PixArt-Sigma Implementation (#7654)

* support PixArt-DMD

---------

Co-authored-by: jschen <chenjunsong4@h-partners.com>
Co-authored-by: badayvedat <badayvedat@gmail.com>
Co-authored-by: Vedat Baday <54285744+badayvedat@users.noreply.github.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: YiYi Xu <yixu310@gmail.com>
Co-authored-by: yiyixuxu <yixu310@gmail,com>

scripts/convert_pixart_sigma_to_diffusers.py

@@ -0,0 +1,223 @@
+import argparse
+import os
+
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from diffusers import AutoencoderKL, DPMSolverMultistepScheduler, PixArtSigmaPipeline, Transformer2DModel
+
+
+ckpt_id = "PixArt-alpha"
+# https://github.com/PixArt-alpha/PixArt-sigma/blob/dd087141864e30ec44f12cb7448dd654be065e88/scripts/inference.py#L158
+interpolation_scale = {256: 0.5, 512: 1, 1024: 2, 2048: 4}
+
+
+def main(args):
+    all_state_dict = torch.load(args.orig_ckpt_path)
+    state_dict = all_state_dict.pop("state_dict")
+    converted_state_dict = {}
+
+    # Patch embeddings.
+    converted_state_dict["pos_embed.proj.weight"] = state_dict.pop("x_embedder.proj.weight")
+    converted_state_dict["pos_embed.proj.bias"] = state_dict.pop("x_embedder.proj.bias")
+
+    # Caption projection.
+    converted_state_dict["caption_projection.linear_1.weight"] = state_dict.pop("y_embedder.y_proj.fc1.weight")
+    converted_state_dict["caption_projection.linear_1.bias"] = state_dict.pop("y_embedder.y_proj.fc1.bias")
+    converted_state_dict["caption_projection.linear_2.weight"] = state_dict.pop("y_embedder.y_proj.fc2.weight")
+    converted_state_dict["caption_projection.linear_2.bias"] = state_dict.pop("y_embedder.y_proj.fc2.bias")
+
+    # AdaLN-single LN
+    converted_state_dict["adaln_single.emb.timestep_embedder.linear_1.weight"] = state_dict.pop(
+        "t_embedder.mlp.0.weight"
+    )
+    converted_state_dict["adaln_single.emb.timestep_embedder.linear_1.bias"] = state_dict.pop("t_embedder.mlp.0.bias")
+    converted_state_dict["adaln_single.emb.timestep_embedder.linear_2.weight"] = state_dict.pop(
+        "t_embedder.mlp.2.weight"
+    )
+    converted_state_dict["adaln_single.emb.timestep_embedder.linear_2.bias"] = state_dict.pop("t_embedder.mlp.2.bias")
+
+    if args.micro_condition:
+        # Resolution.
+        converted_state_dict["adaln_single.emb.resolution_embedder.linear_1.weight"] = state_dict.pop(
+            "csize_embedder.mlp.0.weight"
+        )
+        converted_state_dict["adaln_single.emb.resolution_embedder.linear_1.bias"] = state_dict.pop(
+            "csize_embedder.mlp.0.bias"
+        )
+        converted_state_dict["adaln_single.emb.resolution_embedder.linear_2.weight"] = state_dict.pop(
+            "csize_embedder.mlp.2.weight"
+        )
+        converted_state_dict["adaln_single.emb.resolution_embedder.linear_2.bias"] = state_dict.pop(
+            "csize_embedder.mlp.2.bias"
+        )
+        # Aspect ratio.
+        converted_state_dict["adaln_single.emb.aspect_ratio_embedder.linear_1.weight"] = state_dict.pop(
+            "ar_embedder.mlp.0.weight"
+        )
+        converted_state_dict["adaln_single.emb.aspect_ratio_embedder.linear_1.bias"] = state_dict.pop(
+            "ar_embedder.mlp.0.bias"
+        )
+        converted_state_dict["adaln_single.emb.aspect_ratio_embedder.linear_2.weight"] = state_dict.pop(
+            "ar_embedder.mlp.2.weight"
+        )
+        converted_state_dict["adaln_single.emb.aspect_ratio_embedder.linear_2.bias"] = state_dict.pop(
+            "ar_embedder.mlp.2.bias"
+        )
+    # Shared norm.
+    converted_state_dict["adaln_single.linear.weight"] = state_dict.pop("t_block.1.weight")
+    converted_state_dict["adaln_single.linear.bias"] = state_dict.pop("t_block.1.bias")
+
+    for depth in range(28):
+        # Transformer blocks.
+        converted_state_dict[f"transformer_blocks.{depth}.scale_shift_table"] = state_dict.pop(
+            f"blocks.{depth}.scale_shift_table"
+        )
+        # Attention is all you need 🤘
+
+        # Self attention.
+        q, k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.attn.qkv.weight"), 3, dim=0)
+        q_bias, k_bias, v_bias = torch.chunk(state_dict.pop(f"blocks.{depth}.attn.qkv.bias"), 3, dim=0)
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.bias"] = q_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.bias"] = k_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.weight"] = v
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.bias"] = v_bias
+        # Projection.
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.weight"] = state_dict.pop(
+            f"blocks.{depth}.attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.bias"] = state_dict.pop(
+            f"blocks.{depth}.attn.proj.bias"
+        )
+        if args.qk_norm:
+            converted_state_dict[f"transformer_blocks.{depth}.attn1.q_norm.weight"] = state_dict.pop(
+                f"blocks.{depth}.attn.q_norm.weight"
+            )
+            converted_state_dict[f"transformer_blocks.{depth}.attn1.q_norm.bias"] = state_dict.pop(
+                f"blocks.{depth}.attn.q_norm.bias"
+            )
+            converted_state_dict[f"transformer_blocks.{depth}.attn1.k_norm.weight"] = state_dict.pop(
+                f"blocks.{depth}.attn.k_norm.weight"
+            )
+            converted_state_dict[f"transformer_blocks.{depth}.attn1.k_norm.bias"] = state_dict.pop(
+                f"blocks.{depth}.attn.k_norm.bias"
+            )
+
+        # Feed-forward.
+        converted_state_dict[f"transformer_blocks.{depth}.ff.net.0.proj.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.fc1.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.net.0.proj.bias"] = state_dict.pop(
+            f"blocks.{depth}.mlp.fc1.bias"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.net.2.weight"] = state_dict.pop(
+            f"blocks.{depth}.mlp.fc2.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.ff.net.2.bias"] = state_dict.pop(
+            f"blocks.{depth}.mlp.fc2.bias"
+        )
+
+        # Cross-attention.
+        q = state_dict.pop(f"blocks.{depth}.cross_attn.q_linear.weight")
+        q_bias = state_dict.pop(f"blocks.{depth}.cross_attn.q_linear.bias")
+        k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.cross_attn.kv_linear.weight"), 2, dim=0)
+        k_bias, v_bias = torch.chunk(state_dict.pop(f"blocks.{depth}.cross_attn.kv_linear.bias"), 2, dim=0)
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.weight"] = q
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.bias"] = q_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.weight"] = k
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.bias"] = k_bias
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.weight"] = v
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.bias"] = v_bias
+
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.weight"] = state_dict.pop(
+            f"blocks.{depth}.cross_attn.proj.weight"
+        )
+        converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.bias"] = state_dict.pop(
+            f"blocks.{depth}.cross_attn.proj.bias"
+        )
+
+    # Final block.
+    converted_state_dict["proj_out.weight"] = state_dict.pop("final_layer.linear.weight")
+    converted_state_dict["proj_out.bias"] = state_dict.pop("final_layer.linear.bias")
+    converted_state_dict["scale_shift_table"] = state_dict.pop("final_layer.scale_shift_table")
+
+    # PixArt XL/2
+    transformer = Transformer2DModel(
+        sample_size=args.image_size // 8,
+        num_layers=28,
+        attention_head_dim=72,
+        in_channels=4,
+        out_channels=8,
+        patch_size=2,
+        attention_bias=True,
+        num_attention_heads=16,
+        cross_attention_dim=1152,
+        activation_fn="gelu-approximate",
+        num_embeds_ada_norm=1000,
+        norm_type="ada_norm_single",
+        norm_elementwise_affine=False,
+        norm_eps=1e-6,
+        caption_channels=4096,
+        interpolation_scale=interpolation_scale[args.image_size],
+        use_additional_conditions=args.micro_condition,
+    )
+    transformer.load_state_dict(converted_state_dict, strict=True)
+
+    assert transformer.pos_embed.pos_embed is not None
+    try:
+        state_dict.pop("y_embedder.y_embedding")
+        state_dict.pop("pos_embed")
+    except Exception as e:
+        print(f"Skipping {str(e)}")
+        pass
+    assert len(state_dict) == 0, f"State dict is not empty, {state_dict.keys()}"
+
+    num_model_params = sum(p.numel() for p in transformer.parameters())
+    print(f"Total number of transformer parameters: {num_model_params}")
+
+    if args.only_transformer:
+        transformer.save_pretrained(os.path.join(args.dump_path, "transformer"))
+    else:
+        # pixart-Sigma vae link: https://huggingface.co/PixArt-alpha/pixart_sigma_sdxlvae_T5_diffusers/tree/main/vae
+        vae = AutoencoderKL.from_pretrained(f"{ckpt_id}/pixart_sigma_sdxlvae_T5_diffusers", subfolder="vae")
+
+        scheduler = DPMSolverMultistepScheduler()
+
+        tokenizer = T5Tokenizer.from_pretrained(f"{ckpt_id}/pixart_sigma_sdxlvae_T5_diffusers", subfolder="tokenizer")
+        text_encoder = T5EncoderModel.from_pretrained(
+            f"{ckpt_id}/pixart_sigma_sdxlvae_T5_diffusers", subfolder="text_encoder"
+        )
+
+        pipeline = PixArtSigmaPipeline(
+            tokenizer=tokenizer, text_encoder=text_encoder, transformer=transformer, vae=vae, scheduler=scheduler
+        )
+
+        pipeline.save_pretrained(args.dump_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--micro_condition", action="store_true", help="If use Micro-condition in PixArtMS structure during training."
+    )
+    parser.add_argument("--qk_norm", action="store_true", help="If use qk norm during training.")
+    parser.add_argument(
+        "--orig_ckpt_path", default=None, type=str, required=False, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--image_size",
+        default=1024,
+        type=int,
+        choices=[256, 512, 1024, 2048],
+        required=False,
+        help="Image size of pretrained model, 256, 512, 1024, or 2048.",
+    )
+    parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
+    parser.add_argument("--only_transformer", default=True, type=bool, required=True)
+
+    args = parser.parse_args()
+    main(args)

src/diffusers/__init__.py

@@ -261,6 +261,7 @@ else:
             "PaintByExamplePipeline",
             "PIAPipeline",
             "PixArtAlphaPipeline",
+            "PixArtSigmaPipeline",
             "SemanticStableDiffusionPipeline",
             "ShapEImg2ImgPipeline",
             "ShapEPipeline",
@@ -637,6 +638,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             PaintByExamplePipeline,
             PIAPipeline,
             PixArtAlphaPipeline,
+            PixArtSigmaPipeline,
             SemanticStableDiffusionPipeline,
             ShapEImg2ImgPipeline,
             ShapEPipeline,

src/diffusers/image_processor.py

@@ -994,3 +994,77 @@ class IPAdapterMaskProcessor(VaeImageProcessor):
         )
 
         return mask_downsample
+
+
+class PixArtImageProcessor(VaeImageProcessor):
+    """
+    Image processor for PixArt image resize and crop.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`. Can accept
+            `height` and `width` arguments from [`image_processor.VaeImageProcessor.preprocess`] method.
+        vae_scale_factor (`int`, *optional*, defaults to `8`):
+            VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor.
+        resample (`str`, *optional*, defaults to `lanczos`):
+            Resampling filter to use when resizing the image.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image to [-1,1].
+        do_binarize (`bool`, *optional*, defaults to `False`):
+            Whether to binarize the image to 0/1.
+        do_convert_rgb (`bool`, *optional*, defaults to be `False`):
+            Whether to convert the images to RGB format.
+        do_convert_grayscale (`bool`, *optional*, defaults to be `False`):
+            Whether to convert the images to grayscale format.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        do_resize: bool = True,
+        vae_scale_factor: int = 8,
+        resample: str = "lanczos",
+        do_normalize: bool = True,
+        do_binarize: bool = False,
+        do_convert_grayscale: bool = False,
+    ):
+        super().__init__(
+            do_resize=do_resize,
+            vae_scale_factor=vae_scale_factor,
+            resample=resample,
+            do_normalize=do_normalize,
+            do_binarize=do_binarize,
+            do_convert_grayscale=do_convert_grayscale,
+        )
+
+    @staticmethod
+    def classify_height_width_bin(height: int, width: int, ratios: dict) -> Tuple[int, int]:
+        """Returns binned height and width."""
+        ar = float(height / width)
+        closest_ratio = min(ratios.keys(), key=lambda ratio: abs(float(ratio) - ar))
+        default_hw = ratios[closest_ratio]
+        return int(default_hw[0]), int(default_hw[1])
+
+    @staticmethod
+    def resize_and_crop_tensor(samples: torch.Tensor, new_width: int, new_height: int) -> torch.Tensor:
+        orig_height, orig_width = samples.shape[2], samples.shape[3]
+
+        # Check if resizing is needed
+        if orig_height != new_height or orig_width != new_width:
+            ratio = max(new_height / orig_height, new_width / orig_width)
+            resized_width = int(orig_width * ratio)
+            resized_height = int(orig_height * ratio)
+
+            # Resize
+            samples = F.interpolate(
+                samples, size=(resized_height, resized_width), mode="bilinear", align_corners=False
+            )
+
+            # Center Crop
+            start_x = (resized_width - new_width) // 2
+            end_x = start_x + new_width
+            start_y = (resized_height - new_height) // 2
+            end_y = start_y + new_height
+            samples = samples[:, :, start_y:end_y, start_x:end_x]
+
+        return samples

src/diffusers/models/transformers/transformer_2d.py

@@ -100,6 +100,7 @@ class Transformer2DModel(ModelMixin, ConfigMixin):
         attention_type: str = "default",
         caption_channels: int = None,
         interpolation_scale: float = None,
+        use_additional_conditions: Optional[bool] = None,
     ):
         super().__init__()
 
@@ -124,6 +125,12 @@ class Transformer2DModel(ModelMixin, ConfigMixin):
         self.in_channels = in_channels
         self.out_channels = in_channels if out_channels is None else out_channels
         self.gradient_checkpointing = False
+        if use_additional_conditions is None:
+            if norm_type == "ada_norm_single" and sample_size == 128:
+                use_additional_conditions = True
+            else:
+                use_additional_conditions = False
+        self.use_additional_conditions = use_additional_conditions
 
         # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
         # Define whether input is continuous or discrete depending on configuration
@@ -305,9 +312,7 @@ class Transformer2DModel(ModelMixin, ConfigMixin):
 
         # PixArt-Alpha blocks.
         self.adaln_single = None
-        self.use_additional_conditions = False
         if self.config.norm_type == "ada_norm_single":
-            self.use_additional_conditions = self.config.sample_size == 128
             # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
             # additional conditions until we find better name
             self.adaln_single = AdaLayerNormSingle(

src/diffusers/pipelines/__init__.py

@@ -187,7 +187,7 @@ else:
     _import_structure["musicldm"] = ["MusicLDMPipeline"]
     _import_structure["paint_by_example"] = ["PaintByExamplePipeline"]
     _import_structure["pia"] = ["PIAPipeline"]
-    _import_structure["pixart_alpha"] = ["PixArtAlphaPipeline"]
+    _import_structure["pixart_alpha"] = ["PixArtAlphaPipeline", "PixArtSigmaPipeline"]
     _import_structure["semantic_stable_diffusion"] = ["SemanticStableDiffusionPipeline"]
     _import_structure["shap_e"] = ["ShapEImg2ImgPipeline", "ShapEPipeline"]
     _import_structure["stable_cascade"] = [
@@ -450,7 +450,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .musicldm import MusicLDMPipeline
         from .paint_by_example import PaintByExamplePipeline
         from .pia import PIAPipeline
-        from .pixart_alpha import PixArtAlphaPipeline
+        from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline
         from .semantic_stable_diffusion import SemanticStableDiffusionPipeline
         from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline
         from .stable_cascade import (

src/diffusers/pipelines/pixart_alpha/__init__.py

@@ -23,6 +23,7 @@ except OptionalDependencyNotAvailable:
     _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
 else:
     _import_structure["pipeline_pixart_alpha"] = ["PixArtAlphaPipeline"]
+    _import_structure["pipeline_pixart_sigma"] = ["PixArtSigmaPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
@@ -32,7 +33,13 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     except OptionalDependencyNotAvailable:
         from ...utils.dummy_torch_and_transformers_objects import *
     else:
-        from .pipeline_pixart_alpha import PixArtAlphaPipeline
+        from .pipeline_pixart_alpha import (
+            ASPECT_RATIO_256_BIN,
+            ASPECT_RATIO_512_BIN,
+            ASPECT_RATIO_1024_BIN,
+            PixArtAlphaPipeline,
+        )
+        from .pipeline_pixart_sigma import ASPECT_RATIO_2048_BIN, PixArtSigmaPipeline
 
 else:
     import sys

src/diffusers/pipelines/pixart_alpha/pipeline_pixart_alpha.py

@@ -19,10 +19,9 @@ import urllib.parse as ul
 from typing import Callable, List, Optional, Tuple, Union
 
 import torch
-import torch.nn.functional as F
 from transformers import T5EncoderModel, T5Tokenizer
 
-from ...image_processor import VaeImageProcessor
+from ...image_processor import PixArtImageProcessor
 from ...models import AutoencoderKL, Transformer2DModel
 from ...schedulers import DPMSolverMultistepScheduler
 from ...utils import (
@@ -272,7 +271,7 @@ class PixArtAlphaPipeline(DiffusionPipeline):
         )
 
         self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
-        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor)
 
     # Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/utils.py
     def mask_text_embeddings(self, emb, mask):
@@ -674,38 +673,6 @@ class PixArtAlphaPipeline(DiffusionPipeline):
         latents = latents * self.scheduler.init_noise_sigma
         return latents
 
-    @staticmethod
-    def classify_height_width_bin(height: int, width: int, ratios: dict) -> Tuple[int, int]:
-        """Returns binned height and width."""
-        ar = float(height / width)
-        closest_ratio = min(ratios.keys(), key=lambda ratio: abs(float(ratio) - ar))
-        default_hw = ratios[closest_ratio]
-        return int(default_hw[0]), int(default_hw[1])
-
-    @staticmethod
-    def resize_and_crop_tensor(samples: torch.Tensor, new_width: int, new_height: int) -> torch.Tensor:
-        orig_height, orig_width = samples.shape[2], samples.shape[3]
-
-        # Check if resizing is needed
-        if orig_height != new_height or orig_width != new_width:
-            ratio = max(new_height / orig_height, new_width / orig_width)
-            resized_width = int(orig_width * ratio)
-            resized_height = int(orig_height * ratio)
-
-            # Resize
-            samples = F.interpolate(
-                samples, size=(resized_height, resized_width), mode="bilinear", align_corners=False
-            )
-
-            # Center Crop
-            start_x = (resized_width - new_width) // 2
-            end_x = start_x + new_width
-            start_y = (resized_height - new_height) // 2
-            end_y = start_y + new_height
-            samples = samples[:, :, start_y:end_y, start_x:end_x]
-
-        return samples
-
     @torch.no_grad()
     @replace_example_docstring(EXAMPLE_DOC_STRING)
     def __call__(
@@ -826,7 +793,7 @@ class PixArtAlphaPipeline(DiffusionPipeline):
             else:
                 raise ValueError("Invalid sample size")
             orig_height, orig_width = height, width
-            height, width = self.classify_height_width_bin(height, width, ratios=aspect_ratio_bin)
+            height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin)
 
         self.check_inputs(
             prompt,
@@ -956,7 +923,11 @@ class PixArtAlphaPipeline(DiffusionPipeline):
                     noise_pred = noise_pred
 
                 # compute previous image: x_t -> x_t-1
-                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+                if num_inference_steps == 1:
+                    # For DMD one step sampling: https://arxiv.org/abs/2311.18828
+                    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).pred_original_sample
+                else:
+                    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
 
                 # call the callback, if provided
                 if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
@@ -968,7 +939,7 @@ class PixArtAlphaPipeline(DiffusionPipeline):
         if not output_type == "latent":
             image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
             if use_resolution_binning:
-                image = self.resize_and_crop_tensor(image, orig_width, orig_height)
+                image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height)
         else:
             image = latents
 

src/diffusers/pipelines/pixart_alpha/pipeline_pixart_sigma.py

@@ -0,0 +1,866 @@
+# Copyright 2024 PixArt-Sigma Authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+import inspect
+import re
+import urllib.parse as ul
+from typing import Callable, List, Optional, Tuple, Union
+
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from ...image_processor import PixArtImageProcessor
+from ...models import AutoencoderKL, Transformer2DModel
+from ...schedulers import DPMSolverMultistepScheduler
+from ...utils import (
+    BACKENDS_MAPPING,
+    deprecate,
+    is_bs4_available,
+    is_ftfy_available,
+    logging,
+    replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from .pipeline_pixart_alpha import (
+    ASPECT_RATIO_256_BIN,
+    ASPECT_RATIO_512_BIN,
+    ASPECT_RATIO_1024_BIN,
+)
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+if is_bs4_available():
+    from bs4 import BeautifulSoup
+
+if is_ftfy_available():
+    import ftfy
+
+
+ASPECT_RATIO_2048_BIN = {
+    "0.25": [1024.0, 4096.0],
+    "0.26": [1024.0, 3968.0],
+    "0.27": [1024.0, 3840.0],
+    "0.28": [1024.0, 3712.0],
+    "0.32": [1152.0, 3584.0],
+    "0.33": [1152.0, 3456.0],
+    "0.35": [1152.0, 3328.0],
+    "0.4": [1280.0, 3200.0],
+    "0.42": [1280.0, 3072.0],
+    "0.48": [1408.0, 2944.0],
+    "0.5": [1408.0, 2816.0],
+    "0.52": [1408.0, 2688.0],
+    "0.57": [1536.0, 2688.0],
+    "0.6": [1536.0, 2560.0],
+    "0.68": [1664.0, 2432.0],
+    "0.72": [1664.0, 2304.0],
+    "0.78": [1792.0, 2304.0],
+    "0.82": [1792.0, 2176.0],
+    "0.88": [1920.0, 2176.0],
+    "0.94": [1920.0, 2048.0],
+    "1.0": [2048.0, 2048.0],
+    "1.07": [2048.0, 1920.0],
+    "1.13": [2176.0, 1920.0],
+    "1.21": [2176.0, 1792.0],
+    "1.29": [2304.0, 1792.0],
+    "1.38": [2304.0, 1664.0],
+    "1.46": [2432.0, 1664.0],
+    "1.67": [2560.0, 1536.0],
+    "1.75": [2688.0, 1536.0],
+    "2.0": [2816.0, 1408.0],
+    "2.09": [2944.0, 1408.0],
+    "2.4": [3072.0, 1280.0],
+    "2.5": [3200.0, 1280.0],
+    "2.89": [3328.0, 1152.0],
+    "3.0": [3456.0, 1152.0],
+    "3.11": [3584.0, 1152.0],
+    "3.62": [3712.0, 1024.0],
+    "3.75": [3840.0, 1024.0],
+    "3.88": [3968.0, 1024.0],
+    "4.0": [4096.0, 1024.0],
+}
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import PixArtSigmaPipeline
+
+        >>> # You can replace the checkpoint id with "PixArt-alpha/PixArt-Sigma-XL-2-512-MS" too.
+        >>> pipe = PixArtSigmaPipeline.from_pretrained(
+        ...     "PixArt-alpha/PixArt-Sigma-XL-2-1024-MS", torch_dtype=torch.float16
+        ... )
+        >>> # Enable memory optimizations.
+        >>> # pipe.enable_model_cpu_offload()
+
+        >>> prompt = "A small cactus with a happy face in the Sahara desert."
+        >>> image = pipe(prompt).images[0]
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+                Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+                timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+                must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class PixArtSigmaPipeline(DiffusionPipeline):
+    r"""
+    Pipeline for text-to-image generation using PixArt-Sigma.
+    """
+
+    bad_punct_regex = re.compile(
+        r"["
+        + "#®•©™&@·º½¾¿¡§~"
+        + r"\)"
+        + r"\("
+        + r"\]"
+        + r"\["
+        + r"\}"
+        + r"\{"
+        + r"\|"
+        + "\\"
+        + r"\/"
+        + r"\*"
+        + r"]{1,}"
+    )  # noqa
+
+    _optional_components = ["tokenizer", "text_encoder"]
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+
+    def __init__(
+        self,
+        tokenizer: T5Tokenizer,
+        text_encoder: T5EncoderModel,
+        vae: AutoencoderKL,
+        transformer: Transformer2DModel,
+        scheduler: DPMSolverMultistepScheduler,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+        )
+
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+    # copied from diffusers.pipelines.pixart_alpha.pipeline_pixart_alpha.py
+    def mask_text_embeddings(self, emb, mask):
+        if emb.shape[0] == 1:
+            keep_index = mask.sum().item()
+            return emb[:, :, :keep_index, :], keep_index
+        else:
+            masked_feature = emb * mask[:, None, :, None]
+            return masked_feature, emb.shape[2]
+
+    # Adapted from diffusers.pipelines.deepfloyd_if.pipeline_if.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: str = "",
+        num_images_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        prompt_attention_mask: Optional[torch.FloatTensor] = None,
+        negative_prompt_attention_mask: Optional[torch.FloatTensor] = None,
+        clean_caption: bool = False,
+        max_sequence_length: int = 120,
+        **kwargs,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds`
+                instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For
+                PixArt-Alpha, this should be "".
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                whether to use classifier free guidance or not
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                number of images that should be generated per prompt
+            device: (`torch.device`, *optional*):
+                torch device to place the resulting embeddings on
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Alpha, it's should be the embeddings of the ""
+                string.
+            clean_caption (`bool`, defaults to `False`):
+                If `True`, the function will preprocess and clean the provided caption before encoding.
+            max_sequence_length (`int`, defaults to 120): Maximum sequence length to use for the prompt.
+        """
+
+        if "mask_feature" in kwargs:
+            deprecation_message = "The use of `mask_feature` is deprecated. It is no longer used in any computation and that doesn't affect the end results. It will be removed in a future version."
+            deprecate("mask_feature", "1.0.0", deprecation_message, standard_warn=False)
+
+        if device is None:
+            device = self._execution_device
+
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # See Section 3.1. of the paper.
+        max_length = max_sequence_length
+
+        if prompt_embeds is None:
+            prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
+            text_inputs = self.tokenizer(
+                prompt,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                add_special_tokens=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids
+            untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                text_input_ids, untruncated_ids
+            ):
+                removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1])
+                logger.warning(
+                    "The following part of your input was truncated because CLIP can only handle sequences up to"
+                    f" {max_length} tokens: {removed_text}"
+                )
+
+            prompt_attention_mask = text_inputs.attention_mask
+            prompt_attention_mask = prompt_attention_mask.to(device)
+
+            prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask)
+            prompt_embeds = prompt_embeds[0]
+
+        if self.text_encoder is not None:
+            dtype = self.text_encoder.dtype
+        elif self.transformer is not None:
+            dtype = self.transformer.dtype
+        else:
+            dtype = None
+
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+        prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            uncond_tokens = [negative_prompt] * batch_size
+            uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
+            max_length = prompt_embeds.shape[1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_attention_mask=True,
+                add_special_tokens=True,
+                return_tensors="pt",
+            )
+            negative_prompt_attention_mask = uncond_input.attention_mask
+            negative_prompt_attention_mask = negative_prompt_attention_mask.to(device)
+
+            negative_prompt_embeds = self.text_encoder(
+                uncond_input.input_ids.to(device), attention_mask=negative_prompt_attention_mask
+            )
+            negative_prompt_embeds = negative_prompt_embeds[0]
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+            negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1)
+            negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1)
+        else:
+            negative_prompt_embeds = None
+            negative_prompt_attention_mask = None
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    # copied from diffusers.pipelines.pixart_alpha.pipeline_pixart_alpha.py
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        negative_prompt,
+        callback_steps,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_attention_mask=None,
+        negative_prompt_attention_mask=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and prompt_attention_mask is None:
+            raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
+
+        if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
+            raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+            if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
+                raise ValueError(
+                    "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
+                    f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
+                    f" {negative_prompt_attention_mask.shape}."
+                )
+
+    # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
+    def _text_preprocessing(self, text, clean_caption=False):
+        if clean_caption and not is_bs4_available():
+            logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
+            logger.warning("Setting `clean_caption` to False...")
+            clean_caption = False
+
+        if clean_caption and not is_ftfy_available():
+            logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
+            logger.warning("Setting `clean_caption` to False...")
+            clean_caption = False
+
+        if not isinstance(text, (tuple, list)):
+            text = [text]
+
+        def process(text: str):
+            if clean_caption:
+                text = self._clean_caption(text)
+                text = self._clean_caption(text)
+            else:
+                text = text.lower().strip()
+            return text
+
+        return [process(t) for t in text]
+
+    # Copied from diffusers.pipelines.pixart_alpha.pipeline_pixart_alpha.PixArtAlphaPipeline._clean_caption
+    def _clean_caption(self, caption):
+        caption = str(caption)
+        caption = ul.unquote_plus(caption)
+        caption = caption.strip().lower()
+        caption = re.sub("<person>", "person", caption)
+        # urls:
+        caption = re.sub(
+            r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))",  # noqa
+            "",
+            caption,
+        )  # regex for urls
+        caption = re.sub(
+            r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))",  # noqa
+            "",
+            caption,
+        )  # regex for urls
+        # html:
+        caption = BeautifulSoup(caption, features="html.parser").text
+
+        # @<nickname>
+        caption = re.sub(r"@[\w\d]+\b", "", caption)
+
+        # 31C0—31EF CJK Strokes
+        # 31F0—31FF Katakana Phonetic Extensions
+        # 3200—32FF Enclosed CJK Letters and Months
+        # 3300—33FF CJK Compatibility
+        # 3400—4DBF CJK Unified Ideographs Extension A
+        # 4DC0—4DFF Yijing Hexagram Symbols
+        # 4E00—9FFF CJK Unified Ideographs
+        caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
+        caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
+        caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
+        caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
+        caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
+        caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
+        caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
+        #######################################################
+
+        # все виды тире / all types of dash --> "-"
+        caption = re.sub(
+            r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+",  # noqa
+            "-",
+            caption,
+        )
+
+        # кавычки к одному стандарту
+        caption = re.sub(r"[`´«»“”¨]", '"', caption)
+        caption = re.sub(r"[‘’]", "'", caption)
+
+        # &quot;
+        caption = re.sub(r"&quot;?", "", caption)
+        # &amp
+        caption = re.sub(r"&amp", "", caption)
+
+        # ip adresses:
+        caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
+
+        # article ids:
+        caption = re.sub(r"\d:\d\d\s+$", "", caption)
+
+        # \n
+        caption = re.sub(r"\\n", " ", caption)
+
+        # "#123"
+        caption = re.sub(r"#\d{1,3}\b", "", caption)
+        # "#12345.."
+        caption = re.sub(r"#\d{5,}\b", "", caption)
+        # "123456.."
+        caption = re.sub(r"\b\d{6,}\b", "", caption)
+        # filenames:
+        caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
+
+        #
+        caption = re.sub(r"[\"\']{2,}", r'"', caption)  # """AUSVERKAUFT"""
+        caption = re.sub(r"[\.]{2,}", r" ", caption)  # """AUSVERKAUFT"""
+
+        caption = re.sub(self.bad_punct_regex, r" ", caption)  # ***AUSVERKAUFT***, #AUSVERKAUFT
+        caption = re.sub(r"\s+\.\s+", r" ", caption)  # " . "
+
+        # this-is-my-cute-cat / this_is_my_cute_cat
+        regex2 = re.compile(r"(?:\-|\_)")
+        if len(re.findall(regex2, caption)) > 3:
+            caption = re.sub(regex2, " ", caption)
+
+        caption = ftfy.fix_text(caption)
+        caption = html.unescape(html.unescape(caption))
+
+        caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption)  # jc6640
+        caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption)  # jc6640vc
+        caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption)  # 6640vc231
+
+        caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
+        caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
+        caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
+        caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
+        caption = re.sub(r"\bpage\s+\d+\b", "", caption)
+
+        caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption)  # j2d1a2a...
+
+        caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
+
+        caption = re.sub(r"\b\s+\:\s+", r": ", caption)
+        caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
+        caption = re.sub(r"\s+", " ", caption)
+
+        caption.strip()
+
+        caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
+        caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
+        caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
+        caption = re.sub(r"^\.\S+$", "", caption)
+
+        return caption.strip()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            int(height) // self.vae_scale_factor,
+            int(width) // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: str = "",
+        num_inference_steps: int = 20,
+        timesteps: List[int] = None,
+        guidance_scale: float = 4.5,
+        num_images_per_prompt: Optional[int] = 1,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        prompt_attention_mask: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_attention_mask: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        clean_caption: bool = True,
+        use_resolution_binning: bool = True,
+        max_sequence_length: int = 300,
+        **kwargs,
+    ) -> Union[ImagePipelineOutput, Tuple]:
+        """
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            num_inference_steps (`int`, *optional*, defaults to 100):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+                timesteps are used. Must be in descending order.
+            guidance_scale (`float`, *optional*, defaults to 4.5):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size):
+                The width in pixels of the generated image.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            prompt_attention_mask (`torch.FloatTensor`, *optional*): Pre-generated attention mask for text embeddings.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
+                Pre-generated attention mask for negative text embeddings.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            clean_caption (`bool`, *optional*, defaults to `True`):
+                Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
+                be installed. If the dependencies are not installed, the embeddings will be created from the raw
+                prompt.
+            use_resolution_binning (`bool` defaults to `True`):
+                If set to `True`, the requested height and width are first mapped to the closest resolutions using
+                `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to
+                the requested resolution. Useful for generating non-square images.
+            max_sequence_length (`int` defaults to 120): Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.ImagePipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+                returned where the first element is a list with the generated images
+        """
+        # 1. Check inputs. Raise error if not correct
+        height = height or self.transformer.config.sample_size * self.vae_scale_factor
+        width = width or self.transformer.config.sample_size * self.vae_scale_factor
+        if use_resolution_binning:
+            if self.transformer.config.sample_size == 256:
+                aspect_ratio_bin = ASPECT_RATIO_2048_BIN
+            elif self.transformer.config.sample_size == 128:
+                aspect_ratio_bin = ASPECT_RATIO_1024_BIN
+            elif self.transformer.config.sample_size == 64:
+                aspect_ratio_bin = ASPECT_RATIO_512_BIN
+            elif self.transformer.config.sample_size == 32:
+                aspect_ratio_bin = ASPECT_RATIO_256_BIN
+            else:
+                raise ValueError("Invalid sample size")
+            orig_height, orig_width = height, width
+            height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin)
+
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt,
+            callback_steps,
+            prompt_embeds,
+            negative_prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_attention_mask,
+        )
+
+        # 2. Default height and width to transformer
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt,
+            do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            device=device,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            clean_caption=clean_caption,
+            max_sequence_length=max_sequence_length,
+        )
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+        # 5. Prepare latents.
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            latent_channels,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 6.1 Prepare micro-conditions.
+        added_cond_kwargs = {"resolution": None, "aspect_ratio": None}
+
+        # 7. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                current_timestep = t
+                if not torch.is_tensor(current_timestep):
+                    # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+                    # This would be a good case for the `match` statement (Python 3.10+)
+                    is_mps = latent_model_input.device.type == "mps"
+                    if isinstance(current_timestep, float):
+                        dtype = torch.float32 if is_mps else torch.float64
+                    else:
+                        dtype = torch.int32 if is_mps else torch.int64
+                    current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device)
+                elif len(current_timestep.shape) == 0:
+                    current_timestep = current_timestep[None].to(latent_model_input.device)
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                current_timestep = current_timestep.expand(latent_model_input.shape[0])
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    encoder_attention_mask=prompt_attention_mask,
+                    timestep=current_timestep,
+                    added_cond_kwargs=added_cond_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # learned sigma
+                if self.transformer.config.out_channels // 2 == latent_channels:
+                    noise_pred = noise_pred.chunk(2, dim=1)[0]
+                else:
+                    noise_pred = noise_pred
+
+                # compute previous image: x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+        if not output_type == "latent":
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+            if use_resolution_binning:
+                image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height)
+        else:
+            image = latents
+
+        if not output_type == "latent":
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return ImagePipelineOutput(images=image)

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -737,6 +737,21 @@ class PixArtAlphaPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class PixArtSigmaPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class SemanticStableDiffusionPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/pipelines/pixart_alpha/test_pixart.py

@@ -324,6 +324,10 @@ class PixArtAlphaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
     def test_inference_batch_single_identical(self):
         self._test_inference_batch_single_identical(expected_max_diff=1e-3)
 
+    # PixArt transformer model does not work with sequential offload so skip it for now
+    def test_sequential_offload_forward_pass_twice(self):
+        pass
+
 
 @slow
 @require_torch_gpu

tests/pipelines/pixart_sigma/test_pixart.py

@@ -0,0 +1,430 @@
+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import gc
+import tempfile
+import unittest
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer, T5EncoderModel
+
+from diffusers import (
+    AutoencoderKL,
+    DDIMScheduler,
+    PixArtSigmaPipeline,
+    Transformer2DModel,
+)
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    numpy_cosine_similarity_distance,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import PipelineTesterMixin, to_np
+
+
+enable_full_determinism()
+
+
+class PixArtSigmaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = PixArtSigmaPipeline
+    params = TEXT_TO_IMAGE_PARAMS - {"cross_attention_kwargs"}
+    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+
+    required_optional_params = PipelineTesterMixin.required_optional_params
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = Transformer2DModel(
+            sample_size=8,
+            num_layers=2,
+            patch_size=2,
+            attention_head_dim=8,
+            num_attention_heads=3,
+            caption_channels=32,
+            in_channels=4,
+            cross_attention_dim=24,
+            out_channels=8,
+            attention_bias=True,
+            activation_fn="gelu-approximate",
+            num_embeds_ada_norm=1000,
+            norm_type="ada_norm_single",
+            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")
+
+        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        components = {
+            "transformer": transformer.eval(),
+            "vae": vae.eval(),
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 5.0,
+            "use_resolution_binning": False,
+            "output_type": "np",
+        }
+        return inputs
+
+    def test_sequential_cpu_offload_forward_pass(self):
+        # TODO(PVP, Sayak) need to fix later
+        return
+
+    def test_save_load_optional_components(self):
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        prompt = inputs["prompt"]
+        generator = inputs["generator"]
+        num_inference_steps = inputs["num_inference_steps"]
+        output_type = inputs["output_type"]
+
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = pipe.encode_prompt(prompt)
+
+        # inputs with prompt converted to embeddings
+        inputs = {
+            "prompt_embeds": prompt_embeds,
+            "prompt_attention_mask": prompt_attention_mask,
+            "negative_prompt": None,
+            "negative_prompt_embeds": negative_prompt_embeds,
+            "negative_prompt_attention_mask": negative_prompt_attention_mask,
+            "generator": generator,
+            "num_inference_steps": num_inference_steps,
+            "output_type": output_type,
+            "use_resolution_binning": False,
+        }
+
+        # set all optional components to None
+        for optional_component in pipe._optional_components:
+            setattr(pipe, optional_component, None)
+
+        output = pipe(**inputs)[0]
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            pipe.save_pretrained(tmpdir)
+            pipe_loaded = self.pipeline_class.from_pretrained(tmpdir)
+            pipe_loaded.to(torch_device)
+            pipe_loaded.set_progress_bar_config(disable=None)
+
+        for optional_component in pipe._optional_components:
+            self.assertTrue(
+                getattr(pipe_loaded, optional_component) is None,
+                f"`{optional_component}` did not stay set to None after loading.",
+            )
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        generator = inputs["generator"]
+        num_inference_steps = inputs["num_inference_steps"]
+        output_type = inputs["output_type"]
+
+        # inputs with prompt converted to embeddings
+        inputs = {
+            "prompt_embeds": prompt_embeds,
+            "prompt_attention_mask": prompt_attention_mask,
+            "negative_prompt": None,
+            "negative_prompt_embeds": negative_prompt_embeds,
+            "negative_prompt_attention_mask": negative_prompt_attention_mask,
+            "generator": generator,
+            "num_inference_steps": num_inference_steps,
+            "output_type": output_type,
+            "use_resolution_binning": False,
+        }
+
+        output_loaded = pipe_loaded(**inputs)[0]
+
+        max_diff = np.abs(to_np(output) - to_np(output_loaded)).max()
+        self.assertLess(max_diff, 1e-4)
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs).images
+        image_slice = image[0, -3:, -3:, -1]
+
+        self.assertEqual(image.shape, (1, 8, 8, 3))
+        expected_slice = np.array([0.6319, 0.3526, 0.3806, 0.6327, 0.4639, 0.4830, 0.2583, 0.5331, 0.4852])
+        max_diff = np.abs(image_slice.flatten() - expected_slice).max()
+        self.assertLessEqual(max_diff, 1e-3)
+
+    def test_inference_non_square_images(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        image = pipe(**inputs, height=32, width=48).images
+        image_slice = image[0, -3:, -3:, -1]
+        self.assertEqual(image.shape, (1, 32, 48, 3))
+
+        expected_slice = np.array([0.6493, 0.5370, 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)
+
+    def test_inference_with_embeddings_and_multiple_images(self):
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        prompt = inputs["prompt"]
+        generator = inputs["generator"]
+        num_inference_steps = inputs["num_inference_steps"]
+        output_type = inputs["output_type"]
+
+        prompt_embeds, prompt_attn_mask, negative_prompt_embeds, neg_prompt_attn_mask = pipe.encode_prompt(prompt)
+
+        # inputs with prompt converted to embeddings
+        inputs = {
+            "prompt_embeds": prompt_embeds,
+            "prompt_attention_mask": prompt_attn_mask,
+            "negative_prompt": None,
+            "negative_prompt_embeds": negative_prompt_embeds,
+            "negative_prompt_attention_mask": neg_prompt_attn_mask,
+            "generator": generator,
+            "num_inference_steps": num_inference_steps,
+            "output_type": output_type,
+            "num_images_per_prompt": 2,
+            "use_resolution_binning": False,
+        }
+
+        # set all optional components to None
+        for optional_component in pipe._optional_components:
+            setattr(pipe, optional_component, None)
+
+        output = pipe(**inputs)[0]
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            pipe.save_pretrained(tmpdir)
+            pipe_loaded = self.pipeline_class.from_pretrained(tmpdir)
+            pipe_loaded.to(torch_device)
+            pipe_loaded.set_progress_bar_config(disable=None)
+
+        for optional_component in pipe._optional_components:
+            self.assertTrue(
+                getattr(pipe_loaded, optional_component) is None,
+                f"`{optional_component}` did not stay set to None after loading.",
+            )
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        generator = inputs["generator"]
+        num_inference_steps = inputs["num_inference_steps"]
+        output_type = inputs["output_type"]
+
+        # inputs with prompt converted to embeddings
+        inputs = {
+            "prompt_embeds": prompt_embeds,
+            "prompt_attention_mask": prompt_attn_mask,
+            "negative_prompt": None,
+            "negative_prompt_embeds": negative_prompt_embeds,
+            "negative_prompt_attention_mask": neg_prompt_attn_mask,
+            "generator": generator,
+            "num_inference_steps": num_inference_steps,
+            "output_type": output_type,
+            "num_images_per_prompt": 2,
+            "use_resolution_binning": False,
+        }
+
+        output_loaded = pipe_loaded(**inputs)[0]
+
+        max_diff = np.abs(to_np(output) - to_np(output_loaded)).max()
+        self.assertLess(max_diff, 1e-4)
+
+    def test_inference_with_multiple_images_per_prompt(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["num_images_per_prompt"] = 2
+        image = pipe(**inputs).images
+        image_slice = image[0, -3:, -3:, -1]
+
+        self.assertEqual(image.shape, (2, 8, 8, 3))
+        expected_slice = np.array([0.6319, 0.3526, 0.3806, 0.6327, 0.4639, 0.4830, 0.2583, 0.5331, 0.4852])
+        max_diff = np.abs(image_slice.flatten() - expected_slice).max()
+        self.assertLessEqual(max_diff, 1e-3)
+
+    def test_inference_batch_single_identical(self):
+        self._test_inference_batch_single_identical(expected_max_diff=1e-3)
+
+    # PixArt transformer model does not work with sequential offload so skip it for now
+    def test_sequential_offload_forward_pass_twice(self):
+        pass
+
+
+@slow
+@require_torch_gpu
+class PixArtSigmaPipelineIntegrationTests(unittest.TestCase):
+    ckpt_id_1024 = "PixArt-alpha/PixArt-Sigma-XL-2-1024-MS"
+    ckpt_id_512 = "PixArt-alpha/PixArt-Sigma-XL-2-512-MS"
+    prompt = "A small cactus with a happy face in the Sahara desert."
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_pixart_1024(self):
+        generator = torch.Generator("cpu").manual_seed(0)
+
+        pipe = PixArtSigmaPipeline.from_pretrained(self.ckpt_id_1024, torch_dtype=torch.float16)
+        pipe.enable_model_cpu_offload()
+        prompt = self.prompt
+
+        image = pipe(prompt, generator=generator, num_inference_steps=2, output_type="np").images
+
+        image_slice = image[0, -3:, -3:, -1]
+        expected_slice = np.array([0.0742, 0.0835, 0.2114, 0.0295, 0.0784, 0.2361, 0.1738, 0.2251, 0.3589])
+
+        max_diff = numpy_cosine_similarity_distance(image_slice.flatten(), expected_slice)
+        self.assertLessEqual(max_diff, 1e-4)
+
+    def test_pixart_512(self):
+        generator = torch.Generator("cpu").manual_seed(0)
+
+        pipe = PixArtSigmaPipeline.from_pretrained(self.ckpt_id_512, torch_dtype=torch.float16)
+        pipe.enable_model_cpu_offload()
+
+        prompt = self.prompt
+
+        image = pipe(prompt, generator=generator, num_inference_steps=2, output_type="np").images
+
+        image_slice = image[0, -3:, -3:, -1]
+        expected_slice = np.array([0.3477, 0.3882, 0.4541, 0.3413, 0.3821, 0.4463, 0.4001, 0.4409, 0.4958])
+
+        max_diff = numpy_cosine_similarity_distance(image_slice.flatten(), expected_slice)
+        self.assertLessEqual(max_diff, 1e-4)
+
+    def test_pixart_1024_without_resolution_binning(self):
+        generator = torch.manual_seed(0)
+
+        pipe = PixArtSigmaPipeline.from_pretrained(self.ckpt_id_1024, torch_dtype=torch.float16)
+        pipe.enable_model_cpu_offload()
+
+        prompt = self.prompt
+        height, width = 1024, 768
+        num_inference_steps = 2
+
+        image = pipe(
+            prompt,
+            height=height,
+            width=width,
+            generator=generator,
+            num_inference_steps=num_inference_steps,
+            output_type="np",
+        ).images
+        image_slice = image[0, -3:, -3:, -1]
+
+        generator = torch.manual_seed(0)
+        no_res_bin_image = pipe(
+            prompt,
+            height=height,
+            width=width,
+            generator=generator,
+            num_inference_steps=num_inference_steps,
+            output_type="np",
+            use_resolution_binning=False,
+        ).images
+        no_res_bin_image_slice = no_res_bin_image[0, -3:, -3:, -1]
+
+        assert not np.allclose(image_slice, no_res_bin_image_slice, atol=1e-4, rtol=1e-4)
+
+    def test_pixart_512_without_resolution_binning(self):
+        generator = torch.manual_seed(0)
+
+        pipe = PixArtSigmaPipeline.from_pretrained(self.ckpt_id_512, torch_dtype=torch.float16)
+        pipe.enable_model_cpu_offload()
+
+        prompt = self.prompt
+        height, width = 512, 768
+        num_inference_steps = 2
+
+        image = pipe(
+            prompt,
+            height=height,
+            width=width,
+            generator=generator,
+            num_inference_steps=num_inference_steps,
+            output_type="np",
+        ).images
+        image_slice = image[0, -3:, -3:, -1]
+
+        generator = torch.manual_seed(0)
+        no_res_bin_image = pipe(
+            prompt,
+            height=height,
+            width=width,
+            generator=generator,
+            num_inference_steps=num_inference_steps,
+            output_type="np",
+            use_resolution_binning=False,
+        ).images
+        no_res_bin_image_slice = no_res_bin_image[0, -3:, -3:, -1]
+
+        assert not np.allclose(image_slice, no_res_bin_image_slice, atol=1e-4, rtol=1e-4)