[Feat] PixArt-Alpha

docs/source/en/_toctree.yml

@@ -268,6 +268,8 @@
       title: Parallel Sampling of Diffusion Models
     - local: api/pipelines/pix2pix_zero
       title: Pix2Pix Zero
+    - local: api/pipelines/pixart
+      title: PixArt
     - local: api/pipelines/pndm
       title: PNDM
     - local: api/pipelines/repaint

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

@@ -0,0 +1,36 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# PixArt
+
+![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/pixart/header_collage.png)
+
+[PixArt-α: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis](https://huggingface.co/papers/2310.00426) is Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li.
+
+The abstract from the paper is:
+
+*The most advanced text-to-image (T2I) models require significant training costs (e.g., millions of GPU hours), seriously hindering the fundamental innovation for the AIGC community while increasing CO2 emissions. This paper introduces PIXART-α, a Transformer-based T2I diffusion model whose image generation quality is competitive with state-of-the-art image generators (e.g., Imagen, SDXL, and even Midjourney), reaching near-commercial application standards. Additionally, it supports high-resolution image synthesis up to 1024px resolution with low training cost, as shown in Figure 1 and 2. To achieve this goal, three core designs are proposed: (1) Training strategy decomposition: We devise three distinct training steps that separately optimize pixel dependency, text-image alignment, and image aesthetic quality; (2) Efficient T2I Transformer: We incorporate cross-attention modules into Diffusion Transformer (DiT) to inject text conditions and streamline the computation-intensive class-condition branch; (3) High-informative data: We emphasize the significance of concept density in text-image pairs and leverage a large Vision-Language model to auto-label dense pseudo-captions to assist text-image alignment learning. As a result, PIXART-α's training speed markedly surpasses existing large-scale T2I models, e.g., PIXART-α only takes 10.8% of Stable Diffusion v1.5's training time (675 vs. 6,250 A100 GPU days), saving nearly $300,000 ($26,000 vs. $320,000) and reducing 90% CO2 emissions. Moreover, compared with a larger SOTA model, RAPHAEL, our training cost is merely 1%. Extensive experiments demonstrate that PIXART-α excels in image quality, artistry, and semantic control. We hope PIXART-α will provide new insights to the AIGC community and startups to accelerate building their own high-quality yet low-cost generative models from scratch.*
+
+You can find the original codebase at [PixArt-alpha/PixArt-alpha](https://github.com/PixArt-alpha/PixArt-alpha) and all the available checkpoints at [PixArt-alpha](https://huggingface.co/PixArt-alpha).
+
+Some notes about this pipeline:
+
+* It uses a Transformer backbone (instead of a UNet) for denoising. As such it has a similar architecture as [DiT](./dit.md).
+* It was trained using text conditions computed from T5. This aspect makes the pipeline better at following complex text prompts with intricate details. 
+* It is good at producing high-resolution images at different aspect ratios. To get the best results, the authors recommend some size brackets which can be found [here](https://github.com/PixArt-alpha/PixArt-alpha/blob/08fbbd281ec96866109bdd2cdb75f2f58fb17610/diffusion/data/datasets/utils.py).
+* It rivals the quality of state-of-the-art text-to-image generation systems (as of this writing) such as Stable Diffusion XL, Imagen, and DALL-E 2, while being more efficient than them.
+
+## PixArtAlphaPipeline
+
+[[autodoc]] PixArtAlphaPipeline
+	- all
+	- __call__

scripts/convert_pixart_alpha_to_diffusers.py

@@ -0,0 +1,198 @@
+import argparse
+import os
+
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from diffusers import AutoencoderKL, DPMSolverMultistepScheduler, PixArtAlphaPipeline, Transformer2DModel
+
+
+ckpt_id = "PixArt-alpha/PixArt-alpha"
+# https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/scripts/inference.py#L125
+interpolation_scale = {512: 1, 1024: 2}
+
+
+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.y_embedding"] = state_dict.pop("y_embedder.y_embedding")
+    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.image_size == 1024:
+        # 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"
+        )
+
+        # 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")
+
+    # DiT 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,
+    )
+    transformer.load_state_dict(converted_state_dict, strict=True)
+
+    assert transformer.pos_embed.pos_embed is not None
+    state_dict.pop("pos_embed")
+    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:
+        scheduler = DPMSolverMultistepScheduler()
+
+        vae = AutoencoderKL.from_pretrained(ckpt_id, subfolder="sd-vae-ft-ema")
+
+        tokenizer = T5Tokenizer.from_pretrained(ckpt_id, subfolder="t5-v1_1-xxl")
+        text_encoder = T5EncoderModel.from_pretrained(ckpt_id, subfolder="t5-v1_1-xxl")
+
+        pipeline = PixArtAlphaPipeline(
+            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(
+        "--orig_ckpt_path", default=None, type=str, required=False, help="Path to the checkpoint to convert."
+    )
+    parser.add_argument(
+        "--image_size",
+        default=1024,
+        type=int,
+        choices=[512, 1024],
+        required=False,
+        help="Image size of pretrained model, either 512 or 1024.",
+    )
+    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

@@ -235,6 +235,7 @@
             "LDMTextToImagePipeline",
             "MusicLDMPipeline",
             "PaintByExamplePipeline",
+            "PixArtAlphaPipeline",
             "SemanticStableDiffusionPipeline",
             "ShapEImg2ImgPipeline",
             "ShapEPipeline",
@@ -579,6 +580,7 @@
             LDMTextToImagePipeline,
             MusicLDMPipeline,
             PaintByExamplePipeline,
+            PixArtAlphaPipeline,
             SemanticStableDiffusionPipeline,
             ShapEImg2ImgPipeline,
             ShapEPipeline,