixArt-alpha/PixArt-XL-2-1024-MS", subfolder="scheduler")
+
+pipe = OpenSoraPipeline(
+    tokenizer=tokenizer,
+    text_encoder=text_encoder,
+    vae=vae,
+    transformer=model,
+    scheduler=scheduler,
+)
+
+prompt = ""
+out = pipe(prompt, num_inference_steps=1, min_guidance_scale=1.0, max_guidance_scale=3.0)

tests/lora/utils.py

@@ -15,6 +15,7 @@
 import os
 import tempfile
 import unittest
+from itertools import product
 
 import numpy as np
 import torch
@@ -762,6 +763,218 @@ class PeftLoraLoaderMixinTests:
                 "output with no lora and output with lora disabled should give same results",
             )
 
+    def test_simple_inference_with_text_unet_block_scale(self):
+        """
+        Tests a simple inference with lora attached to text encoder and unet, attaches
+        one adapter and set differnt weights for different blocks (i.e. block lora)
+        """
+        for scheduler_cls in [DDIMScheduler, LCMScheduler]:
+            components, text_lora_config, unet_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)).images
+
+            pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
+            pipe.unet.add_adapter(unet_lora_config, "adapter-1")
+
+            self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder")
+            self.assertTrue(check_if_lora_correctly_set(pipe.unet), "Lora not correctly set in Unet")
+
+            if self.has_two_text_encoders:
+                pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-1")
+                self.assertTrue(
+                    check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
+                )
+
+            weights_1 = {"text_encoder": 2, "unet": {"down": 5}}
+            pipe.set_adapters("adapter-1", weights_1)
+            output_weights_1 = pipe(**inputs, generator=torch.manual_seed(0)).images
+
+            weights_2 = {"unet": {"up": 5}}
+            pipe.set_adapters("adapter-1", weights_2)
+            output_weights_2 = pipe(**inputs, generator=torch.manual_seed(0)).images
+
+            self.assertFalse(
+                np.allclose(output_weights_1, output_weights_2, atol=1e-3, rtol=1e-3),
+                "LoRA weights 1 and 2 should give different results",
+            )
+            self.assertFalse(
+                np.allclose(output_no_lora, output_weights_1, atol=1e-3, rtol=1e-3),
+                "No adapter and LoRA weights 1 should give different results",
+            )
+            self.assertFalse(
+                np.allclose(output_no_lora, output_weights_2, atol=1e-3, rtol=1e-3),
+                "No adapter and LoRA weights 2 should give different results",
+            )
+
+            pipe.disable_lora()
+            output_disabled = pipe(**inputs, generator=torch.manual_seed(0)).images
+
+            self.assertTrue(
+                np.allclose(output_no_lora, output_disabled, atol=1e-3, rtol=1e-3),
+                "output with no lora and output with lora disabled should give same results",
+            )
+
+    def test_simple_inference_with_text_unet_multi_adapter_block_lora(self):
+        """
+        Tests a simple inference with lora attached to text encoder and unet, attaches
+        multiple adapters and set differnt weights for different blocks (i.e. block lora)
+        """
+        for scheduler_cls in [DDIMScheduler, LCMScheduler]:
+            components, text_lora_config, unet_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)).images
+
+            pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
+            pipe.text_encoder.add_adapter(text_lora_config, "adapter-2")
+
+            pipe.unet.add_adapter(unet_lora_config, "adapter-1")
+            pipe.unet.add_adapter(unet_lora_config, "adapter-2")
+
+            self.assertTrue(check_if_lora_correctly_set(pipe.text_encoder), "Lora not correctly set in text encoder")
+            self.assertTrue(check_if_lora_correctly_set(pipe.unet), "Lora not correctly set in Unet")
+
+            if self.has_two_text_encoders:
+                pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-1")
+                pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-2")
+                self.assertTrue(
+                    check_if_lora_correctly_set(pipe.text_encoder_2), "Lora not correctly set in text encoder 2"
+                )
+
+            scales_1 = {"text_encoder": 2, "unet": {"down": 5}}
+            scales_2 = {"unet": {"down": 5, "mid": 5}}
+            pipe.set_adapters("adapter-1", scales_1)
+
+            output_adapter_1 = pipe(**inputs, generator=torch.manual_seed(0)).images
+
+            pipe.set_adapters("adapter-2", scales_2)
+            output_adapter_2 = pipe(**inputs, generator=torch.manual_seed(0)).images
+
+            pipe.set_adapters(["adapter-1", "adapter-2"], [scales_1, scales_2])
+
+            output_adapter_mixed = pipe(**inputs, generator=torch.manual_seed(0)).images
+
+            # Fuse and unfuse should lead to the same results
+            self.assertFalse(
+                np.allclose(output_adapter_1, output_adapter_2, atol=1e-3, rtol=1e-3),
+                "Adapter 1 and 2 should give different results",
+            )
+
+            self.assertFalse(
+                np.allclose(output_adapter_1, output_adapter_mixed, atol=1e-3, rtol=1e-3),
+                "Adapter 1 and mixed adapters should give different results",
+            )
+
+            self.assertFalse(
+                np.allclose(output_adapter_2, output_adapter_mixed, atol=1e-3, rtol=1e-3),
+                "Adapter 2 and mixed adapters should give different results",
+            )
+
+            pipe.disable_lora()
+
+            output_disabled = pipe(**inputs, generator=torch.manual_seed(0)).images
+
+            self.assertTrue(
+                np.allclose(output_no_lora, output_disabled, atol=1e-3, rtol=1e-3),
+                "output with no lora and output with lora disabled should give same results",
+            )
+
+            # a mismatching number of adapter_names and adapter_weights should raise an error
+            with self.assertRaises(ValueError):
+                pipe.set_adapters(["adapter-1", "adapter-2"], [scales_1])
+
+    def test_simple_inference_with_text_unet_block_scale_for_all_dict_options(self):
+        """Tests that any valid combination of lora block scales can be used in pipe.set_adapter"""
+
+        def updown_options(blocks_with_tf, layers_per_block, value):
+            """
+            Generate every possible combination for how a lora weight dict for the up/down part can be.
+            E.g. 2, {"block_1": 2}, {"block_1": [2,2,2]}, {"block_1": 2, "block_2": [2,2,2]}, ...
+            """
+            num_val = value
+            list_val = [value] * layers_per_block
+
+            node_opts = [None, num_val, list_val]
+            node_opts_foreach_block = [node_opts] * len(blocks_with_tf)
+
+            updown_opts = [num_val]
+            for nodes in product(*node_opts_foreach_block):
+                if all(n is None for n in nodes):
+                    continue
+                opt = {}
+                for b, n in zip(blocks_with_tf, nodes):
+                    if n is not None:
+                        opt["block_" + str(b)] = n
+                updown_opts.append(opt)
+            return updown_opts
+
+        def all_possible_dict_opts(unet, value):
+            """
+            Generate every possible combination for how a lora weight dict can be.
+            E.g. 2, {"unet: {"down": 2}}, {"unet: {"down": [2,2,2]}}, {"unet: {"mid": 2, "up": [2,2,2]}}, ...
+            """
+
+            down_blocks_with_tf = [i for i, d in enumerate(unet.down_blocks) if hasattr(d, "attentions")]
+            up_blocks_with_tf = [i for i, u in enumerate(unet.up_blocks) if hasattr(u, "attentions")]
+
+            layers_per_block = unet.config.layers_per_block
+
+            text_encoder_opts = [None, value]
+            text_encoder_2_opts = [None, value]
+            mid_opts = [None, value]
+            down_opts = [None] + updown_options(down_blocks_with_tf, layers_per_block, value)
+            up_opts = [None] + updown_options(up_blocks_with_tf, layers_per_block + 1, value)
+
+            opts = []
+
+            for t1, t2, d, m, u in product(text_encoder_opts, text_encoder_2_opts, down_opts, mid_opts, up_opts):
+                if all(o is None for o in (t1, t2, d, m, u)):
+                    continue
+                opt = {}
+                if t1 is not None:
+                    opt["text_encoder"] = t1
+                if t2 is not None:
+                    opt["text_encoder_2"] = t2
+                if all(o is None for o in (d, m, u)):
+                    # no unet scaling
+                    continue
+                opt["unet"] = {}
+                if d is not None:
+                    opt["unet"]["down"] = d
+                if m is not None:
+                    opt["unet"]["mid"] = m
+                if u is not None:
+                    opt["unet"]["up"] = u
+                opts.append(opt)
+
+            return opts
+
+        components, text_lora_config, unet_lora_config = self.get_dummy_components(self.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)
+
+        pipe.text_encoder.add_adapter(text_lora_config, "adapter-1")
+        pipe.unet.add_adapter(unet_lora_config, "adapter-1")
+
+        if self.has_two_text_encoders:
+            pipe.text_encoder_2.add_adapter(text_lora_config, "adapter-1")
+
+        for scale_dict in all_possible_dict_opts(pipe.unet, value=1234):
+            # test if lora block scales can be set with this scale_dict
+            if not self.has_two_text_encoders and "text_encoder_2" in scale_dict:
+                del scale_dict["text_encoder_2"]
+
+            pipe.set_adapters("adapter-1", scale_dict)  # test will fail if this line throws an error
+
     def test_simple_inference_with_text_unet_multi_adapter_delete_adapter(self):
         """
         Tests a simple inference with lora attached to text encoder and unet, attaches

tests/pipelines/animatediff/test_animatediff.py

@@ -131,6 +131,42 @@ class AnimateDiffPipelineFastTests(
     def test_attention_slicing_forward_pass(self):
         pass
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array(
+                [
+                    0.5541,
+                    0.5802,
+                    0.5074,
+                    0.4583,
+                    0.4729,
+                    0.5374,
+                    0.4051,
+                    0.4495,
+                    0.4480,
+                    0.5292,
+                    0.6322,
+                    0.6265,
+                    0.5455,
+                    0.4771,
+                    0.5795,
+                    0.5845,
+                    0.4172,
+                    0.6066,
+                    0.6535,
+                    0.4113,
+                    0.6833,
+                    0.5736,
+                    0.3589,
+                    0.5730,
+                    0.4205,
+                    0.3786,
+                    0.5323,
+                ]
+            )
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_inference_batch_single_identical(
         self,
         batch_size=2,
@@ -299,6 +335,9 @@ class AnimateDiffPipelineFastTests(
         max_diff = np.abs(to_np(output_with_offload) - to_np(output_without_offload)).max()
         self.assertLess(max_diff, 1e-4, "XFormers attention should not affect the inference results")
 
+    def test_vae_slicing(self):
+        return super().test_vae_slicing(image_count=2)
+
 
 @slow
 @require_torch_gpu

tests/pipelines/animatediff/test_animatediff_video2video.py

@@ -135,6 +135,34 @@ class AnimateDiffVideoToVideoPipelineFastTests(IPAdapterTesterMixin, PipelineTes
     def test_attention_slicing_forward_pass(self):
         pass
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array(
+                [
+                    0.4947,
+                    0.4780,
+                    0.4340,
+                    0.4666,
+                    0.4028,
+                    0.4645,
+                    0.4915,
+                    0.4101,
+                    0.4308,
+                    0.4581,
+                    0.3582,
+                    0.4953,
+                    0.4466,
+                    0.5348,
+                    0.5863,
+                    0.5299,
+                    0.5213,
+                    0.5017,
+                ]
+            )
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_inference_batch_single_identical(
         self,
         batch_size=2,

tests/pipelines/controlnet/test_controlnet.py

@@ -221,6 +221,12 @@ class ControlNetPipelineFastTests(
     def test_attention_slicing_forward_pass(self):
         return self._test_attention_slicing_forward_pass(expected_max_diff=2e-3)
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.5234, 0.3333, 0.1745, 0.7605, 0.6224, 0.4637, 0.6989, 0.7526, 0.4665])
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     @unittest.skipIf(
         torch_device != "cuda" or not is_xformers_available(),
         reason="XFormers attention is only available with CUDA and `xformers` installed",
@@ -455,6 +461,12 @@ class StableDiffusionMultiControlNetPipelineFastTests(
     def test_inference_batch_single_identical(self):
         self._test_inference_batch_single_identical(expected_max_diff=2e-3)
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.2422, 0.3425, 0.4048, 0.5351, 0.3503, 0.2419, 0.4645, 0.4570, 0.3804])
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_save_pretrained_raise_not_implemented_exception(self):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)
@@ -668,6 +680,12 @@ class StableDiffusionMultiControlNetOneModelPipelineFastTests(
     def test_inference_batch_single_identical(self):
         self._test_inference_batch_single_identical(expected_max_diff=2e-3)
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.5264, 0.3203, 0.1602, 0.8235, 0.6332, 0.4593, 0.7226, 0.7777, 0.4780])
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_save_pretrained_raise_not_implemented_exception(self):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)

tests/pipelines/controlnet/test_controlnet_img2img.py

@@ -174,6 +174,12 @@ class ControlNetImg2ImgPipelineFastTests(
     def test_attention_slicing_forward_pass(self):
         return self._test_attention_slicing_forward_pass(expected_max_diff=2e-3)
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.7096, 0.5149, 0.3571, 0.5897, 0.4715, 0.4052, 0.6098, 0.6886, 0.4213])
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     @unittest.skipIf(
         torch_device != "cuda" or not is_xformers_available(),
         reason="XFormers attention is only available with CUDA and `xformers` installed",
@@ -366,6 +372,12 @@ class StableDiffusionMultiControlNetPipelineFastTests(
     def test_inference_batch_single_identical(self):
         self._test_inference_batch_single_identical(expected_max_diff=2e-3)
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.5293, 0.7339, 0.6642, 0.3950, 0.5212, 0.5175, 0.7002, 0.5907, 0.5182])
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_save_pretrained_raise_not_implemented_exception(self):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)

tests/pipelines/controlnet/test_controlnet_sdxl.py

@@ -191,6 +191,15 @@ class StableDiffusionXLControlNetPipelineFastTests(
     def test_attention_slicing_forward_pass(self):
         return self._test_attention_slicing_forward_pass(expected_max_diff=2e-3)
 
+    def test_ip_adapter_single(self, from_ssd1b=False, expected_pipe_slice=None):
+        if not from_ssd1b:
+            expected_pipe_slice = None
+            if torch_device == "cpu":
+                expected_pipe_slice = np.array(
+                    [0.7331, 0.5907, 0.5667, 0.6029, 0.5679, 0.5968, 0.4033, 0.4761, 0.5090]
+                )
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     @unittest.skipIf(
         torch_device != "cuda" or not is_xformers_available(),
         reason="XFormers attention is only available with CUDA and `xformers` installed",
@@ -1047,6 +1056,12 @@ class StableDiffusionSSD1BControlNetPipelineFastTests(StableDiffusionXLControlNe
         # make sure that it's equal
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-4
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.6832, 0.5703, 0.5460, 0.6300, 0.5856, 0.6034, 0.4494, 0.4613, 0.5036])
+        return super().test_ip_adapter_single(from_ssd1b=True, expected_pipe_slice=expected_pipe_slice)
+
     def test_controlnet_sdxl_lcm(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
 

tests/pipelines/controlnet/test_controlnet_sdxl_img2img.py

@@ -170,6 +170,12 @@ class ControlNetPipelineSDXLImg2ImgFastTests(
 
         return inputs
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.6265, 0.5441, 0.5384, 0.5446, 0.5810, 0.5908, 0.5414, 0.5428, 0.5353])
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_stable_diffusion_xl_controlnet_img2img(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         components = self.get_dummy_components()

tests/pipelines/latent_consistency_models/test_latent_consistency_models.py

@@ -108,6 +108,12 @@ class LatentConsistencyModelPipelineFastTests(
         }
         return inputs
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.1403, 0.5072, 0.5316, 0.1202, 0.3865, 0.4211, 0.5363, 0.3557, 0.3645])
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_lcm_onestep(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
 

tests/pipelines/latent_consistency_models/test_latent_consistency_models_img2img.py

@@ -119,6 +119,12 @@ class LatentConsistencyModelImg2ImgPipelineFastTests(
         }
         return inputs
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.4003, 0.3718, 0.2863, 0.5500, 0.5587, 0.3772, 0.4617, 0.4961, 0.4417])
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_lcm_onestep(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
 

tests/pipelines/pia/test_pia.py

@@ -138,6 +138,43 @@ class PIAPipelineFastTests(IPAdapterTesterMixin, PipelineTesterMixin, unittest.T
 
         assert isinstance(pipe.unet, UNetMotionModel)
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array(
+                [
+                    0.5609,
+                    0.5756,
+                    0.4830,
+                    0.4420,
+                    0.4547,
+                    0.5129,
+                    0.3779,
+                    0.4042,
+                    0.3772,
+                    0.4450,
+                    0.5710,
+                    0.5536,
+                    0.4835,
+                    0.4308,
+                    0.5578,
+                    0.5578,
+                    0.4395,
+                    0.5440,
+                    0.6051,
+                    0.4651,
+                    0.6258,
+                    0.5662,
+                    0.3988,
+                    0.5108,
+                    0.4153,
+                    0.3993,
+                    0.4803,
+                ]
+            )
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     @unittest.skip("Attention slicing is not enabled in this pipeline")
     def test_attention_slicing_forward_pass(self):
         pass

tests/pipelines/stable_diffusion/test_stable_diffusion.py

@@ -370,6 +370,12 @@ class StableDiffusionPipelineFastTests(
 
         assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.3203, 0.4555, 0.4711, 0.3505, 0.3973, 0.4650, 0.5137, 0.3392, 0.4045])
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_stable_diffusion_ddim_factor_8(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
 

tests/pipelines/stable_diffusion/test_stable_diffusion_img2img.py

@@ -253,6 +253,12 @@ class StableDiffusionImg2ImgPipelineFastTests(
 
         assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.4932, 0.5092, 0.5135, 0.5517, 0.5626, 0.6621, 0.6490, 0.5021, 0.5441])
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_stable_diffusion_img2img_multiple_init_images(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         components = self.get_dummy_components()

tests/pipelines/stable_diffusion/test_stable_diffusion_inpaint.py

@@ -388,6 +388,15 @@ class StableDiffusionInpaintPipelineFastTests(
         # they should be the same
         assert torch.allclose(intermediate_latent, output_interrupted, atol=1e-4)
 
+    def test_ip_adapter_single(self, from_simple=False, expected_pipe_slice=None):
+        if not from_simple:
+            expected_pipe_slice = None
+            if torch_device == "cpu":
+                expected_pipe_slice = np.array(
+                    [0.4390, 0.5452, 0.3772, 0.5448, 0.6031, 0.4480, 0.5194, 0.4687, 0.4640]
+                )
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
 
 class StableDiffusionSimpleInpaintPipelineFastTests(StableDiffusionInpaintPipelineFastTests):
     pipeline_class = StableDiffusionInpaintPipeline
@@ -475,6 +484,12 @@ class StableDiffusionSimpleInpaintPipelineFastTests(StableDiffusionInpaintPipeli
         }
         return inputs
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.6345, 0.5395, 0.5611, 0.5403, 0.5830, 0.5855, 0.5193, 0.5443, 0.5211])
+        return super().test_ip_adapter_single(from_simple=True, expected_pipe_slice=expected_pipe_slice)
+
     def test_stable_diffusion_inpaint(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         components = self.get_dummy_components()

tests/pipelines/stable_diffusion_2/test_stable_diffusion_attend_and_excite.py

@@ -185,6 +185,9 @@ class StableDiffusionAttendAndExcitePipelineFastTests(
     def test_save_load_optional_components(self):
         super().test_save_load_optional_components(expected_max_difference=4e-4)
 
+    def test_karras_schedulers_shape(self):
+        super().test_karras_schedulers_shape(num_inference_steps_for_strength_for_iterations=3)
+
 
 @require_torch_gpu
 @nightly

tests/pipelines/stable_diffusion_sag/test_stable_diffusion_sag.py

@@ -32,13 +32,15 @@ from diffusers import (
 from diffusers.utils.testing_utils import enable_full_determinism, nightly, require_torch_gpu, 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 PipelineLatentTesterMixin, PipelineTesterMixin
+from ..test_pipelines_common import IPAdapterTesterMixin, PipelineLatentTesterMixin, PipelineTesterMixin
 
 
 enable_full_determinism()
 
 
-class StableDiffusionSAGPipelineFastTests(PipelineLatentTesterMixin, PipelineTesterMixin, unittest.TestCase):
+class StableDiffusionSAGPipelineFastTests(
+    IPAdapterTesterMixin, PipelineLatentTesterMixin, PipelineTesterMixin, unittest.TestCase
+):
     pipeline_class = StableDiffusionSAGPipeline
     params = TEXT_TO_IMAGE_PARAMS
     batch_params = TEXT_TO_IMAGE_BATCH_PARAMS

tests/pipelines/stable_diffusion_xl/test_stable_diffusion_xl.py

@@ -292,6 +292,12 @@ class StableDiffusionXLPipelineFastTests(
         # make sure that it's equal
         assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.5552, 0.5569, 0.4725, 0.4348, 0.4994, 0.4632, 0.5142, 0.5012, 0.4700])
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_attention_slicing_forward_pass(self):
         super().test_attention_slicing_forward_pass(expected_max_diff=3e-3)
 

tests/pipelines/stable_diffusion_xl/test_stable_diffusion_xl_adapter.py

@@ -294,6 +294,15 @@ class StableDiffusionXLAdapterPipelineFastTests(
         }
         return inputs
 
+    def test_ip_adapter_single(self, from_multi=False, expected_pipe_slice=None):
+        if not from_multi:
+            expected_pipe_slice = None
+            if torch_device == "cpu":
+                expected_pipe_slice = np.array(
+                    [0.5753, 0.6022, 0.4728, 0.4986, 0.5708, 0.4645, 0.5194, 0.5134, 0.4730]
+                )
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_stable_diffusion_adapter_default_case(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         components = self.get_dummy_components()
@@ -446,6 +455,12 @@ class StableDiffusionXLMultiAdapterPipelineFastTests(
         )
         assert np.abs(image_slice.flatten() - expected_slice).max() < 5e-3
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.5813, 0.6100, 0.4756, 0.5057, 0.5720, 0.4632, 0.5177, 0.5125, 0.4718])
+        return super().test_ip_adapter_single(from_multi=True, expected_pipe_slice=expected_pipe_slice)
+
     def test_inference_batch_consistent(
         self, batch_sizes=[2, 4, 13], additional_params_copy_to_batched_inputs=["num_inference_steps"]
     ):

tests/pipelines/stable_diffusion_xl/test_stable_diffusion_xl_img2img.py

@@ -311,6 +311,12 @@ class StableDiffusionXLImg2ImgPipelineFastTests(
         # make sure that it's equal
         assert np.abs(image_slice_1.flatten() - image_slice_2.flatten()).max() < 1e-4
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.5174, 0.4512, 0.5006, 0.6273, 0.5160, 0.6825, 0.6655, 0.5840, 0.5675])
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_stable_diffusion_xl_img2img_tiny_autoencoder(self):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         components = self.get_dummy_components()

tests/pipelines/stable_diffusion_xl/test_stable_diffusion_xl_inpaint.py

@@ -223,6 +223,12 @@ class StableDiffusionXLInpaintPipelineFastTests(
         }
         return inputs
 
+    def test_ip_adapter_single(self):
+        expected_pipe_slice = None
+        if torch_device == "cpu":
+            expected_pipe_slice = np.array([0.7971, 0.5371, 0.5973, 0.5642, 0.6689, 0.6894, 0.5770, 0.6063, 0.5261])
+        return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
     def test_components_function(self):
         init_components = self.get_dummy_components()
         init_components.pop("requires_aesthetics_score")

tests/pipelines/test_pipelines_common.py

@@ -37,11 +37,7 @@ from diffusers.pipelines.pipeline_utils import StableDiffusionMixin
 from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import logging
 from diffusers.utils.import_utils import is_accelerate_available, is_accelerate_version, is_xformers_available
-from diffusers.utils.testing_utils import (
-    CaptureLogger,
-    require_torch,
-    torch_device,
-)
+from diffusers.utils.testing_utils import CaptureLogger, require_torch, torch_device
 
 from ..models.autoencoders.test_models_vae import (
     get_asym_autoencoder_kl_config,
@@ -71,7 +67,7 @@ class SDFunctionTesterMixin:
     It provides a set of common tests for PyTorch pipeline that inherit from StableDiffusionMixin, e.g. vae_slicing, vae_tiling, freeu, etc.
     """
 
-    def test_vae_slicing(self):
+    def test_vae_slicing(self, image_count=4):
         device = "cpu"  # ensure determinism for the device-dependent torch.Generator
         components = self.get_dummy_components()
         # components["scheduler"] = LMSDiscreteScheduler.from_config(components["scheduler"].config)
@@ -79,8 +75,6 @@ class SDFunctionTesterMixin:
         pipe = pipe.to(device)
         pipe.set_progress_bar_config(disable=None)
 
-        image_count = 4
-
         inputs = self.get_dummy_inputs(device)
         inputs["prompt"] = [inputs["prompt"]] * image_count
         if "image" in inputs:  # fix batch size mismatch in I2V_Gen pipeline
@@ -241,7 +235,11 @@ class IPAdapterTesterMixin:
         inputs["return_dict"] = False
         return inputs
 
-    def test_ip_adapter_single(self, expected_max_diff: float = 1e-4):
+    def test_ip_adapter_single(self, expected_max_diff: float = 1e-4, expected_pipe_slice=None):
+        # Raising the tolerance for this test when it's run on a CPU because we
+        # compare against static slices and that can be shaky (with a VVVV low probability).
+        expected_max_diff = 9e-4 if torch_device == "cpu" else expected_max_diff
+
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components).to(torch_device)
         pipe.set_progress_bar_config(disable=None)
@@ -249,7 +247,10 @@ class IPAdapterTesterMixin:
 
         # forward pass without ip adapter
         inputs = self._modify_inputs_for_ip_adapter_test(self.get_dummy_inputs(torch_device))
-        output_without_adapter = pipe(**inputs)[0]
+        if expected_pipe_slice is None:
+            output_without_adapter = pipe(**inputs)[0]
+        else:
+            output_without_adapter = expected_pipe_slice
 
         adapter_state_dict = create_ip_adapter_state_dict(pipe.unet)
         pipe.unet._load_ip_adapter_weights(adapter_state_dict)
@@ -259,12 +260,16 @@ class IPAdapterTesterMixin:
         inputs["ip_adapter_image_embeds"] = [self._get_dummy_image_embeds(cross_attention_dim)]
         pipe.set_ip_adapter_scale(0.0)
         output_without_adapter_scale = pipe(**inputs)[0]
+        if expected_pipe_slice is not None:
+            output_without_adapter_scale = output_without_adapter_scale[0, -3:, -3:, -1].flatten()
 
         # forward pass with single ip adapter, but with scale of adapter weights
         inputs = self._modify_inputs_for_ip_adapter_test(self.get_dummy_inputs(torch_device))
         inputs["ip_adapter_image_embeds"] = [self._get_dummy_image_embeds(cross_attention_dim)]
         pipe.set_ip_adapter_scale(42.0)
         output_with_adapter_scale = pipe(**inputs)[0]
+        if expected_pipe_slice is not None:
+            output_with_adapter_scale = output_with_adapter_scale[0, -3:, -3:, -1].flatten()
 
         max_diff_without_adapter_scale = np.abs(output_without_adapter_scale - output_without_adapter).max()
         max_diff_with_adapter_scale = np.abs(output_with_adapter_scale - output_without_adapter).max()
@@ -514,7 +519,9 @@ class PipelineKarrasSchedulerTesterMixin:
     equivalence of dict and tuple outputs, etc.
     """
 
-    def test_karras_schedulers_shape(self):
+    def test_karras_schedulers_shape(
+        self, num_inference_steps_for_strength=4, num_inference_steps_for_strength_for_iterations=5
+    ):
         components = self.get_dummy_components()
         pipe = self.pipeline_class(**components)
 
@@ -527,13 +534,13 @@ class PipelineKarrasSchedulerTesterMixin:
         inputs["num_inference_steps"] = 2
 
         if "strength" in inputs:
-            inputs["num_inference_steps"] = 4
+            inputs["num_inference_steps"] = num_inference_steps_for_strength
             inputs["strength"] = 0.5
 
         outputs = []
         for scheduler_enum in KarrasDiffusionSchedulers:
             if "KDPM2" in scheduler_enum.name:
-                inputs["num_inference_steps"] = 5
+                inputs["num_inference_steps"] = num_inference_steps_for_strength_for_iterations
 
             scheduler_cls = getattr(diffusers, scheduler_enum.name)
             pipe.scheduler = scheduler_cls.from_config(pipe.scheduler.config)
@@ -1137,20 +1144,24 @@ class PipelineTesterMixin:
         self.assertLess(
             max_diff, expected_max_diff, "running CPU offloading 2nd time should not affect the inference results"
         )
-        offloaded_modules = [
-            v
+        offloaded_modules = {
+            k: v
             for k, v in pipe.components.items()
             if isinstance(v, torch.nn.Module) and k not in pipe._exclude_from_cpu_offload
-        ]
-        (
-            self.assertTrue(all(v.device.type == "cpu" for v in offloaded_modules)),
-            f"Not offloaded: {[v for v in offloaded_modules if v.device.type != 'cpu']}",
+        }
+        self.assertTrue(
+            all(v.device.type == "cpu" for v in offloaded_modules.values()),
+            f"Not offloaded: {[k for k, v in offloaded_modules.items() if v.device.type != 'cpu']}",
         )
 
-        offloaded_modules_with_hooks = [v for v in offloaded_modules if hasattr(v, "_hf_hook")]
-        (
-            self.assertTrue(all(isinstance(v, accelerate.hooks.CpuOffload) for v in offloaded_modules_with_hooks)),
-            f"Not installed correct hook: {[v for v in offloaded_modules_with_hooks if not isinstance(v, accelerate.hooks.CpuOffload)]}",
+        offloaded_modules_with_incorrect_hooks = {}
+        for k, v in offloaded_modules.items():
+            if hasattr(v, "_hf_hook") and not isinstance(v._hf_hook, accelerate.hooks.CpuOffload):
+                offloaded_modules_with_incorrect_hooks[k] = type(v._hf_hook)
+
+        self.assertTrue(
+            len(offloaded_modules_with_incorrect_hooks) == 0,
+            f"Not installed correct hook: {offloaded_modules_with_incorrect_hooks}",
         )
 
     @unittest.skipIf(
@@ -1174,7 +1185,7 @@ class PipelineTesterMixin:
         inputs = self.get_dummy_inputs(generator_device)
         output_with_offload = pipe(**inputs)[0]
 
-        pipe.nable_sequential_cpu_offload()
+        pipe.enable_sequential_cpu_offload()
         inputs = self.get_dummy_inputs(generator_device)
         output_with_offload_twice = pipe(**inputs)[0]
 
@@ -1182,22 +1193,23 @@ class PipelineTesterMixin:
         self.assertLess(
             max_diff, expected_max_diff, "running sequential offloading second time should have the inference results"
         )
-        offloaded_modules = [
-            v
+        offloaded_modules = {
+            k: v
             for k, v in pipe.components.items()
             if isinstance(v, torch.nn.Module) and k not in pipe._exclude_from_cpu_offload
-        ]
-        (
-            self.assertTrue(all(v.device.type == "meta" for v in offloaded_modules)),
-            f"Not offloaded: {[v for v in offloaded_modules if v.device.type != 'meta']}",
+        }
+        self.assertTrue(
+            all(v.device.type == "meta" for v in offloaded_modules.values()),
+            f"Not offloaded: {[k for k, v in offloaded_modules.items() if v.device.type != 'meta']}",
         )
+        offloaded_modules_with_incorrect_hooks = {}
+        for k, v in offloaded_modules.items():
+            if hasattr(v, "_hf_hook") and not isinstance(v._hf_hook, accelerate.hooks.AlignDevicesHook):
+                offloaded_modules_with_incorrect_hooks[k] = type(v._hf_hook)
 
-        offloaded_modules_with_hooks = [v for v in offloaded_modules if hasattr(v, "_hf_hook")]
-        (
-            self.assertTrue(
-                all(isinstance(v, accelerate.hooks.AlignDevicesHook) for v in offloaded_modules_with_hooks)
-            ),
-            f"Not installed correct hook: {[v for v in offloaded_modules_with_hooks if not isinstance(v, accelerate.hooks.AlignDevicesHook)]}",
+        self.assertTrue(
+            len(offloaded_modules_with_incorrect_hooks) == 0,
+            f"Not installed correct hook: {offloaded_modules_with_incorrect_hooks}",
         )
 
     @unittest.skipIf(

tests/schedulers/test_scheduler_unipc.py

@@ -24,6 +24,7 @@ class UniPCMultistepSchedulerTest(SchedulerCommonTest):
             "beta_schedule": "linear",
             "solver_order": 2,
             "solver_type": "bh2",
+            "final_sigmas_type": "sigma_min",
         }
 
         config.update(**kwargs)