Merge branch 'main' into test-fixes

* move attend and excite out.

* fix: import

* fix diffedit

.github/workflows/push_tests_fast.yml

@@ -98,6 +98,7 @@ jobs:
     - name: Run example PyTorch CPU tests
       if: ${{ matrix.config.framework == 'pytorch_examples' }}
       run: |
+        python -m pip install peft
         python -m pytest -n 2 --max-worker-restart=0 --dist=loadfile \
           --make-reports=tests_${{ matrix.config.report }} \
           examples

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

@@ -203,7 +203,7 @@ def make_inpaint_condition(image, image_mask):
     image_mask = np.array(image_mask.convert("L")).astype(np.float32) / 255.0
 
     assert image.shape[0:1] == image_mask.shape[0:1]
-    image[image_mask > 0.5] = 1.0  # set as masked pixel
+    image[image_mask > 0.5] = -1.0  # set as masked pixel
     image = np.expand_dims(image, 0).transpose(0, 3, 1, 2)
     image = torch.from_numpy(image)
     return image

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

@@ -41,6 +41,20 @@ Now, define four different `Generator`s and assign each `Generator` a seed (`0`
 generator = [torch.Generator(device="cuda").manual_seed(i) for i in range(4)]
 ```
 
+<Tip warning={true}>
+
+To create a batched seed, you should use a list comprehension that iterates over the length specified in `range()`. This creates a unique `Generator` object for each image in the batch. If you only multiply the `Generator` by the batch size, this only creates one `Generator` object that is used sequentially for each image in the batch.
+
+For example, if you want to use the same seed to create 4 identical images:
+
+```py
+❌ [torch.Generator().manual_seed(seed)] * 4
+
+✅ [torch.Generator().manual_seed(seed) for _ in range(4)]
+```
+
+</Tip>
+
 Generate the images and have a look:
 
 ```python

examples/community/regional_prompting_stable_diffusion.py

@@ -73,7 +73,14 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
         requires_safety_checker: bool = True,
     ):
         super().__init__(
-            vae, text_encoder, tokenizer, unet, scheduler, safety_checker, feature_extractor, requires_safety_checker
+            vae,
+            text_encoder,
+            tokenizer,
+            unet,
+            scheduler,
+            safety_checker,
+            feature_extractor,
+            requires_safety_checker,
         )
         self.register_modules(
             vae=vae,
@@ -102,22 +109,22 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
         return_dict: bool = True,
         rp_args: Dict[str, str] = None,
     ):
-        active = KBRK in prompt[0] if type(prompt) == list else KBRK in prompt  # noqa: E721
+        active = KBRK in prompt[0] if isinstance(prompt, list) else KBRK in prompt
         if negative_prompt is None:
-            negative_prompt = "" if type(prompt) == str else [""] * len(prompt)  # noqa: E721
+            negative_prompt = "" if isinstance(prompt, str) else [""] * len(prompt)
 
         device = self._execution_device
         regions = 0
 
         self.power = int(rp_args["power"]) if "power" in rp_args else 1
 
-        prompts = prompt if type(prompt) == list else [prompt]  # noqa: E721
-        n_prompts = negative_prompt if type(negative_prompt) == list else [negative_prompt]  # noqa: E721
+        prompts = prompt if isinstance(prompt, list) else [prompt]
+        n_prompts = negative_prompt if isinstance(prompt, str) else [negative_prompt]
         self.batch = batch = num_images_per_prompt * len(prompts)
         all_prompts_cn, all_prompts_p = promptsmaker(prompts, num_images_per_prompt)
         all_n_prompts_cn, _ = promptsmaker(n_prompts, num_images_per_prompt)
 
-        cn = len(all_prompts_cn) == len(all_n_prompts_cn)
+        equal = len(all_prompts_cn) == len(all_n_prompts_cn)
 
         if Compel:
             compel = Compel(tokenizer=self.tokenizer, text_encoder=self.text_encoder)
@@ -129,7 +136,7 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
                 return torch.cat(embl)
 
             conds = getcompelembs(all_prompts_cn)
-            unconds = getcompelembs(all_n_prompts_cn) if cn else getcompelembs(n_prompts)
+            unconds = getcompelembs(all_n_prompts_cn)
             embs = getcompelembs(prompts)
             n_embs = getcompelembs(n_prompts)
             prompt = negative_prompt = None
@@ -137,7 +144,7 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
             conds = self.encode_prompt(prompts, device, 1, True)[0]
             unconds = (
                 self.encode_prompt(n_prompts, device, 1, True)[0]
-                if cn
+                if equal
                 else self.encode_prompt(all_n_prompts_cn, device, 1, True)[0]
             )
             embs = n_embs = None
@@ -206,8 +213,11 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
                     else:
                         px, nx = hidden_states.chunk(2)
 
-                    if cn:
-                        hidden_states = torch.cat([px for i in range(regions)] + [nx for i in range(regions)], 0)
+                    if equal:
+                        hidden_states = torch.cat(
+                            [px for i in range(regions)] + [nx for i in range(regions)],
+                            0,
+                        )
                         encoder_hidden_states = torch.cat([conds] + [unconds])
                     else:
                         hidden_states = torch.cat([px for i in range(regions)] + [nx], 0)
@@ -289,9 +299,9 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
                     if any(x in mode for x in ["COL", "ROW"]):
                         reshaped = hidden_states.reshape(hidden_states.size()[0], h, w, hidden_states.size()[2])
                         center = reshaped.shape[0] // 2
-                        px = reshaped[0:center] if cn else reshaped[0:-batch]
-                        nx = reshaped[center:] if cn else reshaped[-batch:]
-                        outs = [px, nx] if cn else [px]
+                        px = reshaped[0:center] if equal else reshaped[0:-batch]
+                        nx = reshaped[center:] if equal else reshaped[-batch:]
+                        outs = [px, nx] if equal else [px]
                         for out in outs:
                             c = 0
                             for i, ocell in enumerate(ocells):
@@ -321,15 +331,16 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
                                             :,
                                         ]
                                     c += 1
-                        px, nx = (px[0:batch], nx[0:batch]) if cn else (px[0:batch], nx)
+                        px, nx = (px[0:batch], nx[0:batch]) if equal else (px[0:batch], nx)
                         hidden_states = torch.cat([nx, px], 0) if revers else torch.cat([px, nx], 0)
                         hidden_states = hidden_states.reshape(xshape)
 
                     #### Regional Prompting Prompt mode
                     elif "PRO" in mode:
-                        center = reshaped.shape[0] // 2
-                        px = reshaped[0:center] if cn else reshaped[0:-batch]
-                        nx = reshaped[center:] if cn else reshaped[-batch:]
+                        px, nx = (
+                            torch.chunk(hidden_states) if equal else hidden_states[0:-batch],
+                            hidden_states[-batch:],
+                        )
 
                         if (h, w) in self.attnmasks and self.maskready:
 
@@ -340,8 +351,8 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
                                         out[b] = out[b] + out[r * batch + b]
                                 return out
 
-                            px, nx = (mask(px), mask(nx)) if cn else (mask(px), nx)
-                        px, nx = (px[0:batch], nx[0:batch]) if cn else (px[0:batch], nx)
+                            px, nx = (mask(px), mask(nx)) if equal else (mask(px), nx)
+                        px, nx = (px[0:batch], nx[0:batch]) if equal else (px[0:batch], nx)
                         hidden_states = torch.cat([nx, px], 0) if revers else torch.cat([px, nx], 0)
                     return hidden_states
 
@@ -378,7 +389,15 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
             save_mask = False
 
         if mode == "PROMPT" and save_mask:
-            saveattnmaps(self, output, height, width, thresholds, num_inference_steps // 2, regions)
+            saveattnmaps(
+                self,
+                output,
+                height,
+                width,
+                thresholds,
+                num_inference_steps // 2,
+                regions,
+            )
 
         return output
 
@@ -437,7 +456,11 @@ def make_cells(ratios):
 def make_emblist(self, prompts):
     with torch.no_grad():
         tokens = self.tokenizer(
-            prompts, max_length=self.tokenizer.model_max_length, padding=True, truncation=True, return_tensors="pt"
+            prompts,
+            max_length=self.tokenizer.model_max_length,
+            padding=True,
+            truncation=True,
+            return_tensors="pt",
         ).input_ids.to(self.device)
         embs = self.text_encoder(tokens, output_hidden_states=True).last_hidden_state.to(self.device, dtype=self.dtype)
     return embs
@@ -563,7 +586,15 @@ def tokendealer(self, all_prompts):
 
 
 def scaled_dot_product_attention(
-    self, query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, scale=None, getattn=False
+    self,
+    query,
+    key,
+    value,
+    attn_mask=None,
+    dropout_p=0.0,
+    is_causal=False,
+    scale=None,
+    getattn=False,
 ) -> torch.Tensor:
     # Efficient implementation equivalent to the following:
     L, S = query.size(-2), key.size(-2)

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -991,6 +991,17 @@ def main(args):
         text_encoder_one.add_adapter(text_lora_config)
         text_encoder_two.add_adapter(text_lora_config)
 
+    # Make sure the trainable params are in float32.
+    if args.mixed_precision == "fp16":
+        models = [unet]
+        if args.train_text_encoder:
+            models.extend([text_encoder_one, text_encoder_two])
+        for model in models:
+            for param in model.parameters():
+                # only upcast trainable parameters (LoRA) into fp32
+                if param.requires_grad:
+                    param.data = param.to(torch.float32)
+
     # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
     def save_model_hook(models, weights, output_dir):
         if accelerator.is_main_process:

examples/research_projects/realfill/requirements.txt

@@ -1,6 +1,6 @@
 diffusers==0.20.1
 accelerate==0.23.0
-transformers==4.34.0
+transformers==4.36.0
 peft==0.5.0
 torch==2.0.1
 torchvision>=0.16

examples/text_to_image/train_text_to_image_lora.py

@@ -460,7 +460,13 @@ def main():
     vae.to(accelerator.device, dtype=weight_dtype)
     text_encoder.to(accelerator.device, dtype=weight_dtype)
 
+    # Add adapter and make sure the trainable params are in float32.
     unet.add_adapter(unet_lora_config)
+    if args.mixed_precision == "fp16":
+        for param in unet.parameters():
+            # only upcast trainable parameters (LoRA) into fp32
+            if param.requires_grad:
+                param.data = param.to(torch.float32)
 
     if args.enable_xformers_memory_efficient_attention:
         if is_xformers_available():
@@ -888,39 +894,42 @@ def main():
                 ignore_patterns=["step_*", "epoch_*"],
             )
 
-    # Final inference
-    # Load previous pipeline
-    pipeline = DiffusionPipeline.from_pretrained(
-        args.pretrained_model_name_or_path, revision=args.revision, variant=args.variant, torch_dtype=weight_dtype
-    )
-    pipeline = pipeline.to(accelerator.device)
+        # Final inference
+        # Load previous pipeline
+        if args.validation_prompt is not None:
+            pipeline = DiffusionPipeline.from_pretrained(
+                args.pretrained_model_name_or_path,
+                revision=args.revision,
+                variant=args.variant,
+                torch_dtype=weight_dtype,
+            )
+            pipeline = pipeline.to(accelerator.device)
 
-    # load attention processors
-    pipeline.unet.load_attn_procs(args.output_dir)
+            # load attention processors
+            pipeline.load_lora_weights(args.output_dir)
 
-    # run inference
-    generator = torch.Generator(device=accelerator.device)
-    if args.seed is not None:
-        generator = generator.manual_seed(args.seed)
-    images = []
-    for _ in range(args.num_validation_images):
-        images.append(pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0])
+            # run inference
+            generator = torch.Generator(device=accelerator.device)
+            if args.seed is not None:
+                generator = generator.manual_seed(args.seed)
+            images = []
+            for _ in range(args.num_validation_images):
+                images.append(pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0])
 
-    if accelerator.is_main_process:
-        for tracker in accelerator.trackers:
-            if len(images) != 0:
-                if tracker.name == "tensorboard":
-                    np_images = np.stack([np.asarray(img) for img in images])
-                    tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
-                if tracker.name == "wandb":
-                    tracker.log(
-                        {
-                            "test": [
-                                wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
-                                for i, image in enumerate(images)
-                            ]
-                        }
-                    )
+            for tracker in accelerator.trackers:
+                if len(images) != 0:
+                    if tracker.name == "tensorboard":
+                        np_images = np.stack([np.asarray(img) for img in images])
+                        tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
+                    if tracker.name == "wandb":
+                        tracker.log(
+                            {
+                                "test": [
+                                    wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
+                                    for i, image in enumerate(images)
+                                ]
+                            }
+                        )
 
     accelerator.end_training()
 

examples/text_to_image/train_text_to_image_lora_sdxl.py

@@ -22,7 +22,6 @@ import os
 import random
 import shutil
 from pathlib import Path
-from typing import Dict
 
 import datasets
 import numpy as np
@@ -436,22 +435,6 @@ DATASET_NAME_MAPPING = {
 }
 
 
-def unet_attn_processors_state_dict(unet) -> Dict[str, torch.tensor]:
-    """
-    Returns:
-        a state dict containing just the attention processor parameters.
-    """
-    attn_processors = unet.attn_processors
-
-    attn_processors_state_dict = {}
-
-    for attn_processor_key, attn_processor in attn_processors.items():
-        for parameter_key, parameter in attn_processor.state_dict().items():
-            attn_processors_state_dict[f"{attn_processor_key}.{parameter_key}"] = parameter
-
-    return attn_processors_state_dict
-
-
 def tokenize_prompt(tokenizer, prompt):
     text_inputs = tokenizer(
         prompt,
@@ -640,6 +623,17 @@ def main(args):
         text_encoder_one.add_adapter(text_lora_config)
         text_encoder_two.add_adapter(text_lora_config)
 
+    # Make sure the trainable params are in float32.
+    if args.mixed_precision == "fp16":
+        models = [unet]
+        if args.train_text_encoder:
+            models.extend([text_encoder_one, text_encoder_two])
+        for model in models:
+            for param in model.parameters():
+                # only upcast trainable parameters (LoRA) into fp32
+                if param.requires_grad:
+                    param.data = param.to(torch.float32)
+
     # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
     def save_model_hook(models, weights, output_dir):
         if accelerator.is_main_process:
@@ -1187,6 +1181,9 @@ def main(args):
         torch.cuda.empty_cache()
 
         # Final inference
+        # Make sure vae.dtype is consistent with the unet.dtype
+        if args.mixed_precision == "fp16":
+            vae.to(weight_dtype)
         # Load previous pipeline
         pipeline = StableDiffusionXLPipeline.from_pretrained(
             args.pretrained_model_name_or_path,

src/diffusers/models/controlnetxs.py

@@ -23,9 +23,7 @@ from torch.nn.modules.normalization import GroupNorm
 
 from ..configuration_utils import ConfigMixin, register_to_config
 from ..utils import BaseOutput, logging
-from .attention_processor import (
-    AttentionProcessor,
-)
+from .attention_processor import USE_PEFT_BACKEND, AttentionProcessor
 from .autoencoders import AutoencoderKL
 from .lora import LoRACompatibleConv
 from .modeling_utils import ModelMixin
@@ -817,11 +815,23 @@ def increase_block_input_in_encoder_resnet(unet: UNet2DConditionModel, block_no,
     norm_kwargs = {a: getattr(old_norm1, a) for a in norm_args}
     norm_kwargs["num_channels"] += by  # surgery done here
     # conv1
-    conv1_args = (
-        "in_channels out_channels kernel_size stride padding dilation groups bias padding_mode lora_layer".split(" ")
-    )
+    conv1_args = [
+        "in_channels",
+        "out_channels",
+        "kernel_size",
+        "stride",
+        "padding",
+        "dilation",
+        "groups",
+        "bias",
+        "padding_mode",
+    ]
+    if not USE_PEFT_BACKEND:
+        conv1_args.append("lora_layer")
+
     for a in conv1_args:
         assert hasattr(old_conv1, a)
+
     conv1_kwargs = {a: getattr(old_conv1, a) for a in conv1_args}
     conv1_kwargs["bias"] = "bias" in conv1_kwargs  # as param, bias is a boolean, but as attr, it's a tensor.
     conv1_kwargs["in_channels"] += by  # surgery done here
@@ -839,25 +849,42 @@ def increase_block_input_in_encoder_resnet(unet: UNet2DConditionModel, block_no,
     }
     # swap old with new modules
     unet.down_blocks[block_no].resnets[resnet_idx].norm1 = GroupNorm(**norm_kwargs)
-    unet.down_blocks[block_no].resnets[resnet_idx].conv1 = LoRACompatibleConv(**conv1_kwargs)
-    unet.down_blocks[block_no].resnets[resnet_idx].conv_shortcut = LoRACompatibleConv(**conv_shortcut_args_kwargs)
+    unet.down_blocks[block_no].resnets[resnet_idx].conv1 = (
+        nn.Conv2d(**conv1_kwargs) if USE_PEFT_BACKEND else LoRACompatibleConv(**conv1_kwargs)
+    )
+    unet.down_blocks[block_no].resnets[resnet_idx].conv_shortcut = (
+        nn.Conv2d(**conv_shortcut_args_kwargs) if USE_PEFT_BACKEND else LoRACompatibleConv(**conv_shortcut_args_kwargs)
+    )
     unet.down_blocks[block_no].resnets[resnet_idx].in_channels += by  # surgery done here
 
 
 def increase_block_input_in_encoder_downsampler(unet: UNet2DConditionModel, block_no, by):
     """Increase channels sizes to allow for additional concatted information from base model"""
     old_down = unet.down_blocks[block_no].downsamplers[0].conv
-    # conv1
-    args = "in_channels out_channels kernel_size stride padding dilation groups bias padding_mode lora_layer".split(
-        " "
-    )
+
+    args = [
+        "in_channels",
+        "out_channels",
+        "kernel_size",
+        "stride",
+        "padding",
+        "dilation",
+        "groups",
+        "bias",
+        "padding_mode",
+    ]
+    if not USE_PEFT_BACKEND:
+        args.append("lora_layer")
+
     for a in args:
         assert hasattr(old_down, a)
     kwargs = {a: getattr(old_down, a) for a in args}
     kwargs["bias"] = "bias" in kwargs  # as param, bias is a boolean, but as attr, it's a tensor.
     kwargs["in_channels"] += by  # surgery done here
     # swap old with new modules
-    unet.down_blocks[block_no].downsamplers[0].conv = LoRACompatibleConv(**kwargs)
+    unet.down_blocks[block_no].downsamplers[0].conv = (
+        nn.Conv2d(**kwargs) if USE_PEFT_BACKEND else LoRACompatibleConv(**kwargs)
+    )
     unet.down_blocks[block_no].downsamplers[0].channels += by  # surgery done here
 
 
@@ -871,12 +898,20 @@ def increase_block_input_in_mid_resnet(unet: UNet2DConditionModel, by):
         assert hasattr(old_norm1, a)
     norm_kwargs = {a: getattr(old_norm1, a) for a in norm_args}
     norm_kwargs["num_channels"] += by  # surgery done here
-    # conv1
-    conv1_args = (
-        "in_channels out_channels kernel_size stride padding dilation groups bias padding_mode lora_layer".split(" ")
-    )
-    for a in conv1_args:
-        assert hasattr(old_conv1, a)
+    conv1_args = [
+        "in_channels",
+        "out_channels",
+        "kernel_size",
+        "stride",
+        "padding",
+        "dilation",
+        "groups",
+        "bias",
+        "padding_mode",
+    ]
+    if not USE_PEFT_BACKEND:
+        conv1_args.append("lora_layer")
+
     conv1_kwargs = {a: getattr(old_conv1, a) for a in conv1_args}
     conv1_kwargs["bias"] = "bias" in conv1_kwargs  # as param, bias is a boolean, but as attr, it's a tensor.
     conv1_kwargs["in_channels"] += by  # surgery done here
@@ -894,8 +929,12 @@ def increase_block_input_in_mid_resnet(unet: UNet2DConditionModel, by):
     }
     # swap old with new modules
     unet.mid_block.resnets[0].norm1 = GroupNorm(**norm_kwargs)
-    unet.mid_block.resnets[0].conv1 = LoRACompatibleConv(**conv1_kwargs)
-    unet.mid_block.resnets[0].conv_shortcut = LoRACompatibleConv(**conv_shortcut_args_kwargs)
+    unet.mid_block.resnets[0].conv1 = (
+        nn.Conv2d(**conv1_kwargs) if USE_PEFT_BACKEND else LoRACompatibleConv(**conv1_kwargs)
+    )
+    unet.mid_block.resnets[0].conv_shortcut = (
+        nn.Conv2d(**conv_shortcut_args_kwargs) if USE_PEFT_BACKEND else LoRACompatibleConv(**conv_shortcut_args_kwargs)
+    )
     unet.mid_block.resnets[0].in_channels += by  # surgery done here
 
 

src/diffusers/models/downsampling.py

@@ -0,0 +1,318 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..utils import USE_PEFT_BACKEND
+from .lora import LoRACompatibleConv
+from .upsampling import upfirdn2d_native
+
+
+class Downsample1D(nn.Module):
+    """A 1D downsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+        padding (`int`, default `1`):
+            padding for the convolution.
+        name (`str`, default `conv`):
+            name of the downsampling 1D layer.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        use_conv: bool = False,
+        out_channels: Optional[int] = None,
+        padding: int = 1,
+        name: str = "conv",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.padding = padding
+        stride = 2
+        self.name = name
+
+        if use_conv:
+            self.conv = nn.Conv1d(self.channels, self.out_channels, 3, stride=stride, padding=padding)
+        else:
+            assert self.channels == self.out_channels
+            self.conv = nn.AvgPool1d(kernel_size=stride, stride=stride)
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        assert inputs.shape[1] == self.channels
+        return self.conv(inputs)
+
+
+class Downsample2D(nn.Module):
+    """A 2D downsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+        padding (`int`, default `1`):
+            padding for the convolution.
+        name (`str`, default `conv`):
+            name of the downsampling 2D layer.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        use_conv: bool = False,
+        out_channels: Optional[int] = None,
+        padding: int = 1,
+        name: str = "conv",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.padding = padding
+        stride = 2
+        self.name = name
+        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+
+        if use_conv:
+            conv = conv_cls(self.channels, self.out_channels, 3, stride=stride, padding=padding)
+        else:
+            assert self.channels == self.out_channels
+            conv = nn.AvgPool2d(kernel_size=stride, stride=stride)
+
+        # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
+        if name == "conv":
+            self.Conv2d_0 = conv
+            self.conv = conv
+        elif name == "Conv2d_0":
+            self.conv = conv
+        else:
+            self.conv = conv
+
+    def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor:
+        assert hidden_states.shape[1] == self.channels
+
+        if self.use_conv and self.padding == 0:
+            pad = (0, 1, 0, 1)
+            hidden_states = F.pad(hidden_states, pad, mode="constant", value=0)
+
+        assert hidden_states.shape[1] == self.channels
+
+        if not USE_PEFT_BACKEND:
+            if isinstance(self.conv, LoRACompatibleConv):
+                hidden_states = self.conv(hidden_states, scale)
+            else:
+                hidden_states = self.conv(hidden_states)
+        else:
+            hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class FirDownsample2D(nn.Module):
+    """A 2D FIR downsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+        fir_kernel (`tuple`, default `(1, 3, 3, 1)`):
+            kernel for the FIR filter.
+    """
+
+    def __init__(
+        self,
+        channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        use_conv: bool = False,
+        fir_kernel: Tuple[int, int, int, int] = (1, 3, 3, 1),
+    ):
+        super().__init__()
+        out_channels = out_channels if out_channels else channels
+        if use_conv:
+            self.Conv2d_0 = nn.Conv2d(channels, out_channels, kernel_size=3, stride=1, padding=1)
+        self.fir_kernel = fir_kernel
+        self.use_conv = use_conv
+        self.out_channels = out_channels
+
+    def _downsample_2d(
+        self,
+        hidden_states: torch.FloatTensor,
+        weight: Optional[torch.FloatTensor] = None,
+        kernel: Optional[torch.FloatTensor] = None,
+        factor: int = 2,
+        gain: float = 1,
+    ) -> torch.FloatTensor:
+        """Fused `Conv2d()` followed by `downsample_2d()`.
+        Padding is performed only once at the beginning, not between the operations. The fused op is considerably more
+        efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of
+        arbitrary order.
+
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+            weight (`torch.FloatTensor`, *optional*):
+                Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`. Grouped convolution can be
+                performed by `inChannels = x.shape[0] // numGroups`.
+            kernel (`torch.FloatTensor`, *optional*):
+                FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
+                corresponds to average pooling.
+            factor (`int`, *optional*, default to `2`):
+                Integer downsampling factor.
+            gain (`float`, *optional*, default to `1.0`):
+                Scaling factor for signal magnitude.
+
+        Returns:
+            output (`torch.FloatTensor`):
+                Tensor of the shape `[N, C, H // factor, W // factor]` or `[N, H // factor, W // factor, C]`, and same
+                datatype as `x`.
+        """
+
+        assert isinstance(factor, int) and factor >= 1
+        if kernel is None:
+            kernel = [1] * factor
+
+        # setup kernel
+        kernel = torch.tensor(kernel, dtype=torch.float32)
+        if kernel.ndim == 1:
+            kernel = torch.outer(kernel, kernel)
+        kernel /= torch.sum(kernel)
+
+        kernel = kernel * gain
+
+        if self.use_conv:
+            _, _, convH, convW = weight.shape
+            pad_value = (kernel.shape[0] - factor) + (convW - 1)
+            stride_value = [factor, factor]
+            upfirdn_input = upfirdn2d_native(
+                hidden_states,
+                torch.tensor(kernel, device=hidden_states.device),
+                pad=((pad_value + 1) // 2, pad_value // 2),
+            )
+            output = F.conv2d(upfirdn_input, weight, stride=stride_value, padding=0)
+        else:
+            pad_value = kernel.shape[0] - factor
+            output = upfirdn2d_native(
+                hidden_states,
+                torch.tensor(kernel, device=hidden_states.device),
+                down=factor,
+                pad=((pad_value + 1) // 2, pad_value // 2),
+            )
+
+        return output
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        if self.use_conv:
+            downsample_input = self._downsample_2d(hidden_states, weight=self.Conv2d_0.weight, kernel=self.fir_kernel)
+            hidden_states = downsample_input + self.Conv2d_0.bias.reshape(1, -1, 1, 1)
+        else:
+            hidden_states = self._downsample_2d(hidden_states, kernel=self.fir_kernel, factor=2)
+
+        return hidden_states
+
+
+# downsample/upsample layer used in k-upscaler, might be able to use FirDownsample2D/DirUpsample2D instead
+class KDownsample2D(nn.Module):
+    r"""A 2D K-downsampling layer.
+
+    Parameters:
+        pad_mode (`str`, *optional*, default to `"reflect"`): the padding mode to use.
+    """
+
+    def __init__(self, pad_mode: str = "reflect"):
+        super().__init__()
+        self.pad_mode = pad_mode
+        kernel_1d = torch.tensor([[1 / 8, 3 / 8, 3 / 8, 1 / 8]])
+        self.pad = kernel_1d.shape[1] // 2 - 1
+        self.register_buffer("kernel", kernel_1d.T @ kernel_1d, persistent=False)
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        inputs = F.pad(inputs, (self.pad,) * 4, self.pad_mode)
+        weight = inputs.new_zeros(
+            [
+                inputs.shape[1],
+                inputs.shape[1],
+                self.kernel.shape[0],
+                self.kernel.shape[1],
+            ]
+        )
+        indices = torch.arange(inputs.shape[1], device=inputs.device)
+        kernel = self.kernel.to(weight)[None, :].expand(inputs.shape[1], -1, -1)
+        weight[indices, indices] = kernel
+        return F.conv2d(inputs, weight, stride=2)
+
+
+def downsample_2d(
+    hidden_states: torch.FloatTensor,
+    kernel: Optional[torch.FloatTensor] = None,
+    factor: int = 2,
+    gain: float = 1,
+) -> torch.FloatTensor:
+    r"""Downsample2D a batch of 2D images with the given filter.
+    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and downsamples each image with the
+    given filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the
+    specified `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its
+    shape is a multiple of the downsampling factor.
+
+    Args:
+        hidden_states (`torch.FloatTensor`)
+            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        kernel (`torch.FloatTensor`, *optional*):
+            FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
+            corresponds to average pooling.
+        factor (`int`, *optional*, default to `2`):
+            Integer downsampling factor.
+        gain (`float`, *optional*, default to `1.0`):
+            Scaling factor for signal magnitude.
+
+    Returns:
+        output (`torch.FloatTensor`):
+            Tensor of the shape `[N, C, H // factor, W // factor]`
+    """
+
+    assert isinstance(factor, int) and factor >= 1
+    if kernel is None:
+        kernel = [1] * factor
+
+    kernel = torch.tensor(kernel, dtype=torch.float32)
+    if kernel.ndim == 1:
+        kernel = torch.outer(kernel, kernel)
+    kernel /= torch.sum(kernel)
+
+    kernel = kernel * gain
+    pad_value = kernel.shape[0] - factor
+    output = upfirdn2d_native(
+        hidden_states,
+        kernel.to(device=hidden_states.device),
+        down=factor,
+        pad=((pad_value + 1) // 2, pad_value // 2),
+    )
+    return output

src/diffusers/models/embeddings.py

@@ -729,7 +729,7 @@ class PositionNet(nn.Module):
         return objs
 
 
-class CombinedTimestepSizeEmbeddings(nn.Module):
+class PixArtAlphaCombinedTimestepSizeEmbeddings(nn.Module):
     """
     For PixArt-Alpha.
 
@@ -746,45 +746,27 @@ class CombinedTimestepSizeEmbeddings(nn.Module):
 
         self.use_additional_conditions = use_additional_conditions
         if use_additional_conditions:
-            self.use_additional_conditions = True
             self.additional_condition_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
             self.resolution_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=size_emb_dim)
             self.aspect_ratio_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=size_emb_dim)
 
-    def apply_condition(self, size: torch.Tensor, batch_size: int, embedder: nn.Module):
-        if size.ndim == 1:
-            size = size[:, None]
-
-        if size.shape[0] != batch_size:
-            size = size.repeat(batch_size // size.shape[0], 1)
-            if size.shape[0] != batch_size:
-                raise ValueError(f"`batch_size` should be {size.shape[0]} but found {batch_size}.")
-
-        current_batch_size, dims = size.shape[0], size.shape[1]
-        size = size.reshape(-1)
-        size_freq = self.additional_condition_proj(size).to(size.dtype)
-
-        size_emb = embedder(size_freq)
-        size_emb = size_emb.reshape(current_batch_size, dims * self.outdim)
-        return size_emb
-
     def forward(self, timestep, resolution, aspect_ratio, batch_size, hidden_dtype):
         timesteps_proj = self.time_proj(timestep)
         timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype))  # (N, D)
 
         if self.use_additional_conditions:
-            resolution = self.apply_condition(resolution, batch_size=batch_size, embedder=self.resolution_embedder)
-            aspect_ratio = self.apply_condition(
-                aspect_ratio, batch_size=batch_size, embedder=self.aspect_ratio_embedder
-            )
-            conditioning = timesteps_emb + torch.cat([resolution, aspect_ratio], dim=1)
+            resolution_emb = self.additional_condition_proj(resolution.flatten()).to(hidden_dtype)
+            resolution_emb = self.resolution_embedder(resolution_emb).reshape(batch_size, -1)
+            aspect_ratio_emb = self.additional_condition_proj(aspect_ratio.flatten()).to(hidden_dtype)
+            aspect_ratio_emb = self.aspect_ratio_embedder(aspect_ratio_emb).reshape(batch_size, -1)
+            conditioning = timesteps_emb + torch.cat([resolution_emb, aspect_ratio_emb], dim=1)
         else:
             conditioning = timesteps_emb
 
         return conditioning
 
 
-class CaptionProjection(nn.Module):
+class PixArtAlphaTextProjection(nn.Module):
     """
     Projects caption embeddings. Also handles dropout for classifier-free guidance.
 
@@ -796,9 +778,8 @@ class CaptionProjection(nn.Module):
         self.linear_1 = nn.Linear(in_features=in_features, out_features=hidden_size, bias=True)
         self.act_1 = nn.GELU(approximate="tanh")
         self.linear_2 = nn.Linear(in_features=hidden_size, out_features=hidden_size, bias=True)
-        self.register_buffer("y_embedding", nn.Parameter(torch.randn(num_tokens, in_features) / in_features**0.5))
 
-    def forward(self, caption, force_drop_ids=None):
+    def forward(self, caption):
         hidden_states = self.linear_1(caption)
         hidden_states = self.act_1(hidden_states)
         hidden_states = self.linear_2(hidden_states)

src/diffusers/models/normalization.py

@@ -20,7 +20,7 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 from .activations import get_activation
-from .embeddings import CombinedTimestepLabelEmbeddings, CombinedTimestepSizeEmbeddings
+from .embeddings import CombinedTimestepLabelEmbeddings, PixArtAlphaCombinedTimestepSizeEmbeddings
 
 
 class AdaLayerNorm(nn.Module):
@@ -91,7 +91,7 @@ class AdaLayerNormSingle(nn.Module):
     def __init__(self, embedding_dim: int, use_additional_conditions: bool = False):
         super().__init__()
 
-        self.emb = CombinedTimestepSizeEmbeddings(
+        self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings(
             embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions
         )
 

src/diffusers/models/resnet.py

@@ -23,562 +23,23 @@ import torch.nn.functional as F
 from ..utils import USE_PEFT_BACKEND
 from .activations import get_activation
 from .attention_processor import SpatialNorm
+from .downsampling import (  # noqa
+    Downsample1D,
+    Downsample2D,
+    FirDownsample2D,
+    KDownsample2D,
+    downsample_2d,
+)
 from .lora import LoRACompatibleConv, LoRACompatibleLinear
 from .normalization import AdaGroupNorm
-
-
-class Upsample1D(nn.Module):
-    """A 1D upsampling layer with an optional convolution.
-
-    Parameters:
-        channels (`int`):
-            number of channels in the inputs and outputs.
-        use_conv (`bool`, default `False`):
-            option to use a convolution.
-        use_conv_transpose (`bool`, default `False`):
-            option to use a convolution transpose.
-        out_channels (`int`, optional):
-            number of output channels. Defaults to `channels`.
-        name (`str`, default `conv`):
-            name of the upsampling 1D layer.
-    """
-
-    def __init__(
-        self,
-        channels: int,
-        use_conv: bool = False,
-        use_conv_transpose: bool = False,
-        out_channels: Optional[int] = None,
-        name: str = "conv",
-    ):
-        super().__init__()
-        self.channels = channels
-        self.out_channels = out_channels or channels
-        self.use_conv = use_conv
-        self.use_conv_transpose = use_conv_transpose
-        self.name = name
-
-        self.conv = None
-        if use_conv_transpose:
-            self.conv = nn.ConvTranspose1d(channels, self.out_channels, 4, 2, 1)
-        elif use_conv:
-            self.conv = nn.Conv1d(self.channels, self.out_channels, 3, padding=1)
-
-    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
-        assert inputs.shape[1] == self.channels
-        if self.use_conv_transpose:
-            return self.conv(inputs)
-
-        outputs = F.interpolate(inputs, scale_factor=2.0, mode="nearest")
-
-        if self.use_conv:
-            outputs = self.conv(outputs)
-
-        return outputs
-
-
-class Downsample1D(nn.Module):
-    """A 1D downsampling layer with an optional convolution.
-
-    Parameters:
-        channels (`int`):
-            number of channels in the inputs and outputs.
-        use_conv (`bool`, default `False`):
-            option to use a convolution.
-        out_channels (`int`, optional):
-            number of output channels. Defaults to `channels`.
-        padding (`int`, default `1`):
-            padding for the convolution.
-        name (`str`, default `conv`):
-            name of the downsampling 1D layer.
-    """
-
-    def __init__(
-        self,
-        channels: int,
-        use_conv: bool = False,
-        out_channels: Optional[int] = None,
-        padding: int = 1,
-        name: str = "conv",
-    ):
-        super().__init__()
-        self.channels = channels
-        self.out_channels = out_channels or channels
-        self.use_conv = use_conv
-        self.padding = padding
-        stride = 2
-        self.name = name
-
-        if use_conv:
-            self.conv = nn.Conv1d(self.channels, self.out_channels, 3, stride=stride, padding=padding)
-        else:
-            assert self.channels == self.out_channels
-            self.conv = nn.AvgPool1d(kernel_size=stride, stride=stride)
-
-    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
-        assert inputs.shape[1] == self.channels
-        return self.conv(inputs)
-
-
-class Upsample2D(nn.Module):
-    """A 2D upsampling layer with an optional convolution.
-
-    Parameters:
-        channels (`int`):
-            number of channels in the inputs and outputs.
-        use_conv (`bool`, default `False`):
-            option to use a convolution.
-        use_conv_transpose (`bool`, default `False`):
-            option to use a convolution transpose.
-        out_channels (`int`, optional):
-            number of output channels. Defaults to `channels`.
-        name (`str`, default `conv`):
-            name of the upsampling 2D layer.
-    """
-
-    def __init__(
-        self,
-        channels: int,
-        use_conv: bool = False,
-        use_conv_transpose: bool = False,
-        out_channels: Optional[int] = None,
-        name: str = "conv",
-    ):
-        super().__init__()
-        self.channels = channels
-        self.out_channels = out_channels or channels
-        self.use_conv = use_conv
-        self.use_conv_transpose = use_conv_transpose
-        self.name = name
-        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
-
-        conv = None
-        if use_conv_transpose:
-            conv = nn.ConvTranspose2d(channels, self.out_channels, 4, 2, 1)
-        elif use_conv:
-            conv = conv_cls(self.channels, self.out_channels, 3, padding=1)
-
-        # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
-        if name == "conv":
-            self.conv = conv
-        else:
-            self.Conv2d_0 = conv
-
-    def forward(
-        self,
-        hidden_states: torch.FloatTensor,
-        output_size: Optional[int] = None,
-        scale: float = 1.0,
-    ) -> torch.FloatTensor:
-        assert hidden_states.shape[1] == self.channels
-
-        if self.use_conv_transpose:
-            return self.conv(hidden_states)
-
-        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
-        # TODO(Suraj): Remove this cast once the issue is fixed in PyTorch
-        # https://github.com/pytorch/pytorch/issues/86679
-        dtype = hidden_states.dtype
-        if dtype == torch.bfloat16:
-            hidden_states = hidden_states.to(torch.float32)
-
-        # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
-        if hidden_states.shape[0] >= 64:
-            hidden_states = hidden_states.contiguous()
-
-        # if `output_size` is passed we force the interpolation output
-        # size and do not make use of `scale_factor=2`
-        if output_size is None:
-            hidden_states = F.interpolate(hidden_states, scale_factor=2.0, mode="nearest")
-        else:
-            hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest")
-
-        # If the input is bfloat16, we cast back to bfloat16
-        if dtype == torch.bfloat16:
-            hidden_states = hidden_states.to(dtype)
-
-        # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
-        if self.use_conv:
-            if self.name == "conv":
-                if isinstance(self.conv, LoRACompatibleConv) and not USE_PEFT_BACKEND:
-                    hidden_states = self.conv(hidden_states, scale)
-                else:
-                    hidden_states = self.conv(hidden_states)
-            else:
-                if isinstance(self.Conv2d_0, LoRACompatibleConv) and not USE_PEFT_BACKEND:
-                    hidden_states = self.Conv2d_0(hidden_states, scale)
-                else:
-                    hidden_states = self.Conv2d_0(hidden_states)
-
-        return hidden_states
-
-
-class Downsample2D(nn.Module):
-    """A 2D downsampling layer with an optional convolution.
-
-    Parameters:
-        channels (`int`):
-            number of channels in the inputs and outputs.
-        use_conv (`bool`, default `False`):
-            option to use a convolution.
-        out_channels (`int`, optional):
-            number of output channels. Defaults to `channels`.
-        padding (`int`, default `1`):
-            padding for the convolution.
-        name (`str`, default `conv`):
-            name of the downsampling 2D layer.
-    """
-
-    def __init__(
-        self,
-        channels: int,
-        use_conv: bool = False,
-        out_channels: Optional[int] = None,
-        padding: int = 1,
-        name: str = "conv",
-    ):
-        super().__init__()
-        self.channels = channels
-        self.out_channels = out_channels or channels
-        self.use_conv = use_conv
-        self.padding = padding
-        stride = 2
-        self.name = name
-        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
-
-        if use_conv:
-            conv = conv_cls(self.channels, self.out_channels, 3, stride=stride, padding=padding)
-        else:
-            assert self.channels == self.out_channels
-            conv = nn.AvgPool2d(kernel_size=stride, stride=stride)
-
-        # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
-        if name == "conv":
-            self.Conv2d_0 = conv
-            self.conv = conv
-        elif name == "Conv2d_0":
-            self.conv = conv
-        else:
-            self.conv = conv
-
-    def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor:
-        assert hidden_states.shape[1] == self.channels
-
-        if self.use_conv and self.padding == 0:
-            pad = (0, 1, 0, 1)
-            hidden_states = F.pad(hidden_states, pad, mode="constant", value=0)
-
-        assert hidden_states.shape[1] == self.channels
-
-        if not USE_PEFT_BACKEND:
-            if isinstance(self.conv, LoRACompatibleConv):
-                hidden_states = self.conv(hidden_states, scale)
-            else:
-                hidden_states = self.conv(hidden_states)
-        else:
-            hidden_states = self.conv(hidden_states)
-
-        return hidden_states
-
-
-class FirUpsample2D(nn.Module):
-    """A 2D FIR upsampling layer with an optional convolution.
-
-    Parameters:
-        channels (`int`, optional):
-            number of channels in the inputs and outputs.
-        use_conv (`bool`, default `False`):
-            option to use a convolution.
-        out_channels (`int`, optional):
-            number of output channels. Defaults to `channels`.
-        fir_kernel (`tuple`, default `(1, 3, 3, 1)`):
-            kernel for the FIR filter.
-    """
-
-    def __init__(
-        self,
-        channels: Optional[int] = None,
-        out_channels: Optional[int] = None,
-        use_conv: bool = False,
-        fir_kernel: Tuple[int, int, int, int] = (1, 3, 3, 1),
-    ):
-        super().__init__()
-        out_channels = out_channels if out_channels else channels
-        if use_conv:
-            self.Conv2d_0 = nn.Conv2d(channels, out_channels, kernel_size=3, stride=1, padding=1)
-        self.use_conv = use_conv
-        self.fir_kernel = fir_kernel
-        self.out_channels = out_channels
-
-    def _upsample_2d(
-        self,
-        hidden_states: torch.FloatTensor,
-        weight: Optional[torch.FloatTensor] = None,
-        kernel: Optional[torch.FloatTensor] = None,
-        factor: int = 2,
-        gain: float = 1,
-    ) -> torch.FloatTensor:
-        """Fused `upsample_2d()` followed by `Conv2d()`.
-
-        Padding is performed only once at the beginning, not between the operations. The fused op is considerably more
-        efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of
-        arbitrary order.
-
-        Args:
-            hidden_states (`torch.FloatTensor`):
-                Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
-            weight (`torch.FloatTensor`, *optional*):
-                Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`. Grouped convolution can be
-                performed by `inChannels = x.shape[0] // numGroups`.
-            kernel (`torch.FloatTensor`, *optional*):
-                FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
-                corresponds to nearest-neighbor upsampling.
-            factor (`int`, *optional*): Integer upsampling factor (default: 2).
-            gain (`float`, *optional*): Scaling factor for signal magnitude (default: 1.0).
-
-        Returns:
-            output (`torch.FloatTensor`):
-                Tensor of the shape `[N, C, H * factor, W * factor]` or `[N, H * factor, W * factor, C]`, and same
-                datatype as `hidden_states`.
-        """
-
-        assert isinstance(factor, int) and factor >= 1
-
-        # Setup filter kernel.
-        if kernel is None:
-            kernel = [1] * factor
-
-        # setup kernel
-        kernel = torch.tensor(kernel, dtype=torch.float32)
-        if kernel.ndim == 1:
-            kernel = torch.outer(kernel, kernel)
-        kernel /= torch.sum(kernel)
-
-        kernel = kernel * (gain * (factor**2))
-
-        if self.use_conv:
-            convH = weight.shape[2]
-            convW = weight.shape[3]
-            inC = weight.shape[1]
-
-            pad_value = (kernel.shape[0] - factor) - (convW - 1)
-
-            stride = (factor, factor)
-            # Determine data dimensions.
-            output_shape = (
-                (hidden_states.shape[2] - 1) * factor + convH,
-                (hidden_states.shape[3] - 1) * factor + convW,
-            )
-            output_padding = (
-                output_shape[0] - (hidden_states.shape[2] - 1) * stride[0] - convH,
-                output_shape[1] - (hidden_states.shape[3] - 1) * stride[1] - convW,
-            )
-            assert output_padding[0] >= 0 and output_padding[1] >= 0
-            num_groups = hidden_states.shape[1] // inC
-
-            # Transpose weights.
-            weight = torch.reshape(weight, (num_groups, -1, inC, convH, convW))
-            weight = torch.flip(weight, dims=[3, 4]).permute(0, 2, 1, 3, 4)
-            weight = torch.reshape(weight, (num_groups * inC, -1, convH, convW))
-
-            inverse_conv = F.conv_transpose2d(
-                hidden_states,
-                weight,
-                stride=stride,
-                output_padding=output_padding,
-                padding=0,
-            )
-
-            output = upfirdn2d_native(
-                inverse_conv,
-                torch.tensor(kernel, device=inverse_conv.device),
-                pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2 + 1),
-            )
-        else:
-            pad_value = kernel.shape[0] - factor
-            output = upfirdn2d_native(
-                hidden_states,
-                torch.tensor(kernel, device=hidden_states.device),
-                up=factor,
-                pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2),
-            )
-
-        return output
-
-    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
-        if self.use_conv:
-            height = self._upsample_2d(hidden_states, self.Conv2d_0.weight, kernel=self.fir_kernel)
-            height = height + self.Conv2d_0.bias.reshape(1, -1, 1, 1)
-        else:
-            height = self._upsample_2d(hidden_states, kernel=self.fir_kernel, factor=2)
-
-        return height
-
-
-class FirDownsample2D(nn.Module):
-    """A 2D FIR downsampling layer with an optional convolution.
-
-    Parameters:
-        channels (`int`):
-            number of channels in the inputs and outputs.
-        use_conv (`bool`, default `False`):
-            option to use a convolution.
-        out_channels (`int`, optional):
-            number of output channels. Defaults to `channels`.
-        fir_kernel (`tuple`, default `(1, 3, 3, 1)`):
-            kernel for the FIR filter.
-    """
-
-    def __init__(
-        self,
-        channels: Optional[int] = None,
-        out_channels: Optional[int] = None,
-        use_conv: bool = False,
-        fir_kernel: Tuple[int, int, int, int] = (1, 3, 3, 1),
-    ):
-        super().__init__()
-        out_channels = out_channels if out_channels else channels
-        if use_conv:
-            self.Conv2d_0 = nn.Conv2d(channels, out_channels, kernel_size=3, stride=1, padding=1)
-        self.fir_kernel = fir_kernel
-        self.use_conv = use_conv
-        self.out_channels = out_channels
-
-    def _downsample_2d(
-        self,
-        hidden_states: torch.FloatTensor,
-        weight: Optional[torch.FloatTensor] = None,
-        kernel: Optional[torch.FloatTensor] = None,
-        factor: int = 2,
-        gain: float = 1,
-    ) -> torch.FloatTensor:
-        """Fused `Conv2d()` followed by `downsample_2d()`.
-        Padding is performed only once at the beginning, not between the operations. The fused op is considerably more
-        efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of
-        arbitrary order.
-
-        Args:
-            hidden_states (`torch.FloatTensor`):
-                Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
-            weight (`torch.FloatTensor`, *optional*):
-                Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`. Grouped convolution can be
-                performed by `inChannels = x.shape[0] // numGroups`.
-            kernel (`torch.FloatTensor`, *optional*):
-                FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
-                corresponds to average pooling.
-            factor (`int`, *optional*, default to `2`):
-                Integer downsampling factor.
-            gain (`float`, *optional*, default to `1.0`):
-                Scaling factor for signal magnitude.
-
-        Returns:
-            output (`torch.FloatTensor`):
-                Tensor of the shape `[N, C, H // factor, W // factor]` or `[N, H // factor, W // factor, C]`, and same
-                datatype as `x`.
-        """
-
-        assert isinstance(factor, int) and factor >= 1
-        if kernel is None:
-            kernel = [1] * factor
-
-        # setup kernel
-        kernel = torch.tensor(kernel, dtype=torch.float32)
-        if kernel.ndim == 1:
-            kernel = torch.outer(kernel, kernel)
-        kernel /= torch.sum(kernel)
-
-        kernel = kernel * gain
-
-        if self.use_conv:
-            _, _, convH, convW = weight.shape
-            pad_value = (kernel.shape[0] - factor) + (convW - 1)
-            stride_value = [factor, factor]
-            upfirdn_input = upfirdn2d_native(
-                hidden_states,
-                torch.tensor(kernel, device=hidden_states.device),
-                pad=((pad_value + 1) // 2, pad_value // 2),
-            )
-            output = F.conv2d(upfirdn_input, weight, stride=stride_value, padding=0)
-        else:
-            pad_value = kernel.shape[0] - factor
-            output = upfirdn2d_native(
-                hidden_states,
-                torch.tensor(kernel, device=hidden_states.device),
-                down=factor,
-                pad=((pad_value + 1) // 2, pad_value // 2),
-            )
-
-        return output
-
-    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
-        if self.use_conv:
-            downsample_input = self._downsample_2d(hidden_states, weight=self.Conv2d_0.weight, kernel=self.fir_kernel)
-            hidden_states = downsample_input + self.Conv2d_0.bias.reshape(1, -1, 1, 1)
-        else:
-            hidden_states = self._downsample_2d(hidden_states, kernel=self.fir_kernel, factor=2)
-
-        return hidden_states
-
-
-# downsample/upsample layer used in k-upscaler, might be able to use FirDownsample2D/DirUpsample2D instead
-class KDownsample2D(nn.Module):
-    r"""A 2D K-downsampling layer.
-
-    Parameters:
-        pad_mode (`str`, *optional*, default to `"reflect"`): the padding mode to use.
-    """
-
-    def __init__(self, pad_mode: str = "reflect"):
-        super().__init__()
-        self.pad_mode = pad_mode
-        kernel_1d = torch.tensor([[1 / 8, 3 / 8, 3 / 8, 1 / 8]])
-        self.pad = kernel_1d.shape[1] // 2 - 1
-        self.register_buffer("kernel", kernel_1d.T @ kernel_1d, persistent=False)
-
-    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
-        inputs = F.pad(inputs, (self.pad,) * 4, self.pad_mode)
-        weight = inputs.new_zeros(
-            [
-                inputs.shape[1],
-                inputs.shape[1],
-                self.kernel.shape[0],
-                self.kernel.shape[1],
-            ]
-        )
-        indices = torch.arange(inputs.shape[1], device=inputs.device)
-        kernel = self.kernel.to(weight)[None, :].expand(inputs.shape[1], -1, -1)
-        weight[indices, indices] = kernel
-        return F.conv2d(inputs, weight, stride=2)
-
-
-class KUpsample2D(nn.Module):
-    r"""A 2D K-upsampling layer.
-
-    Parameters:
-        pad_mode (`str`, *optional*, default to `"reflect"`): the padding mode to use.
-    """
-
-    def __init__(self, pad_mode: str = "reflect"):
-        super().__init__()
-        self.pad_mode = pad_mode
-        kernel_1d = torch.tensor([[1 / 8, 3 / 8, 3 / 8, 1 / 8]]) * 2
-        self.pad = kernel_1d.shape[1] // 2 - 1
-        self.register_buffer("kernel", kernel_1d.T @ kernel_1d, persistent=False)
-
-    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
-        inputs = F.pad(inputs, ((self.pad + 1) // 2,) * 4, self.pad_mode)
-        weight = inputs.new_zeros(
-            [
-                inputs.shape[1],
-                inputs.shape[1],
-                self.kernel.shape[0],
-                self.kernel.shape[1],
-            ]
-        )
-        indices = torch.arange(inputs.shape[1], device=inputs.device)
-        kernel = self.kernel.to(weight)[None, :].expand(inputs.shape[1], -1, -1)
-        weight[indices, indices] = kernel
-        return F.conv_transpose2d(inputs, weight, stride=2, padding=self.pad * 2 + 1)
+from .upsampling import (  # noqa
+    FirUpsample2D,
+    KUpsample2D,
+    Upsample1D,
+    Upsample2D,
+    upfirdn2d_native,
+    upsample_2d,
+)
 
 
 class ResnetBlock2D(nn.Module):
@@ -894,151 +355,6 @@ class ResidualTemporalBlock1D(nn.Module):
         return out + self.residual_conv(inputs)
 
 
-def upsample_2d(
-    hidden_states: torch.FloatTensor,
-    kernel: Optional[torch.FloatTensor] = None,
-    factor: int = 2,
-    gain: float = 1,
-) -> torch.FloatTensor:
-    r"""Upsample2D a batch of 2D images with the given filter.
-    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and upsamples each image with the given
-    filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the specified
-    `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its shape is
-    a: multiple of the upsampling factor.
-
-    Args:
-        hidden_states (`torch.FloatTensor`):
-            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
-        kernel (`torch.FloatTensor`, *optional*):
-            FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
-            corresponds to nearest-neighbor upsampling.
-        factor (`int`, *optional*, default to `2`):
-            Integer upsampling factor.
-        gain (`float`, *optional*, default to `1.0`):
-            Scaling factor for signal magnitude (default: 1.0).
-
-    Returns:
-        output (`torch.FloatTensor`):
-            Tensor of the shape `[N, C, H * factor, W * factor]`
-    """
-    assert isinstance(factor, int) and factor >= 1
-    if kernel is None:
-        kernel = [1] * factor
-
-    kernel = torch.tensor(kernel, dtype=torch.float32)
-    if kernel.ndim == 1:
-        kernel = torch.outer(kernel, kernel)
-    kernel /= torch.sum(kernel)
-
-    kernel = kernel * (gain * (factor**2))
-    pad_value = kernel.shape[0] - factor
-    output = upfirdn2d_native(
-        hidden_states,
-        kernel.to(device=hidden_states.device),
-        up=factor,
-        pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2),
-    )
-    return output
-
-
-def downsample_2d(
-    hidden_states: torch.FloatTensor,
-    kernel: Optional[torch.FloatTensor] = None,
-    factor: int = 2,
-    gain: float = 1,
-) -> torch.FloatTensor:
-    r"""Downsample2D a batch of 2D images with the given filter.
-    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and downsamples each image with the
-    given filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the
-    specified `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its
-    shape is a multiple of the downsampling factor.
-
-    Args:
-        hidden_states (`torch.FloatTensor`)
-            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
-        kernel (`torch.FloatTensor`, *optional*):
-            FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
-            corresponds to average pooling.
-        factor (`int`, *optional*, default to `2`):
-            Integer downsampling factor.
-        gain (`float`, *optional*, default to `1.0`):
-            Scaling factor for signal magnitude.
-
-    Returns:
-        output (`torch.FloatTensor`):
-            Tensor of the shape `[N, C, H // factor, W // factor]`
-    """
-
-    assert isinstance(factor, int) and factor >= 1
-    if kernel is None:
-        kernel = [1] * factor
-
-    kernel = torch.tensor(kernel, dtype=torch.float32)
-    if kernel.ndim == 1:
-        kernel = torch.outer(kernel, kernel)
-    kernel /= torch.sum(kernel)
-
-    kernel = kernel * gain
-    pad_value = kernel.shape[0] - factor
-    output = upfirdn2d_native(
-        hidden_states,
-        kernel.to(device=hidden_states.device),
-        down=factor,
-        pad=((pad_value + 1) // 2, pad_value // 2),
-    )
-    return output
-
-
-def upfirdn2d_native(
-    tensor: torch.Tensor,
-    kernel: torch.Tensor,
-    up: int = 1,
-    down: int = 1,
-    pad: Tuple[int, int] = (0, 0),
-) -> torch.Tensor:
-    up_x = up_y = up
-    down_x = down_y = down
-    pad_x0 = pad_y0 = pad[0]
-    pad_x1 = pad_y1 = pad[1]
-
-    _, channel, in_h, in_w = tensor.shape
-    tensor = tensor.reshape(-1, in_h, in_w, 1)
-
-    _, in_h, in_w, minor = tensor.shape
-    kernel_h, kernel_w = kernel.shape
-
-    out = tensor.view(-1, in_h, 1, in_w, 1, minor)
-    out = F.pad(out, [0, 0, 0, up_x - 1, 0, 0, 0, up_y - 1])
-    out = out.view(-1, in_h * up_y, in_w * up_x, minor)
-
-    out = F.pad(out, [0, 0, max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)])
-    out = out.to(tensor.device)  # Move back to mps if necessary
-    out = out[
-        :,
-        max(-pad_y0, 0) : out.shape[1] - max(-pad_y1, 0),
-        max(-pad_x0, 0) : out.shape[2] - max(-pad_x1, 0),
-        :,
-    ]
-
-    out = out.permute(0, 3, 1, 2)
-    out = out.reshape([-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1])
-    w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w)
-    out = F.conv2d(out, w)
-    out = out.reshape(
-        -1,
-        minor,
-        in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1,
-        in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1,
-    )
-    out = out.permute(0, 2, 3, 1)
-    out = out[:, ::down_y, ::down_x, :]
-
-    out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h) // down_y + 1
-    out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w) // down_x + 1
-
-    return out.view(-1, channel, out_h, out_w)
-
-
 class TemporalConvLayer(nn.Module):
     """
     Temporal convolutional layer that can be used for video (sequence of images) input Code mostly copied from:

src/diffusers/models/transformer_2d.py

@@ -22,7 +22,7 @@ from ..configuration_utils import ConfigMixin, register_to_config
 from ..models.embeddings import ImagePositionalEmbeddings
 from ..utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version
 from .attention import BasicTransformerBlock
-from .embeddings import CaptionProjection, PatchEmbed
+from .embeddings import PatchEmbed, PixArtAlphaTextProjection
 from .lora import LoRACompatibleConv, LoRACompatibleLinear
 from .modeling_utils import ModelMixin
 from .normalization import AdaLayerNormSingle
@@ -235,7 +235,7 @@ class Transformer2DModel(ModelMixin, ConfigMixin):
 
         self.caption_projection = None
         if caption_channels is not None:
-            self.caption_projection = CaptionProjection(in_features=caption_channels, hidden_size=inner_dim)
+            self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
 
         self.gradient_checkpointing = False
 

src/diffusers/models/upsampling.py

@@ -0,0 +1,426 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..utils import USE_PEFT_BACKEND
+from .lora import LoRACompatibleConv
+
+
+class Upsample1D(nn.Module):
+    """A 1D upsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        use_conv_transpose (`bool`, default `False`):
+            option to use a convolution transpose.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+        name (`str`, default `conv`):
+            name of the upsampling 1D layer.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        use_conv: bool = False,
+        use_conv_transpose: bool = False,
+        out_channels: Optional[int] = None,
+        name: str = "conv",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+
+        self.conv = None
+        if use_conv_transpose:
+            self.conv = nn.ConvTranspose1d(channels, self.out_channels, 4, 2, 1)
+        elif use_conv:
+            self.conv = nn.Conv1d(self.channels, self.out_channels, 3, padding=1)
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        assert inputs.shape[1] == self.channels
+        if self.use_conv_transpose:
+            return self.conv(inputs)
+
+        outputs = F.interpolate(inputs, scale_factor=2.0, mode="nearest")
+
+        if self.use_conv:
+            outputs = self.conv(outputs)
+
+        return outputs
+
+
+class Upsample2D(nn.Module):
+    """A 2D upsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        use_conv_transpose (`bool`, default `False`):
+            option to use a convolution transpose.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+        name (`str`, default `conv`):
+            name of the upsampling 2D layer.
+    """
+
+    def __init__(
+        self,
+        channels: int,
+        use_conv: bool = False,
+        use_conv_transpose: bool = False,
+        out_channels: Optional[int] = None,
+        name: str = "conv",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+
+        conv = None
+        if use_conv_transpose:
+            conv = nn.ConvTranspose2d(channels, self.out_channels, 4, 2, 1)
+        elif use_conv:
+            conv = conv_cls(self.channels, self.out_channels, 3, padding=1)
+
+        # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
+        if name == "conv":
+            self.conv = conv
+        else:
+            self.Conv2d_0 = conv
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        output_size: Optional[int] = None,
+        scale: float = 1.0,
+    ) -> torch.FloatTensor:
+        assert hidden_states.shape[1] == self.channels
+
+        if self.use_conv_transpose:
+            return self.conv(hidden_states)
+
+        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
+        # TODO(Suraj): Remove this cast once the issue is fixed in PyTorch
+        # https://github.com/pytorch/pytorch/issues/86679
+        dtype = hidden_states.dtype
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(torch.float32)
+
+        # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+        if hidden_states.shape[0] >= 64:
+            hidden_states = hidden_states.contiguous()
+
+        # if `output_size` is passed we force the interpolation output
+        # size and do not make use of `scale_factor=2`
+        if output_size is None:
+            hidden_states = F.interpolate(hidden_states, scale_factor=2.0, mode="nearest")
+        else:
+            hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest")
+
+        # If the input is bfloat16, we cast back to bfloat16
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(dtype)
+
+        # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
+        if self.use_conv:
+            if self.name == "conv":
+                if isinstance(self.conv, LoRACompatibleConv) and not USE_PEFT_BACKEND:
+                    hidden_states = self.conv(hidden_states, scale)
+                else:
+                    hidden_states = self.conv(hidden_states)
+            else:
+                if isinstance(self.Conv2d_0, LoRACompatibleConv) and not USE_PEFT_BACKEND:
+                    hidden_states = self.Conv2d_0(hidden_states, scale)
+                else:
+                    hidden_states = self.Conv2d_0(hidden_states)
+
+        return hidden_states
+
+
+class FirUpsample2D(nn.Module):
+    """A 2D FIR upsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`, optional):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+        fir_kernel (`tuple`, default `(1, 3, 3, 1)`):
+            kernel for the FIR filter.
+    """
+
+    def __init__(
+        self,
+        channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        use_conv: bool = False,
+        fir_kernel: Tuple[int, int, int, int] = (1, 3, 3, 1),
+    ):
+        super().__init__()
+        out_channels = out_channels if out_channels else channels
+        if use_conv:
+            self.Conv2d_0 = nn.Conv2d(channels, out_channels, kernel_size=3, stride=1, padding=1)
+        self.use_conv = use_conv
+        self.fir_kernel = fir_kernel
+        self.out_channels = out_channels
+
+    def _upsample_2d(
+        self,
+        hidden_states: torch.FloatTensor,
+        weight: Optional[torch.FloatTensor] = None,
+        kernel: Optional[torch.FloatTensor] = None,
+        factor: int = 2,
+        gain: float = 1,
+    ) -> torch.FloatTensor:
+        """Fused `upsample_2d()` followed by `Conv2d()`.
+
+        Padding is performed only once at the beginning, not between the operations. The fused op is considerably more
+        efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of
+        arbitrary order.
+
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+            weight (`torch.FloatTensor`, *optional*):
+                Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`. Grouped convolution can be
+                performed by `inChannels = x.shape[0] // numGroups`.
+            kernel (`torch.FloatTensor`, *optional*):
+                FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
+                corresponds to nearest-neighbor upsampling.
+            factor (`int`, *optional*): Integer upsampling factor (default: 2).
+            gain (`float`, *optional*): Scaling factor for signal magnitude (default: 1.0).
+
+        Returns:
+            output (`torch.FloatTensor`):
+                Tensor of the shape `[N, C, H * factor, W * factor]` or `[N, H * factor, W * factor, C]`, and same
+                datatype as `hidden_states`.
+        """
+
+        assert isinstance(factor, int) and factor >= 1
+
+        # Setup filter kernel.
+        if kernel is None:
+            kernel = [1] * factor
+
+        # setup kernel
+        kernel = torch.tensor(kernel, dtype=torch.float32)
+        if kernel.ndim == 1:
+            kernel = torch.outer(kernel, kernel)
+        kernel /= torch.sum(kernel)
+
+        kernel = kernel * (gain * (factor**2))
+
+        if self.use_conv:
+            convH = weight.shape[2]
+            convW = weight.shape[3]
+            inC = weight.shape[1]
+
+            pad_value = (kernel.shape[0] - factor) - (convW - 1)
+
+            stride = (factor, factor)
+            # Determine data dimensions.
+            output_shape = (
+                (hidden_states.shape[2] - 1) * factor + convH,
+                (hidden_states.shape[3] - 1) * factor + convW,
+            )
+            output_padding = (
+                output_shape[0] - (hidden_states.shape[2] - 1) * stride[0] - convH,
+                output_shape[1] - (hidden_states.shape[3] - 1) * stride[1] - convW,
+            )
+            assert output_padding[0] >= 0 and output_padding[1] >= 0
+            num_groups = hidden_states.shape[1] // inC
+
+            # Transpose weights.
+            weight = torch.reshape(weight, (num_groups, -1, inC, convH, convW))
+            weight = torch.flip(weight, dims=[3, 4]).permute(0, 2, 1, 3, 4)
+            weight = torch.reshape(weight, (num_groups * inC, -1, convH, convW))
+
+            inverse_conv = F.conv_transpose2d(
+                hidden_states,
+                weight,
+                stride=stride,
+                output_padding=output_padding,
+                padding=0,
+            )
+
+            output = upfirdn2d_native(
+                inverse_conv,
+                torch.tensor(kernel, device=inverse_conv.device),
+                pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2 + 1),
+            )
+        else:
+            pad_value = kernel.shape[0] - factor
+            output = upfirdn2d_native(
+                hidden_states,
+                torch.tensor(kernel, device=hidden_states.device),
+                up=factor,
+                pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2),
+            )
+
+        return output
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        if self.use_conv:
+            height = self._upsample_2d(hidden_states, self.Conv2d_0.weight, kernel=self.fir_kernel)
+            height = height + self.Conv2d_0.bias.reshape(1, -1, 1, 1)
+        else:
+            height = self._upsample_2d(hidden_states, kernel=self.fir_kernel, factor=2)
+
+        return height
+
+
+class KUpsample2D(nn.Module):
+    r"""A 2D K-upsampling layer.
+
+    Parameters:
+        pad_mode (`str`, *optional*, default to `"reflect"`): the padding mode to use.
+    """
+
+    def __init__(self, pad_mode: str = "reflect"):
+        super().__init__()
+        self.pad_mode = pad_mode
+        kernel_1d = torch.tensor([[1 / 8, 3 / 8, 3 / 8, 1 / 8]]) * 2
+        self.pad = kernel_1d.shape[1] // 2 - 1
+        self.register_buffer("kernel", kernel_1d.T @ kernel_1d, persistent=False)
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        inputs = F.pad(inputs, ((self.pad + 1) // 2,) * 4, self.pad_mode)
+        weight = inputs.new_zeros(
+            [
+                inputs.shape[1],
+                inputs.shape[1],
+                self.kernel.shape[0],
+                self.kernel.shape[1],
+            ]
+        )
+        indices = torch.arange(inputs.shape[1], device=inputs.device)
+        kernel = self.kernel.to(weight)[None, :].expand(inputs.shape[1], -1, -1)
+        weight[indices, indices] = kernel
+        return F.conv_transpose2d(inputs, weight, stride=2, padding=self.pad * 2 + 1)
+
+
+def upfirdn2d_native(
+    tensor: torch.Tensor,
+    kernel: torch.Tensor,
+    up: int = 1,
+    down: int = 1,
+    pad: Tuple[int, int] = (0, 0),
+) -> torch.Tensor:
+    up_x = up_y = up
+    down_x = down_y = down
+    pad_x0 = pad_y0 = pad[0]
+    pad_x1 = pad_y1 = pad[1]
+
+    _, channel, in_h, in_w = tensor.shape
+    tensor = tensor.reshape(-1, in_h, in_w, 1)
+
+    _, in_h, in_w, minor = tensor.shape
+    kernel_h, kernel_w = kernel.shape
+
+    out = tensor.view(-1, in_h, 1, in_w, 1, minor)
+    out = F.pad(out, [0, 0, 0, up_x - 1, 0, 0, 0, up_y - 1])
+    out = out.view(-1, in_h * up_y, in_w * up_x, minor)
+
+    out = F.pad(out, [0, 0, max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)])
+    out = out.to(tensor.device)  # Move back to mps if necessary
+    out = out[
+        :,
+        max(-pad_y0, 0) : out.shape[1] - max(-pad_y1, 0),
+        max(-pad_x0, 0) : out.shape[2] - max(-pad_x1, 0),
+        :,
+    ]
+
+    out = out.permute(0, 3, 1, 2)
+    out = out.reshape([-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1])
+    w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w)
+    out = F.conv2d(out, w)
+    out = out.reshape(
+        -1,
+        minor,
+        in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1,
+        in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1,
+    )
+    out = out.permute(0, 2, 3, 1)
+    out = out[:, ::down_y, ::down_x, :]
+
+    out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h) // down_y + 1
+    out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w) // down_x + 1
+
+    return out.view(-1, channel, out_h, out_w)
+
+
+def upsample_2d(
+    hidden_states: torch.FloatTensor,
+    kernel: Optional[torch.FloatTensor] = None,
+    factor: int = 2,
+    gain: float = 1,
+) -> torch.FloatTensor:
+    r"""Upsample2D a batch of 2D images with the given filter.
+    Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and upsamples each image with the given
+    filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the specified
+    `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its shape is
+    a: multiple of the upsampling factor.
+
+    Args:
+        hidden_states (`torch.FloatTensor`):
+            Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+        kernel (`torch.FloatTensor`, *optional*):
+            FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
+            corresponds to nearest-neighbor upsampling.
+        factor (`int`, *optional*, default to `2`):
+            Integer upsampling factor.
+        gain (`float`, *optional*, default to `1.0`):
+            Scaling factor for signal magnitude (default: 1.0).
+
+    Returns:
+        output (`torch.FloatTensor`):
+            Tensor of the shape `[N, C, H * factor, W * factor]`
+    """
+    assert isinstance(factor, int) and factor >= 1
+    if kernel is None:
+        kernel = [1] * factor
+
+    kernel = torch.tensor(kernel, dtype=torch.float32)
+    if kernel.ndim == 1:
+        kernel = torch.outer(kernel, kernel)
+    kernel /= torch.sum(kernel)
+
+    kernel = kernel * (gain * (factor**2))
+    pad_value = kernel.shape[0] - factor
+    output = upfirdn2d_native(
+        hidden_states,
+        kernel.to(device=hidden_states.device),
+        up=factor,
+        pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2),
+    )
+    return output

src/diffusers/pipelines/__init__.py

@@ -179,12 +179,7 @@ else:
     _import_structure["stable_diffusion"].extend(
         [
             "CLIPImageProjection",
-            "StableDiffusionAttendAndExcitePipeline",
             "StableDiffusionDepth2ImgPipeline",
-            "StableDiffusionDiffEditPipeline",
-            "StableDiffusionGLIGENPipeline",
-            "StableDiffusionGLIGENPipeline",
-            "StableDiffusionGLIGENTextImagePipeline",
             "StableDiffusionImageVariationPipeline",
             "StableDiffusionImg2ImgPipeline",
             "StableDiffusionInpaintPipeline",
@@ -193,13 +188,18 @@ else:
             "StableDiffusionLDM3DPipeline",
             "StableDiffusionPanoramaPipeline",
             "StableDiffusionPipeline",
-            "StableDiffusionSAGPipeline",
             "StableDiffusionUpscalePipeline",
             "StableUnCLIPImg2ImgPipeline",
             "StableUnCLIPPipeline",
         ]
     )
+    _import_structure["stable_diffusion_attend_and_excite"] = ["StableDiffusionAttendAndExcitePipeline"]
     _import_structure["stable_diffusion_safe"] = ["StableDiffusionPipelineSafe"]
+    _import_structure["stable_diffusion_sag"] = ["StableDiffusionSAGPipeline"]
+    _import_structure["stable_diffusion_gligen"] = [
+        "StableDiffusionGLIGENPipeline",
+        "StableDiffusionGLIGENTextImagePipeline",
+    ]
     _import_structure["stable_video_diffusion"] = ["StableVideoDiffusionPipeline"]
     _import_structure["stable_diffusion_xl"].extend(
         [
@@ -209,6 +209,7 @@ else:
             "StableDiffusionXLPipeline",
         ]
     )
+    _import_structure["stable_diffusion_diffedit"] = ["StableDiffusionDiffEditPipeline"]
     _import_structure["t2i_adapter"] = [
         "StableDiffusionAdapterPipeline",
         "StableDiffusionXLAdapterPipeline",
@@ -268,7 +269,7 @@ except OptionalDependencyNotAvailable:
 
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_k_diffusion_objects))
 else:
-    _import_structure["stable_diffusion"].extend(["StableDiffusionKDiffusionPipeline"])
+    _import_structure["stable_diffusion_k_diffusion"] = ["StableDiffusionKDiffusionPipeline"]
 try:
     if not is_flax_available():
         raise OptionalDependencyNotAvailable()
@@ -420,11 +421,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline
         from .stable_diffusion import (
             CLIPImageProjection,
-            StableDiffusionAttendAndExcitePipeline,
             StableDiffusionDepth2ImgPipeline,
-            StableDiffusionDiffEditPipeline,
-            StableDiffusionGLIGENPipeline,
-            StableDiffusionGLIGENTextImagePipeline,
             StableDiffusionImageVariationPipeline,
             StableDiffusionImg2ImgPipeline,
             StableDiffusionInpaintPipeline,
@@ -433,12 +430,15 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusionLDM3DPipeline,
             StableDiffusionPanoramaPipeline,
             StableDiffusionPipeline,
-            StableDiffusionSAGPipeline,
             StableDiffusionUpscalePipeline,
             StableUnCLIPImg2ImgPipeline,
             StableUnCLIPPipeline,
         )
+        from .stable_diffusion_attend_and_excite import StableDiffusionAttendAndExcitePipeline
+        from .stable_diffusion_diffedit import StableDiffusionDiffEditPipeline
+        from .stable_diffusion_gligen import StableDiffusionGLIGENPipeline, StableDiffusionGLIGENTextImagePipeline
         from .stable_diffusion_safe import StableDiffusionPipelineSafe
+        from .stable_diffusion_sag import StableDiffusionSAGPipeline
         from .stable_diffusion_xl import (
             StableDiffusionXLImg2ImgPipeline,
             StableDiffusionXLInpaintPipeline,
@@ -498,7 +498,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         except OptionalDependencyNotAvailable:
             from ..utils.dummy_torch_and_transformers_and_k_diffusion_objects import *
         else:
-            from .stable_diffusion import StableDiffusionKDiffusionPipeline
+            from .stable_diffusion_k_diffusion import StableDiffusionKDiffusionPipeline
 
         try:
             if not is_flax_available():

src/diffusers/pipelines/pixart_alpha/pipeline_pixart_alpha.py

@@ -853,6 +853,11 @@ class PixArtAlphaPipeline(DiffusionPipeline):
             aspect_ratio = torch.tensor([float(height / width)]).repeat(batch_size * num_images_per_prompt, 1)
             resolution = resolution.to(dtype=prompt_embeds.dtype, device=device)
             aspect_ratio = aspect_ratio.to(dtype=prompt_embeds.dtype, device=device)
+
+            if do_classifier_free_guidance:
+                resolution = torch.cat([resolution, resolution], dim=0)
+                aspect_ratio = torch.cat([aspect_ratio, aspect_ratio], dim=0)
+
             added_cond_kwargs = {"resolution": resolution, "aspect_ratio": aspect_ratio}
 
         # 7. Denoising loop

src/diffusers/pipelines/stable_diffusion/__init__.py

@@ -44,7 +44,6 @@ else:
     _import_structure["pipeline_stable_diffusion_model_editing"] = ["StableDiffusionModelEditingPipeline"]
     _import_structure["pipeline_stable_diffusion_panorama"] = ["StableDiffusionPanoramaPipeline"]
     _import_structure["pipeline_stable_diffusion_paradigms"] = ["StableDiffusionParadigmsPipeline"]
-    _import_structure["pipeline_stable_diffusion_sag"] = ["StableDiffusionSAGPipeline"]
     _import_structure["pipeline_stable_diffusion_upscale"] = ["StableDiffusionUpscalePipeline"]
     _import_structure["pipeline_stable_unclip"] = ["StableUnCLIPPipeline"]
     _import_structure["pipeline_stable_unclip_img2img"] = ["StableUnCLIPImg2ImgPipeline"]
@@ -67,37 +66,19 @@ try:
 except OptionalDependencyNotAvailable:
     from ...utils.dummy_torch_and_transformers_objects import (
         StableDiffusionDepth2ImgPipeline,
-        StableDiffusionDiffEditPipeline,
         StableDiffusionPix2PixZeroPipeline,
     )
 
     _dummy_objects.update(
         {
             "StableDiffusionDepth2ImgPipeline": StableDiffusionDepth2ImgPipeline,
-            "StableDiffusionDiffEditPipeline": StableDiffusionDiffEditPipeline,
             "StableDiffusionPix2PixZeroPipeline": StableDiffusionPix2PixZeroPipeline,
         }
     )
 else:
     _import_structure["pipeline_stable_diffusion_depth2img"] = ["StableDiffusionDepth2ImgPipeline"]
-    _import_structure["pipeline_stable_diffusion_diffedit"] = ["StableDiffusionDiffEditPipeline"]
     _import_structure["pipeline_stable_diffusion_pix2pix_zero"] = ["StableDiffusionPix2PixZeroPipeline"]
-try:
-    if not (
-        is_torch_available()
-        and is_transformers_available()
-        and is_k_diffusion_available()
-        and is_k_diffusion_version(">=", "0.0.12")
-    ):
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    from ...utils import (
-        dummy_torch_and_transformers_and_k_diffusion_objects,
-    )
 
-    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_k_diffusion_objects))
-else:
-    _import_structure["pipeline_stable_diffusion_k_diffusion"] = ["StableDiffusionKDiffusionPipeline"]
 try:
     if not (is_transformers_available() and is_onnx_available()):
         raise OptionalDependencyNotAvailable()
@@ -139,13 +120,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             StableDiffusionPipelineOutput,
             StableDiffusionSafetyChecker,
         )
-        from .pipeline_stable_diffusion_attend_and_excite import (
-            StableDiffusionAttendAndExcitePipeline,
-        )
-        from .pipeline_stable_diffusion_gligen import StableDiffusionGLIGENPipeline
-        from .pipeline_stable_diffusion_gligen_text_image import (
-            StableDiffusionGLIGENTextImagePipeline,
-        )
         from .pipeline_stable_diffusion_img2img import StableDiffusionImg2ImgPipeline
         from .pipeline_stable_diffusion_inpaint import StableDiffusionInpaintPipeline
         from .pipeline_stable_diffusion_instruct_pix2pix import (
@@ -156,7 +130,6 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         )
         from .pipeline_stable_diffusion_ldm3d import StableDiffusionLDM3DPipeline
         from .pipeline_stable_diffusion_panorama import StableDiffusionPanoramaPipeline
-        from .pipeline_stable_diffusion_sag import StableDiffusionSAGPipeline
         from .pipeline_stable_diffusion_upscale import StableDiffusionUpscalePipeline
         from .pipeline_stable_unclip import StableUnCLIPPipeline
         from .pipeline_stable_unclip_img2img import StableUnCLIPImg2ImgPipeline
@@ -181,29 +154,12 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     except OptionalDependencyNotAvailable:
         from ...utils.dummy_torch_and_transformers_objects import (
             StableDiffusionDepth2ImgPipeline,
-            StableDiffusionDiffEditPipeline,
             StableDiffusionPix2PixZeroPipeline,
         )
     else:
         from .pipeline_stable_diffusion_depth2img import (
             StableDiffusionDepth2ImgPipeline,
         )
-        from .pipeline_stable_diffusion_diffedit import StableDiffusionDiffEditPipeline
-
-    try:
-        if not (
-            is_torch_available()
-            and is_transformers_available()
-            and is_k_diffusion_available()
-            and is_k_diffusion_version(">=", "0.0.12")
-        ):
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        from ...utils.dummy_torch_and_transformers_and_k_diffusion_objects import *
-    else:
-        from .pipeline_stable_diffusion_k_diffusion import (
-            StableDiffusionKDiffusionPipeline,
-        )
 
     try:
         if not (is_transformers_available() and is_onnx_available()):

src/diffusers/pipelines/stable_diffusion_attend_and_excite/__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_stable_diffusion_attend_and_excite"] = ["StableDiffusionAttendAndExcitePipeline"]
+
+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_stable_diffusion_attend_and_excite import StableDiffusionAttendAndExcitePipeline
+
+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/stable_diffusion_attend_and_excite/pipeline_stable_diffusion_attend_and_excite.py

@@ -37,8 +37,8 @@ from ...utils import (
 )
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline
-from . import StableDiffusionPipelineOutput
-from .safety_checker import StableDiffusionSafetyChecker
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
 
 logger = logging.get_logger(__name__)

src/diffusers/pipelines/stable_diffusion_diffedit/__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_stable_diffusion_diffedit"] = ["StableDiffusionDiffEditPipeline"]
+
+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_stable_diffusion_diffedit import StableDiffusionDiffEditPipeline
+
+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/stable_diffusion_diffedit/pipeline_stable_diffusion_diffedit.py

@@ -40,8 +40,8 @@ from ...utils import (
 )
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline
-from . import StableDiffusionPipelineOutput
-from .safety_checker import StableDiffusionSafetyChecker
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

src/diffusers/pipelines/stable_diffusion_gligen/__init__.py

@@ -0,0 +1,50 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_stable_diffusion_gligen"] = ["StableDiffusionGLIGENPipeline"]
+    _import_structure["pipeline_stable_diffusion_gligen_text_image"] = ["StableDiffusionGLIGENTextImagePipeline"]
+
+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_stable_diffusion_gligen import StableDiffusionGLIGENPipeline
+        from .pipeline_stable_diffusion_gligen_text_image import StableDiffusionGLIGENTextImagePipeline
+
+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/stable_diffusion_gligen/pipeline_stable_diffusion_gligen.py

@@ -36,8 +36,8 @@ from ...utils import (
 )
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline
-from . import StableDiffusionPipelineOutput
-from .safety_checker import StableDiffusionSafetyChecker
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

src/diffusers/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py

@@ -35,9 +35,9 @@ from ...schedulers import KarrasDiffusionSchedulers
 from ...utils import USE_PEFT_BACKEND, logging, replace_example_docstring, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline
-from . import StableDiffusionPipelineOutput
-from .clip_image_project_model import CLIPImageProjection
-from .safety_checker import StableDiffusionSafetyChecker
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.clip_image_project_model import CLIPImageProjection
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

src/diffusers/pipelines/stable_diffusion_k_diffusion/__init__.py

@@ -0,0 +1,60 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_k_diffusion_available,
+    is_k_diffusion_version,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+    if not (
+        is_transformers_available()
+        and is_torch_available()
+        and is_k_diffusion_available()
+        and is_k_diffusion_version(">=", "0.0.12")
+    ):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_and_k_diffusion_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_k_diffusion_objects))
+else:
+    _import_structure["pipeline_stable_diffusion_k_diffusion"] = ["StableDiffusionKDiffusionPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (
+            is_transformers_available()
+            and is_torch_available()
+            and is_k_diffusion_available()
+            and is_k_diffusion_version(">=", "0.0.12")
+        ):
+            raise OptionalDependencyNotAvailable()
+
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_and_k_diffusion_objects import *
+    else:
+        from .pipeline_stable_diffusion_k_diffusion import StableDiffusionKDiffusionPipeline
+
+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/stable_diffusion_k_diffusion/pipeline_stable_diffusion_k_diffusion.py

@@ -27,7 +27,7 @@ from ...schedulers import LMSDiscreteScheduler
 from ...utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline
-from . import StableDiffusionPipelineOutput
+from ..stable_diffusion import StableDiffusionPipelineOutput
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

src/diffusers/pipelines/stable_diffusion_sag/__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_stable_diffusion_sag"] = ["StableDiffusionSAGPipeline"]
+
+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_stable_diffusion_sag import StableDiffusionSAGPipeline
+
+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/stable_diffusion_sag/pipeline_stable_diffusion_sag.py

@@ -34,8 +34,8 @@ from ...utils import (
 )
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline
-from . import StableDiffusionPipelineOutput
-from .safety_checker import StableDiffusionSafetyChecker
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

src/diffusers/schedulers/scheduling_euler_ancestral_discrete.py

@@ -92,6 +92,43 @@ def betas_for_alpha_bar(
     return torch.tensor(betas, dtype=torch.float32)
 
 
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas):
+    """
+    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+    Args:
+        betas (`torch.FloatTensor`):
+            the betas that the scheduler is being initialized with.
+
+    Returns:
+        `torch.FloatTensor`: rescaled betas with zero terminal SNR
+    """
+    # Convert betas to alphas_bar_sqrt
+    alphas = 1.0 - betas
+    alphas_cumprod = torch.cumprod(alphas, dim=0)
+    alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+    # Store old values.
+    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+    # Shift so the last timestep is zero.
+    alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+    # Scale so the first timestep is back to the old value.
+    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+    # Convert alphas_bar_sqrt to betas
+    alphas_bar = alphas_bar_sqrt**2  # Revert sqrt
+    alphas = alphas_bar[1:] / alphas_bar[:-1]  # Revert cumprod
+    alphas = torch.cat([alphas_bar[0:1], alphas])
+    betas = 1 - alphas
+
+    return betas
+
+
 class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
     """
     Ancestral sampling with Euler method steps.
@@ -122,6 +159,10 @@ class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
             An offset added to the inference steps. You can use a combination of `offset=1` and
             `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
             Diffusion.
+        rescale_betas_zero_snr (`bool`, defaults to `False`):
+            Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+            dark samples instead of limiting it to samples with medium brightness. Loosely related to
+            [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
     """
 
     _compatibles = [e.name for e in KarrasDiffusionSchedulers]
@@ -138,6 +179,7 @@ class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
         prediction_type: str = "epsilon",
         timestep_spacing: str = "linspace",
         steps_offset: int = 0,
+        rescale_betas_zero_snr: bool = False,
     ):
         if trained_betas is not None:
             self.betas = torch.tensor(trained_betas, dtype=torch.float32)
@@ -152,9 +194,17 @@ class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
         else:
             raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
 
+        if rescale_betas_zero_snr:
+            self.betas = rescale_zero_terminal_snr(self.betas)
+
         self.alphas = 1.0 - self.betas
         self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
 
+        if rescale_betas_zero_snr:
+            # Close to 0 without being 0 so first sigma is not inf
+            # FP16 smallest positive subnormal works well here
+            self.alphas_cumprod[-1] = 2**-24
+
         sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
         sigmas = np.concatenate([sigmas[::-1], [0.0]]).astype(np.float32)
         self.sigmas = torch.from_numpy(sigmas)
@@ -327,6 +377,9 @@ class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
 
         sigma = self.sigmas[self.step_index]
 
+        # Upcast to avoid precision issues when computing prev_sample
+        sample = sample.to(torch.float32)
+
         # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
         if self.config.prediction_type == "epsilon":
             pred_original_sample = sample - sigma * model_output
@@ -357,6 +410,9 @@ class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
 
         prev_sample = prev_sample + noise * sigma_up
 
+        # Cast sample back to model compatible dtype
+        prev_sample = prev_sample.to(model_output.dtype)
+
         # upon completion increase step index by one
         self._step_index += 1
 

tests/lora/test_lora_layers_peft.py

@@ -13,6 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import copy
+import importlib
 import os
 import tempfile
 import time
@@ -24,6 +25,7 @@ import torch.nn as nn
 import torch.nn.functional as F
 from huggingface_hub import hf_hub_download
 from huggingface_hub.repocard import RepoCard
+from packaging import version
 from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
 
 from diffusers import (
@@ -1983,10 +1985,26 @@ class LoraSDXLIntegrationTests(unittest.TestCase):
         fused_te_2_state_dict = pipe.text_encoder_2.state_dict()
         unet_state_dict = pipe.unet.state_dict()
 
+        peft_ge_070 = version.parse(importlib.metadata.version("peft")) >= version.parse("0.7.0")
+
+        def remap_key(key, sd):
+            # some keys have moved around for PEFT >= 0.7.0, but they should still be loaded correctly
+            if (key in sd) or (not peft_ge_070):
+                return key
+
+            # instead of linear.weight, we now have linear.base_layer.weight, etc.
+            if key.endswith(".weight"):
+                key = key[:-7] + ".base_layer.weight"
+            elif key.endswith(".bias"):
+                key = key[:-5] + ".base_layer.bias"
+            return key
+
         for key, value in text_encoder_1_sd.items():
+            key = remap_key(key, fused_te_state_dict)
             self.assertTrue(torch.allclose(fused_te_state_dict[key], value))
 
         for key, value in text_encoder_2_sd.items():
+            key = remap_key(key, fused_te_2_state_dict)
             self.assertTrue(torch.allclose(fused_te_2_state_dict[key], value))
 
         for key, value in unet_state_dict.items():

tests/pipelines/controlnetxs/test_controlnetxs.py

@@ -34,6 +34,7 @@ from diffusers.utils.testing_utils import (
     enable_full_determinism,
     load_image,
     load_numpy,
+    numpy_cosine_similarity_distance,
     require_python39_or_higher,
     require_torch_2,
     require_torch_gpu,
@@ -273,7 +274,9 @@ class ControlNetXSPipelineSlowTests(unittest.TestCase):
 
         original_image = image[-3:, -3:, -1].flatten()
         expected_image = np.array([0.1274, 0.1401, 0.147, 0.1185, 0.1555, 0.1492, 0.1565, 0.1474, 0.1701])
-        assert np.allclose(original_image, expected_image, atol=1e-04)
+
+        max_diff = numpy_cosine_similarity_distance(original_image, expected_image)
+        assert max_diff < 1e-4
 
     def test_depth(self):
         controlnet = ControlNetXSModel.from_pretrained("UmerHA/ConrolNetXS-SD2.1-depth")
@@ -298,7 +301,9 @@ class ControlNetXSPipelineSlowTests(unittest.TestCase):
 
         original_image = image[-3:, -3:, -1].flatten()
         expected_image = np.array([0.1098, 0.1025, 0.1211, 0.1129, 0.1165, 0.1262, 0.1185, 0.1261, 0.1703])
-        assert np.allclose(original_image, expected_image, atol=1e-04)
+
+        max_diff = numpy_cosine_similarity_distance(original_image, expected_image)
+        assert max_diff < 1e-4
 
     @require_python39_or_higher
     @require_torch_2

tests/pipelines/pixart/test_pixart.py

@@ -64,7 +64,9 @@ class PixArtAlphaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
             norm_elementwise_affine=False,
             norm_eps=1e-6,
         )
+        torch.manual_seed(0)
         vae = AutoencoderKL()
+
         scheduler = DDIMScheduler()
         text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
 
@@ -186,7 +188,7 @@ class PixArtAlphaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         image_slice = image[0, -3:, -3:, -1]
 
         self.assertEqual(image.shape, (1, 8, 8, 3))
-        expected_slice = np.array([0.5303, 0.2658, 0.7979, 0.1182, 0.3304, 0.4608, 0.5195, 0.4261, 0.4675])
+        expected_slice = np.array([0.6319, 0.3526, 0.3806, 0.6327, 0.4639, 0.483, 0.2583, 0.5331, 0.4852])
         max_diff = np.abs(image_slice.flatten() - expected_slice).max()
         self.assertLessEqual(max_diff, 1e-3)
 
@@ -203,7 +205,7 @@ class PixArtAlphaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         image_slice = image[0, -3:, -3:, -1]
         self.assertEqual(image.shape, (1, 32, 48, 3))
 
-        expected_slice = np.array([0.3859, 0.2987, 0.2333, 0.5243, 0.6721, 0.4436, 0.5292, 0.5373, 0.4416])
+        expected_slice = np.array([0.6493, 0.537, 0.4081, 0.4762, 0.3695, 0.4711, 0.3026, 0.5218, 0.5263])
         max_diff = np.abs(image_slice.flatten() - expected_slice).max()
         self.assertLessEqual(max_diff, 1e-3)
 
@@ -293,7 +295,7 @@ class PixArtAlphaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         image_slice = image[0, -3:, -3:, -1]
 
         self.assertEqual(image.shape, (2, 8, 8, 3))
-        expected_slice = np.array([0.5303, 0.2658, 0.7979, 0.1182, 0.3304, 0.4608, 0.5195, 0.4261, 0.4675])
+        expected_slice = np.array([0.6319, 0.3526, 0.3806, 0.6327, 0.4639, 0.483, 0.2583, 0.5331, 0.4852])
         max_diff = np.abs(image_slice.flatten() - expected_slice).max()
         self.assertLessEqual(max_diff, 1e-3)
 

tests/pipelines/stable_diffusion/test_onnx_stable_diffusion_inpaint_legacy.py

@@ -1,97 +0,0 @@
-# coding=utf-8
-# Copyright 2023 HuggingFace Inc.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import unittest
-
-import numpy as np
-
-from diffusers import OnnxStableDiffusionInpaintPipelineLegacy
-from diffusers.utils.testing_utils import (
-    is_onnx_available,
-    load_image,
-    load_numpy,
-    nightly,
-    require_onnxruntime,
-    require_torch_gpu,
-)
-
-
-if is_onnx_available():
-    import onnxruntime as ort
-
-
-@nightly
-@require_onnxruntime
-@require_torch_gpu
-class StableDiffusionOnnxInpaintLegacyPipelineIntegrationTests(unittest.TestCase):
-    @property
-    def gpu_provider(self):
-        return (
-            "CUDAExecutionProvider",
-            {
-                "gpu_mem_limit": "15000000000",  # 15GB
-                "arena_extend_strategy": "kSameAsRequested",
-            },
-        )
-
-    @property
-    def gpu_options(self):
-        options = ort.SessionOptions()
-        options.enable_mem_pattern = False
-        return options
-
-    def test_inference(self):
-        init_image = load_image(
-            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
-            "/in_paint/overture-creations-5sI6fQgYIuo.png"
-        )
-        mask_image = load_image(
-            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
-            "/in_paint/overture-creations-5sI6fQgYIuo_mask.png"
-        )
-        expected_image = load_numpy(
-            "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
-            "/in_paint/red_cat_sitting_on_a_park_bench_onnx.npy"
-        )
-
-        # using the PNDM scheduler by default
-        pipe = OnnxStableDiffusionInpaintPipelineLegacy.from_pretrained(
-            "CompVis/stable-diffusion-v1-4",
-            revision="onnx",
-            safety_checker=None,
-            feature_extractor=None,
-            provider=self.gpu_provider,
-            sess_options=self.gpu_options,
-        )
-        pipe.set_progress_bar_config(disable=None)
-
-        prompt = "A red cat sitting on a park bench"
-
-        generator = np.random.RandomState(0)
-        output = pipe(
-            prompt=prompt,
-            image=init_image,
-            mask_image=mask_image,
-            strength=0.75,
-            guidance_scale=7.5,
-            num_inference_steps=15,
-            generator=generator,
-            output_type="np",
-        )
-
-        image = output.images[0]
-
-        assert image.shape == (512, 512, 3)
-        assert np.abs(expected_image - image).max() < 1e-2

tests/schedulers/test_scheduler_euler_ancestral.py

@@ -37,6 +37,10 @@ class EulerAncestralDiscreteSchedulerTest(SchedulerCommonTest):
         for prediction_type in ["epsilon", "v_prediction"]:
             self.check_over_configs(prediction_type=prediction_type)
 
+    def test_rescale_betas_zero_snr(self):
+        for rescale_betas_zero_snr in [True, False]:
+            self.check_over_configs(rescale_betas_zero_snr=rescale_betas_zero_snr)
+
     def test_full_loop_no_noise(self):
         scheduler_class = self.scheduler_classes[0]
         scheduler_config = self.get_scheduler_config()