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

+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,
@@ -956,6 +923,10 @@ class PixArtAlphaPipeline(DiffusionPipeline):
                     noise_pred = noise_pred
 
                 # compute previous image: x_t -> x_t-1
+                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
@@ -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/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

+# 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)