Merge branch 'main' into ltx-rename

.github/workflows/push_tests_mps.yml

@@ -46,7 +46,7 @@ jobs:
       shell: arch -arch arm64 bash {0}
       run: |
         ${CONDA_RUN} python -m pip install --upgrade pip uv
-        ${CONDA_RUN} python -m uv pip install -e [quality,test]
+        ${CONDA_RUN} python -m uv pip install -e ".[quality,test]"
         ${CONDA_RUN} python -m uv pip install torch torchvision torchaudio
         ${CONDA_RUN} python -m uv pip install accelerate@git+https://github.com/huggingface/accelerate.git
         ${CONDA_RUN} python -m uv pip install transformers --upgrade

docs/source/en/api/models/autoencoder_kl_hunyuan_video.md

@@ -18,7 +18,7 @@ The model can be loaded with the following code snippet.
 ```python
 from diffusers import AutoencoderKLHunyuanVideo
 
-vae = AutoencoderKLHunyuanVideo.from_pretrained("tencent/HunyuanVideo", torch_dtype=torch.float16)
+vae = AutoencoderKLHunyuanVideo.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder="vae", torch_dtype=torch.float16)
 ```
 
 ## AutoencoderKLHunyuanVideo

docs/source/en/api/models/hunyuan_video_transformer_3d.md

@@ -18,7 +18,7 @@ The model can be loaded with the following code snippet.
 ```python
 from diffusers import HunyuanVideoTransformer3DModel
 
-transformer = HunyuanVideoTransformer3DModel.from_pretrained("tencent/HunyuanVideo", torch_dtype=torch.bfloat16)
+transformer = HunyuanVideoTransformer3DModel.from_pretrained("hunyuanvideo-community/HunyuanVideo", subfolder="transformer", torch_dtype=torch.bfloat16)
 ```
 
 ## HunyuanVideoTransformer3DModel

docs/source/en/api/models/sana_transformer2d.md

@@ -22,7 +22,7 @@ The model can be loaded with the following code snippet.
 ```python
 from diffusers import SanaTransformer2DModel
 
-transformer = SanaTransformer2DModel.from_pretrained("Efficient-Large-Model/Sana_1600M_1024px_diffusers", subfolder="transformer", torch_dtype=torch.float16)
+transformer = SanaTransformer2DModel.from_pretrained("Efficient-Large-Model/Sana_1600M_1024px_BF16_diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
 ```
 
 ## SanaTransformer2DModel

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

@@ -29,7 +29,7 @@ Recommendations for inference:
 - Transformer should be in `torch.bfloat16`.
 - VAE should be in `torch.float16`.
 - `num_frames` should be of the form `4 * k + 1`, for example `49` or `129`.
-- For smaller resolution images, try lower values of `shift` (between `2.0` to `5.0`) in the [Scheduler](https://huggingface.co/docs/diffusers/main/en/api/schedulers/flow_match_euler_discrete#diffusers.FlowMatchEulerDiscreteScheduler.shift). For larger resolution images, try higher values (between `7.0` and `12.0`). The default value is `7.0` for HunyuanVideo.
+- For smaller resolution videos, try lower values of `shift` (between `2.0` to `5.0`) in the [Scheduler](https://huggingface.co/docs/diffusers/main/en/api/schedulers/flow_match_euler_discrete#diffusers.FlowMatchEulerDiscreteScheduler.shift). For larger resolution images, try higher values (between `7.0` and `12.0`). The default value is `7.0` for HunyuanVideo.
 - For more information about supported resolutions and other details, please refer to the original repository [here](https://github.com/Tencent/HunyuanVideo/).
 
 ## HunyuanVideoPipeline

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

@@ -32,9 +32,9 @@ Available models:
 
 | Model | Recommended dtype |
 |:-----:|:-----------------:|
+| [`Efficient-Large-Model/Sana_1600M_1024px_BF16_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px_BF16_diffusers) | `torch.bfloat16` |
 | [`Efficient-Large-Model/Sana_1600M_1024px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px_diffusers) | `torch.float16` |
 | [`Efficient-Large-Model/Sana_1600M_1024px_MultiLing_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px_MultiLing_diffusers) | `torch.float16` |
-| [`Efficient-Large-Model/Sana_1600M_1024px_BF16_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px_BF16_diffusers) | `torch.bfloat16` |
 | [`Efficient-Large-Model/Sana_1600M_512px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_512px_diffusers) | `torch.float16` |
 | [`Efficient-Large-Model/Sana_1600M_512px_MultiLing_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_1600M_512px_MultiLing_diffusers) | `torch.float16` |
 | [`Efficient-Large-Model/Sana_600M_1024px_diffusers`](https://huggingface.co/Efficient-Large-Model/Sana_600M_1024px_diffusers) | `torch.float16` |

docs/source/en/quantization/torchao.md

@@ -27,7 +27,7 @@ The example below only quantizes the weights to int8.
 ```python
 from diffusers import FluxPipeline, FluxTransformer2DModel, TorchAoConfig
 
-model_id = "black-forest-labs/Flux.1-Dev"
+model_id = "black-forest-labs/FLUX.1-dev"
 dtype = torch.bfloat16
 
 quantization_config = TorchAoConfig("int8wo")
@@ -45,7 +45,9 @@ pipe = FluxPipeline.from_pretrained(
 pipe.to("cuda")
 
 prompt = "A cat holding a sign that says hello world"
-image = pipe(prompt, num_inference_steps=28, guidance_scale=0.0).images[0]
+image = pipe(
+    prompt, num_inference_steps=50, guidance_scale=4.5, max_sequence_length=512
+).images[0]
 image.save("output.png")
 ```
 

scripts/convert_flux_xlabs_ipadapter_to_diffusers.py

@@ -0,0 +1,97 @@
+import argparse
+from contextlib import nullcontext
+
+import safetensors.torch
+from accelerate import init_empty_weights
+from huggingface_hub import hf_hub_download
+
+from diffusers.utils.import_utils import is_accelerate_available, is_transformers_available
+
+
+if is_transformers_available():
+    from transformers import CLIPVisionModelWithProjection
+
+    vision = True
+else:
+    vision = False
+
+"""
+python scripts/convert_flux_xlabs_ipadapter_to_diffusers.py  \
+--original_state_dict_repo_id "XLabs-AI/flux-ip-adapter" \
+--filename "flux-ip-adapter.safetensors"
+--output_path "flux-ip-adapter-hf/"
+"""
+
+
+CTX = init_empty_weights if is_accelerate_available else nullcontext
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--original_state_dict_repo_id", default=None, type=str)
+parser.add_argument("--filename", default="flux.safetensors", type=str)
+parser.add_argument("--checkpoint_path", default=None, type=str)
+parser.add_argument("--output_path", type=str)
+parser.add_argument("--vision_pretrained_or_path", default="openai/clip-vit-large-patch14", type=str)
+
+args = parser.parse_args()
+
+
+def load_original_checkpoint(args):
+    if args.original_state_dict_repo_id is not None:
+        ckpt_path = hf_hub_download(repo_id=args.original_state_dict_repo_id, filename=args.filename)
+    elif args.checkpoint_path is not None:
+        ckpt_path = args.checkpoint_path
+    else:
+        raise ValueError(" please provide either `original_state_dict_repo_id` or a local `checkpoint_path`")
+
+    original_state_dict = safetensors.torch.load_file(ckpt_path)
+    return original_state_dict
+
+
+def convert_flux_ipadapter_checkpoint_to_diffusers(original_state_dict, num_layers):
+    converted_state_dict = {}
+
+    # image_proj
+    ## norm
+    converted_state_dict["image_proj.norm.weight"] = original_state_dict.pop("ip_adapter_proj_model.norm.weight")
+    converted_state_dict["image_proj.norm.bias"] = original_state_dict.pop("ip_adapter_proj_model.norm.bias")
+    ## proj
+    converted_state_dict["image_proj.proj.weight"] = original_state_dict.pop("ip_adapter_proj_model.norm.weight")
+    converted_state_dict["image_proj.proj.bias"] = original_state_dict.pop("ip_adapter_proj_model.norm.bias")
+
+    # double transformer blocks
+    for i in range(num_layers):
+        block_prefix = f"ip_adapter.{i}."
+        # to_k_ip
+        converted_state_dict[f"{block_prefix}to_k_ip.bias"] = original_state_dict.pop(
+            f"double_blocks.{i}.processor.ip_adapter_double_stream_k_proj.bias"
+        )
+        converted_state_dict[f"{block_prefix}to_k_ip.weight"] = original_state_dict.pop(
+            f"double_blocks.{i}.processor.ip_adapter_double_stream_k_proj.weight"
+        )
+        # to_v_ip
+        converted_state_dict[f"{block_prefix}to_v_ip.bias"] = original_state_dict.pop(
+            f"double_blocks.{i}.processor.ip_adapter_double_stream_v_proj.bias"
+        )
+        converted_state_dict[f"{block_prefix}to_k_ip.weight"] = original_state_dict.pop(
+            f"double_blocks.{i}.processor.ip_adapter_double_stream_v_proj.weight"
+        )
+
+    return converted_state_dict
+
+
+def main(args):
+    original_ckpt = load_original_checkpoint(args)
+
+    num_layers = 19
+    converted_ip_adapter_state_dict = convert_flux_ipadapter_checkpoint_to_diffusers(original_ckpt, num_layers)
+
+    print("Saving Flux IP-Adapter in Diffusers format.")
+    safetensors.torch.save_file(converted_ip_adapter_state_dict, f"{args.output_path}/model.safetensors")
+
+    if vision:
+        model = CLIPVisionModelWithProjection.from_pretrained(args.vision_pretrained_or_path)
+        model.save_pretrained(f"{args.output_path}/image_encoder")
+
+
+if __name__ == "__main__":
+    main(args)

scripts/convert_sana_to_diffusers.py

@@ -88,13 +88,18 @@ def main(args):
     # y norm
     converted_state_dict["caption_norm.weight"] = state_dict.pop("attention_y_norm.weight")
 
+    # scheduler
     flow_shift = 3.0
+
+    # model config
     if args.model_type == "SanaMS_1600M_P1_D20":
         layer_num = 20
     elif args.model_type == "SanaMS_600M_P1_D28":
         layer_num = 28
     else:
         raise ValueError(f"{args.model_type} is not supported.")
+    # Positional embedding interpolation scale.
+    interpolation_scale = {512: None, 1024: None, 2048: 1.0}
 
     for depth in range(layer_num):
         # Transformer blocks.
@@ -176,6 +181,7 @@ def main(args):
             patch_size=1,
             norm_elementwise_affine=False,
             norm_eps=1e-6,
+            interpolation_scale=interpolation_scale[args.image_size],
         )
 
     if is_accelerate_available():

src/diffusers/loaders/__init__.py

@@ -55,7 +55,7 @@ def text_encoder_attn_modules(text_encoder):
 
 if is_torch_available():
     _import_structure["single_file_model"] = ["FromOriginalModelMixin"]
-
+    _import_structure["transformer_flux"] = ["FluxTransformer2DLoadersMixin"]
     _import_structure["transformer_sd3"] = ["SD3Transformer2DLoadersMixin"]
     _import_structure["unet"] = ["UNet2DConditionLoadersMixin"]
     _import_structure["utils"] = ["AttnProcsLayers"]
@@ -77,6 +77,7 @@ def text_encoder_attn_modules(text_encoder):
         _import_structure["textual_inversion"] = ["TextualInversionLoaderMixin"]
         _import_structure["ip_adapter"] = [
             "IPAdapterMixin",
+            "FluxIPAdapterMixin",
             "SD3IPAdapterMixin",
         ]
 
@@ -86,12 +87,14 @@ def text_encoder_attn_modules(text_encoder):
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     if is_torch_available():
         from .single_file_model import FromOriginalModelMixin
+        from .transformer_flux import FluxTransformer2DLoadersMixin
         from .transformer_sd3 import SD3Transformer2DLoadersMixin
         from .unet import UNet2DConditionLoadersMixin
         from .utils import AttnProcsLayers
 
         if is_transformers_available():
             from .ip_adapter import (
+                FluxIPAdapterMixin,
                 IPAdapterMixin,
                 SD3IPAdapterMixin,
             )