init-1

.gitignore

@@ -0,0 +1,34 @@
+output/
+models/
+._____temp/
+.vscode/
+*__pycache__/*
+*.pyc
+video/
+*.safetensors
+*.pth
+omni/
+pcd/
+
+models/
+models_ms/
+!diffsynth/models**
+diffsynth/models/__pycache__/
+ckpt/DVD/
+outputs/
+inference_results/
+*.mp4
+ckpt/DVD
+.msc
+.mv
+ckpt/.cache
+ckpt/.gitattributes
+ckpt/README.md
+overlap_plots/
+test_script/test_from_trained_all_vid_test.py
+test_script/test_single_video_batch.py
+DVD.egg-info/
+infer_bash/video_test.sh
+ckpt/test/
+!demo/robot_navi.mp4
+!demo/drone.mp4

README.md

@@ -5,10 +5,25 @@ colorFrom: green
 colorTo: yellow
 sdk: gradio
 sdk_version: 6.9.0
+python_version: 3.10.20
 app_file: app.py
 pinned: false
 license: apache-2.0
 short_description: Official demo of DVD (https://dvd-project.github.io/)
+tags:
+  - video diffusion
+  - video depth estimation
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+
+If you find our work useful in your research, please consider citing our paper🌹:
+
+```
+@article{zhang2026dvd,
+  title={DVD: Deterministic Video Depth Estimation with Generative Priors},
+  author={Zhang, Hongfei and Chen, Harold Haodong and Liao, Chenfei and He, Jing and Zhang, Zixin and Li, Haodong and Liang, Yihao and Chen, Kanghao and Ren, Bin and Zheng, Xu and Yang, Shuai and Zhou, Kun and Li, Yinchuan and Sebe, Nicu and Chen, Ying-Cong},
+  journal={arXiv preprint arXiv:2603.12250},
+  year={2026}
+}
+```

app.py

@@ -0,0 +1,77 @@
+import spaces  # must be first!
+import os
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parent
+GRADIO_TMP = REPO_ROOT / ".gradio_cache"
+GRADIO_TMP.mkdir(parents=True, exist_ok=True)
+
+os.environ["GRADIO_TEMP_DIR"] = str(GRADIO_TMP)
+print(f"Gradio temp/cache dir: {GRADIO_TMP}")
+
+import torch
+from argparse import Namespace
+import subprocess
+from test_script.test_single_video import *
+
+import gradio as gr
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+yaml_args = OmegaConf.load(f"{REPO_ROOT}/ckpt/model_config.yaml")
+pipeline = None
+
+
+@spaces.GPU
+def fn(input_video):
+    global pipeline, yaml_args, device
+    if pipeline is None:
+        if not os.path.exists(f"{REPO_ROOT}/ckpt/model.safetensors"):
+            subprocess.run(["bash", f"{REPO_ROOT}/infer_bash/download_ckpt.sh"], check=True)
+        pipeline = load_model(f"{REPO_ROOT}/ckpt", yaml_args)
+    
+    input_video_basename = os.path.basename(input_video)
+    input_tensor, orig_size, origin_fps = load_video_data(Namespace(
+        input_video=input_video,
+        height=480,
+        width=640,
+    ))
+    depth = predict_depth(pipeline, input_tensor, orig_size, Namespace(
+        window_size=81,
+        overlap=21
+    ))
+    output_video = save_results(depth, origin_fps, Namespace(
+        input_video=input_video,
+        output_dir=REPO_ROOT,
+        grayscale=False
+    ))
+
+    return output_video
+
+
+if __name__ == "__main__":
+    inputs = [
+        gr.Video(label="Input Video", autoplay=True),
+    ]
+    outputs = [
+        gr.Video(label="Output Video", autoplay=True),
+    ]
+
+    demo = gr.Interface(
+        fn=fn,
+        title="DVD: Deterministic Video Depth Estimation with Generative Priors",
+        description="""
+            <strong>Please consider starring <span style="color: orange">&#9733;</span> our <a href="https://github.com/EnVision-Research/DVD" target="_blank" rel="noopener noreferrer">GitHub Repo</a> if you find this demo useful!</strong>
+        """,
+        inputs=inputs,
+        outputs=outputs,
+        examples=[
+            [f"{REPO_ROOT}/demo/drone.mp4"],
+            [f"{REPO_ROOT}/demo/robot_navi.mp4"]
+        ]
+    )
+
+    demo.queue(default_concurrency_limit=1)
+    demo.launch(
+        # server_name="0.0.0.0",
+        # server_port=1324,
+    )

ckpt/model_config.yaml

@@ -0,0 +1,7 @@
+model_id_with_origin_paths: Wan-AI/Wan2.1-T2V-1.3B:diffusion_pytorch_model*.safetensors,Wan-AI/Wan2.1-T2V-1.3B:Wan2.1_VAE.pth
+trainable_models: dit
+mode: regression
+denoise_step: 0.5
+training_target: x
+lora_base_model: dit
+lora_rank: 512

configs/img_config/data_diode_all.yaml

@@ -0,0 +1,8 @@
+name: diode
+disp_name: diode_val_all
+# dir: diode
+# dataset_dir: diode
+# filename_ls_path: data_split/diode/diode_val_all_filename_list.txt 
+processing_res: 640
+dir: diode
+filename: data_split/diode/diode_val_all_filename_list.txt 

configs/img_config/data_eth3d.yaml

@@ -0,0 +1,11 @@
+name: eth3d
+disp_name: eth3d_full
+# dataset_dir: eth3d
+# dir: eth3d/eth3d.tar
+# filename_ls_path: data_split/eth3d/eth3d_filename_list.txt
+dir: eth3d
+# dir: eth3d/eth3d.tar
+filename: data_split/eth3d/eth3d_filename_list.txt
+resize_to_hw: [480, 720]
+# processing_res: 768
+# alignment_max_res: 1024

configs/img_config/data_kitti_eigen_test.yaml

@@ -0,0 +1,8 @@
+name: kitti
+disp_name: kitti_eigen_test_full
+# dataset_dir: kitti
+# filename_ls_path: data_split/kitti/eigen_test_files_with_gt.txt
+kitti_bm_crop: true
+valid_mask_crop: eigen
+dir: kitti
+filename: data_split/kitti/eigen_test_files_with_gt.txt

configs/img_config/data_nyu_test.yaml

@@ -0,0 +1,8 @@
+name: nyu_v2
+disp_name: nyu_test_full
+# dataset_dir: nyuv2
+# filename_ls_path: data_split/nyu/labeled/filename_list_test.txt
+eigen_valid_mask: true
+
+dir: nyuv2
+filename: data_split/nyu/labeled/filename_list_test.txt

configs/img_config/data_scannet_val.yaml

@@ -0,0 +1,6 @@
+name: scannet
+disp_name: scannet_val_800_1
+# dataset_dir: scannet
+# filename_ls_path: data_split/scannet/scannet_val_sampled_list_800_1.txt
+dir: scannet
+filename: data_split/scannet/scannet_val_sampled_list_800_1.txt

configs/scannetv1_test.txt

@@ -0,0 +1,312 @@
+scene0568_00
+scene0568_01
+scene0568_02
+scene0304_00
+scene0488_00
+scene0488_01
+scene0412_00
+scene0412_01
+scene0217_00
+scene0019_00
+scene0019_01
+scene0414_00
+scene0575_00
+scene0575_01
+scene0575_02
+scene0426_00
+scene0426_01
+scene0426_02
+scene0426_03
+scene0549_00
+scene0549_01
+scene0578_00
+scene0578_01
+scene0578_02
+scene0665_00
+scene0665_01
+scene0050_00
+scene0050_01
+scene0050_02
+scene0257_00
+scene0025_00
+scene0025_01
+scene0025_02
+scene0583_00
+scene0583_01
+scene0583_02
+scene0701_00
+scene0701_01
+scene0701_02
+scene0580_00
+scene0580_01
+scene0565_00
+scene0169_00
+scene0169_01
+scene0655_00
+scene0655_01
+scene0655_02
+scene0063_00
+scene0221_00
+scene0221_01
+scene0591_00
+scene0591_01
+scene0591_02
+scene0678_00
+scene0678_01
+scene0678_02
+scene0462_00
+scene0427_00
+scene0595_00
+scene0193_00
+scene0193_01
+scene0164_00
+scene0164_01
+scene0164_02
+scene0164_03
+scene0598_00
+scene0598_01
+scene0598_02
+scene0599_00
+scene0599_01
+scene0599_02
+scene0328_00
+scene0300_00
+scene0300_01
+scene0354_00
+scene0458_00
+scene0458_01
+scene0423_00
+scene0423_01
+scene0423_02
+scene0307_00
+scene0307_01
+scene0307_02
+scene0606_00
+scene0606_01
+scene0606_02
+scene0432_00
+scene0432_01
+scene0608_00
+scene0608_01
+scene0608_02
+scene0651_00
+scene0651_01
+scene0651_02
+scene0430_00
+scene0430_01
+scene0689_00
+scene0357_00
+scene0357_01
+scene0574_00
+scene0574_01
+scene0574_02
+scene0329_00
+scene0329_01
+scene0329_02
+scene0153_00
+scene0153_01
+scene0616_00
+scene0616_01
+scene0671_00
+scene0671_01
+scene0618_00
+scene0382_00
+scene0382_01
+scene0490_00
+scene0621_00
+scene0607_00
+scene0607_01
+scene0149_00
+scene0695_00
+scene0695_01
+scene0695_02
+scene0695_03
+scene0389_00
+scene0377_00
+scene0377_01
+scene0377_02
+scene0342_00
+scene0139_00
+scene0629_00
+scene0629_01
+scene0629_02
+scene0496_00
+scene0633_00
+scene0633_01
+scene0518_00
+scene0652_00
+scene0406_00
+scene0406_01
+scene0406_02
+scene0144_00
+scene0144_01
+scene0494_00
+scene0278_00
+scene0278_01
+scene0316_00
+scene0609_00
+scene0609_01
+scene0609_02
+scene0609_03
+scene0084_00
+scene0084_01
+scene0084_02
+scene0696_00
+scene0696_01
+scene0696_02
+scene0351_00
+scene0351_01
+scene0643_00
+scene0644_00
+scene0645_00
+scene0645_01
+scene0645_02
+scene0081_00
+scene0081_01
+scene0081_02
+scene0647_00
+scene0647_01
+scene0535_00
+scene0353_00
+scene0353_01
+scene0353_02
+scene0559_00
+scene0559_01
+scene0559_02
+scene0593_00
+scene0593_01
+scene0246_00
+scene0653_00
+scene0653_01
+scene0064_00
+scene0064_01
+scene0356_00
+scene0356_01
+scene0356_02
+scene0030_00
+scene0030_01
+scene0030_02
+scene0222_00
+scene0222_01
+scene0338_00
+scene0338_01
+scene0338_02
+scene0378_00
+scene0378_01
+scene0378_02
+scene0660_00
+scene0553_00
+scene0553_01
+scene0553_02
+scene0527_00
+scene0663_00
+scene0663_01
+scene0663_02
+scene0664_00
+scene0664_01
+scene0664_02
+scene0334_00
+scene0334_01
+scene0334_02
+scene0046_00
+scene0046_01
+scene0046_02
+scene0203_00
+scene0203_01
+scene0203_02
+scene0088_00
+scene0088_01
+scene0088_02
+scene0088_03
+scene0086_00
+scene0086_01
+scene0086_02
+scene0670_00
+scene0670_01
+scene0256_00
+scene0256_01
+scene0256_02
+scene0249_00
+scene0441_00
+scene0658_00
+scene0704_00
+scene0704_01
+scene0187_00
+scene0187_01
+scene0131_00
+scene0131_01
+scene0131_02
+scene0207_00
+scene0207_01
+scene0207_02
+scene0461_00
+scene0011_00
+scene0011_01
+scene0343_00
+scene0251_00
+scene0077_00
+scene0077_01
+scene0684_00
+scene0684_01
+scene0550_00
+scene0686_00
+scene0686_01
+scene0686_02
+scene0208_00
+scene0500_00
+scene0500_01
+scene0552_00
+scene0552_01
+scene0648_00
+scene0648_01
+scene0435_00
+scene0435_01
+scene0435_02
+scene0435_03
+scene0690_00
+scene0690_01
+scene0693_00
+scene0693_01
+scene0693_02
+scene0700_00
+scene0700_01
+scene0700_02
+scene0699_00
+scene0231_00
+scene0231_01
+scene0231_02
+scene0697_00
+scene0697_01
+scene0697_02
+scene0697_03
+scene0474_00
+scene0474_01
+scene0474_02
+scene0474_03
+scene0474_04
+scene0474_05
+scene0355_00
+scene0355_01
+scene0146_00
+scene0146_01
+scene0146_02
+scene0196_00
+scene0702_00
+scene0702_01
+scene0702_02
+scene0314_00
+scene0277_00
+scene0277_01
+scene0277_02
+scene0095_00
+scene0095_01
+scene0015_00
+scene0100_00
+scene0100_01
+scene0100_02
+scene0558_00
+scene0558_01
+scene0558_02
+scene0685_00
+scene0685_01
+scene0685_02

configs/vid_config/img_sintel.yaml

@@ -0,0 +1,3 @@
+name: sintel
+dir: Sintel/training
+stack_scene_depth: false

configs/vid_config/vid_bonn.yaml

@@ -0,0 +1,2 @@
+name: bonn
+dir: rgbd_bonn_dataset

configs/vid_config/vid_kitti.yaml

@@ -0,0 +1,2 @@
+name: kitti
+dir: kitti_depth

configs/vid_config/vid_scannet.yaml

@@ -0,0 +1,3 @@
+name: scannet
+dir: scannet
+split_ls: 'configs/scannetv1_test.txt'

configs/vid_config/vid_sintel.yaml

@@ -0,0 +1,3 @@
+name: sintel
+dir: Sintel/training
+stack_scene_depth: true

diffsynth/__init__.py

@@ -0,0 +1,4 @@
+from .data import *
+from .models import *
+from .pipelines import *
+from .schedulers import *

diffsynth/configs/model_config.py

@@ -0,0 +1,705 @@
+from typing_extensions import Literal, TypeAlias
+
+from ..models.wan_video_dit import WanModel
+from ..models.wan_video_image_encoder import WanImageEncoder
+from ..models.wan_video_motion_controller import WanMotionControllerModel
+from ..models.wan_video_text_encoder import WanTextEncoder
+from ..models.wan_video_vace import VaceWanModel
+from ..models.wan_video_vae import WanVideoVAE
+
+model_loader_configs = [
+    # These configs are provided for detecting model type automatically.
+    # The format is (state_dict_keys_hash, state_dict_keys_hash_with_shape, model_names, model_classes, model_resource)
+     (None, "9269f8db9040a9d860eaca435be61814", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "aafcfd9672c3a2456dc46e1cb6e52c70", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "6bfcfb3b342cb286ce886889d519a77e", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "6d6ccde6845b95ad9114ab993d917893", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "6bfcfb3b342cb286ce886889d519a77e", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "349723183fc063b2bfc10bb2835cf677", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "efa44cddf936c70abd0ea28b6cbe946c", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "3ef3b1f8e1dab83d5b71fd7b617f859f", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "70ddad9d3a133785da5ea371aae09504", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "26bde73488a92e64cc20b0a7485b9e5b", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "ac6a5aa74f4a0aab6f64eb9a72f19901", ["wan_video_dit"], [WanModel], "civitai"), 
+    (None, "b61c605c2adbd23124d152ed28e049ae", ["wan_video_dit"], [WanModel], "civitai"), 
+    (None, "a61453409b67cd3246cf0c3bebad47ba", ["wan_video_dit", "wan_video_vace"], [WanModel, VaceWanModel], "civitai"),
+    (None, "7a513e1f257a861512b1afd387a8ecd9", ["wan_video_dit", "wan_video_vace"], [WanModel, VaceWanModel], "civitai"),
+    (None, "cb104773c6c2cb6df4f9529ad5c60d0b", ["wan_video_dit"], [WanModel], "diffusers"),
+    (None, "9c8818c2cbea55eca56c7b447df170da", ["wan_video_text_encoder"], [WanTextEncoder], "civitai"),
+    (None, "5941c53e207d62f20f9025686193c40b", ["wan_video_image_encoder"], [WanImageEncoder], "civitai"),
+    (None, "1378ea763357eea97acdef78e65d6d96", ["wan_video_vae"], [WanVideoVAE], "civitai"),
+    (None, "ccc42284ea13e1ad04693284c7a09be6", ["wan_video_vae"], [WanVideoVAE], "civitai"),
+    (None, "dbd5ec76bbf977983f972c151d545389", ["wan_video_motion_controller"], [WanMotionControllerModel], "civitai"),
+]
+huggingface_model_loader_configs = [
+    # These configs are provided for detecting model type automatically.
+    # The format is (architecture_in_huggingface_config, huggingface_lib, model_name, redirected_architecture)
+    ("ChatGLMModel", "diffsynth.models.kolors_text_encoder", "kolors_text_encoder", None),
+    ("MarianMTModel", "transformers.models.marian.modeling_marian", "translator", None),
+    ("BloomForCausalLM", "transformers.models.bloom.modeling_bloom", "beautiful_prompt", None),
+    ("Qwen2ForCausalLM", "transformers.models.qwen2.modeling_qwen2", "qwen_prompt", None),
+    ("LlamaForCausalLM", "transformers.models.llama.modeling_llama", "omost_prompt", None),
+    ("T5EncoderModel", "diffsynth.models.flux_text_encoder", "flux_text_encoder_2", "FluxTextEncoder2"),
+    ("CogVideoXTransformer3DModel", "diffsynth.models.cog_dit", "cog_dit", "CogDiT"),
+    ("SiglipModel", "transformers.models.siglip.modeling_siglip", "siglip_vision_model", "SiglipVisionModel"),
+    ("LlamaForCausalLM", "diffsynth.models.hunyuan_video_text_encoder", "hunyuan_video_text_encoder_2", "HunyuanVideoLLMEncoder"),
+    ("LlavaForConditionalGeneration", "diffsynth.models.hunyuan_video_text_encoder", "hunyuan_video_text_encoder_2", "HunyuanVideoMLLMEncoder"),
+    ("Step1Model", "diffsynth.models.stepvideo_text_encoder", "stepvideo_text_encoder_2", "STEP1TextEncoder"),
+    ("Qwen2_5_VLForConditionalGeneration", "diffsynth.models.qwenvl", "qwenvl", "Qwen25VL_7b_Embedder"),
+]
+patch_model_loader_configs = [
+    # These configs are provided for detecting model type automatically.
+    # The format is (state_dict_keys_hash_with_shape, model_name, model_class, extra_kwargs)
+]
+
+preset_models_on_huggingface = {
+    "HunyuanDiT": [
+        ("Tencent-Hunyuan/HunyuanDiT", "t2i/clip_text_encoder/pytorch_model.bin", "models/HunyuanDiT/t2i/clip_text_encoder"),
+        ("Tencent-Hunyuan/HunyuanDiT", "t2i/mt5/pytorch_model.bin", "models/HunyuanDiT/t2i/mt5"),
+        ("Tencent-Hunyuan/HunyuanDiT", "t2i/model/pytorch_model_ema.pt", "models/HunyuanDiT/t2i/model"),
+        ("Tencent-Hunyuan/HunyuanDiT", "t2i/sdxl-vae-fp16-fix/diffusion_pytorch_model.bin", "models/HunyuanDiT/t2i/sdxl-vae-fp16-fix"),
+    ],
+    "stable-video-diffusion-img2vid-xt": [
+        ("stabilityai/stable-video-diffusion-img2vid-xt", "svd_xt.safetensors", "models/stable_video_diffusion"),
+    ],
+    "ExVideo-SVD-128f-v1": [
+        ("ECNU-CILab/ExVideo-SVD-128f-v1", "model.fp16.safetensors", "models/stable_video_diffusion"),
+    ],
+    # Stable Diffusion
+    "StableDiffusion_v15": [
+        ("benjamin-paine/stable-diffusion-v1-5", "v1-5-pruned-emaonly.safetensors", "models/stable_diffusion"),
+    ],
+    "DreamShaper_8": [
+        ("Yntec/Dreamshaper8", "dreamshaper_8.safetensors", "models/stable_diffusion"),
+    ],
+    # Textual Inversion
+    "TextualInversion_VeryBadImageNegative_v1.3": [
+        ("gemasai/verybadimagenegative_v1.3", "verybadimagenegative_v1.3.pt", "models/textual_inversion"),
+    ],
+    # Stable Diffusion XL
+    "StableDiffusionXL_v1": [
+        ("stabilityai/stable-diffusion-xl-base-1.0", "sd_xl_base_1.0.safetensors", "models/stable_diffusion_xl"),
+    ],
+    "BluePencilXL_v200": [
+        ("frankjoshua/bluePencilXL_v200", "bluePencilXL_v200.safetensors", "models/stable_diffusion_xl"),
+    ],
+    "StableDiffusionXL_Turbo": [
+        ("stabilityai/sdxl-turbo", "sd_xl_turbo_1.0_fp16.safetensors", "models/stable_diffusion_xl_turbo"),
+    ],
+    # Stable Diffusion 3
+    "StableDiffusion3": [
+        ("stabilityai/stable-diffusion-3-medium", "sd3_medium_incl_clips_t5xxlfp16.safetensors", "models/stable_diffusion_3"),
+    ],
+    "StableDiffusion3_without_T5": [
+        ("stabilityai/stable-diffusion-3-medium", "sd3_medium_incl_clips.safetensors", "models/stable_diffusion_3"),
+    ],
+    # ControlNet
+    "ControlNet_v11f1p_sd15_depth": [
+        ("lllyasviel/ControlNet-v1-1", "control_v11f1p_sd15_depth.pth", "models/ControlNet"),
+        ("lllyasviel/Annotators", "dpt_hybrid-midas-501f0c75.pt", "models/Annotators")
+    ],
+    "ControlNet_v11p_sd15_softedge": [
+        ("lllyasviel/ControlNet-v1-1", "control_v11p_sd15_softedge.pth", "models/ControlNet"),
+        ("lllyasviel/Annotators", "ControlNetHED.pth", "models/Annotators")
+    ],
+    "ControlNet_v11f1e_sd15_tile": [
+        ("lllyasviel/ControlNet-v1-1", "control_v11f1e_sd15_tile.pth", "models/ControlNet")
+    ],
+    "ControlNet_v11p_sd15_lineart": [
+        ("lllyasviel/ControlNet-v1-1", "control_v11p_sd15_lineart.pth", "models/ControlNet"),
+        ("lllyasviel/Annotators", "sk_model.pth", "models/Annotators"),
+        ("lllyasviel/Annotators", "sk_model2.pth", "models/Annotators")
+    ],
+    "ControlNet_union_sdxl_promax": [
+        ("xinsir/controlnet-union-sdxl-1.0", "diffusion_pytorch_model_promax.safetensors", "models/ControlNet/controlnet_union"),
+        ("lllyasviel/Annotators", "dpt_hybrid-midas-501f0c75.pt", "models/Annotators")
+    ],
+    # AnimateDiff
+    "AnimateDiff_v2": [
+        ("guoyww/animatediff", "mm_sd_v15_v2.ckpt", "models/AnimateDiff"),
+    ],
+    "AnimateDiff_xl_beta": [
+        ("guoyww/animatediff", "mm_sdxl_v10_beta.ckpt", "models/AnimateDiff"),
+    ],
+
+    # Qwen Prompt
+    "QwenPrompt": [
+        ("Qwen/Qwen2-1.5B-Instruct", "config.json", "models/QwenPrompt/qwen2-1.5b-instruct"),
+        ("Qwen/Qwen2-1.5B-Instruct", "generation_config.json", "models/QwenPrompt/qwen2-1.5b-instruct"),
+        ("Qwen/Qwen2-1.5B-Instruct", "model.safetensors", "models/QwenPrompt/qwen2-1.5b-instruct"),
+        ("Qwen/Qwen2-1.5B-Instruct", "special_tokens_map.json", "models/QwenPrompt/qwen2-1.5b-instruct"),
+        ("Qwen/Qwen2-1.5B-Instruct", "tokenizer.json", "models/QwenPrompt/qwen2-1.5b-instruct"),
+        ("Qwen/Qwen2-1.5B-Instruct", "tokenizer_config.json", "models/QwenPrompt/qwen2-1.5b-instruct"),
+        ("Qwen/Qwen2-1.5B-Instruct", "merges.txt", "models/QwenPrompt/qwen2-1.5b-instruct"),
+        ("Qwen/Qwen2-1.5B-Instruct", "vocab.json", "models/QwenPrompt/qwen2-1.5b-instruct"),
+    ],
+    # Beautiful Prompt
+    "BeautifulPrompt": [
+        ("alibaba-pai/pai-bloom-1b1-text2prompt-sd", "config.json", "models/BeautifulPrompt/pai-bloom-1b1-text2prompt-sd"),
+        ("alibaba-pai/pai-bloom-1b1-text2prompt-sd", "generation_config.json", "models/BeautifulPrompt/pai-bloom-1b1-text2prompt-sd"),
+        ("alibaba-pai/pai-bloom-1b1-text2prompt-sd", "model.safetensors", "models/BeautifulPrompt/pai-bloom-1b1-text2prompt-sd"),
+        ("alibaba-pai/pai-bloom-1b1-text2prompt-sd", "special_tokens_map.json", "models/BeautifulPrompt/pai-bloom-1b1-text2prompt-sd"),
+        ("alibaba-pai/pai-bloom-1b1-text2prompt-sd", "tokenizer.json", "models/BeautifulPrompt/pai-bloom-1b1-text2prompt-sd"),
+        ("alibaba-pai/pai-bloom-1b1-text2prompt-sd", "tokenizer_config.json", "models/BeautifulPrompt/pai-bloom-1b1-text2prompt-sd"),
+    ],
+    # Omost prompt
+    "OmostPrompt":[
+        ("lllyasviel/omost-llama-3-8b-4bits", "model-00001-of-00002.safetensors", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+        ("lllyasviel/omost-llama-3-8b-4bits", "model-00002-of-00002.safetensors", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+        ("lllyasviel/omost-llama-3-8b-4bits", "tokenizer.json", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+        ("lllyasviel/omost-llama-3-8b-4bits", "tokenizer_config.json", "models/OmostPrompt/omost-llama-3-8b-4bits"),  
+        ("lllyasviel/omost-llama-3-8b-4bits", "config.json", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+        ("lllyasviel/omost-llama-3-8b-4bits", "generation_config.json", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+        ("lllyasviel/omost-llama-3-8b-4bits", "model.safetensors.index.json", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+        ("lllyasviel/omost-llama-3-8b-4bits", "special_tokens_map.json", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+    ],
+    # Translator
+    "opus-mt-zh-en": [
+        ("Helsinki-NLP/opus-mt-zh-en", "config.json", "models/translator/opus-mt-zh-en"),
+        ("Helsinki-NLP/opus-mt-zh-en", "generation_config.json", "models/translator/opus-mt-zh-en"),
+        ("Helsinki-NLP/opus-mt-zh-en", "metadata.json", "models/translator/opus-mt-zh-en"),
+        ("Helsinki-NLP/opus-mt-zh-en", "pytorch_model.bin", "models/translator/opus-mt-zh-en"),
+        ("Helsinki-NLP/opus-mt-zh-en", "source.spm", "models/translator/opus-mt-zh-en"),
+        ("Helsinki-NLP/opus-mt-zh-en", "target.spm", "models/translator/opus-mt-zh-en"),
+        ("Helsinki-NLP/opus-mt-zh-en", "tokenizer_config.json", "models/translator/opus-mt-zh-en"),
+        ("Helsinki-NLP/opus-mt-zh-en", "vocab.json", "models/translator/opus-mt-zh-en"),
+    ],
+    # IP-Adapter
+    "IP-Adapter-SD": [
+        ("h94/IP-Adapter", "models/image_encoder/model.safetensors", "models/IpAdapter/stable_diffusion/image_encoder"),
+        ("h94/IP-Adapter", "models/ip-adapter_sd15.bin", "models/IpAdapter/stable_diffusion"),
+    ],
+    "IP-Adapter-SDXL": [
+        ("h94/IP-Adapter", "sdxl_models/image_encoder/model.safetensors", "models/IpAdapter/stable_diffusion_xl/image_encoder"),
+        ("h94/IP-Adapter", "sdxl_models/ip-adapter_sdxl.bin", "models/IpAdapter/stable_diffusion_xl"),
+    ],
+    "SDXL-vae-fp16-fix": [
+        ("madebyollin/sdxl-vae-fp16-fix", "diffusion_pytorch_model.safetensors", "models/sdxl-vae-fp16-fix")
+    ],
+    # Kolors
+    "Kolors": [
+        ("Kwai-Kolors/Kolors", "text_encoder/config.json", "models/kolors/Kolors/text_encoder"),
+        ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model.bin.index.json", "models/kolors/Kolors/text_encoder"),
+        ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00001-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+        ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00002-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+        ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00003-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+        ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00004-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+        ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00005-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+        ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00006-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+        ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00007-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+        ("Kwai-Kolors/Kolors", "unet/diffusion_pytorch_model.safetensors", "models/kolors/Kolors/unet"),
+        ("Kwai-Kolors/Kolors", "vae/diffusion_pytorch_model.safetensors", "models/kolors/Kolors/vae"),
+    ],
+    # FLUX
+    "FLUX.1-dev": [
+        ("black-forest-labs/FLUX.1-dev", "text_encoder/model.safetensors", "models/FLUX/FLUX.1-dev/text_encoder"),
+        ("black-forest-labs/FLUX.1-dev", "text_encoder_2/config.json", "models/FLUX/FLUX.1-dev/text_encoder_2"),
+        ("black-forest-labs/FLUX.1-dev", "text_encoder_2/model-00001-of-00002.safetensors", "models/FLUX/FLUX.1-dev/text_encoder_2"),
+        ("black-forest-labs/FLUX.1-dev", "text_encoder_2/model-00002-of-00002.safetensors", "models/FLUX/FLUX.1-dev/text_encoder_2"),
+        ("black-forest-labs/FLUX.1-dev", "text_encoder_2/model.safetensors.index.json", "models/FLUX/FLUX.1-dev/text_encoder_2"),
+        ("black-forest-labs/FLUX.1-dev", "ae.safetensors", "models/FLUX/FLUX.1-dev"),
+        ("black-forest-labs/FLUX.1-dev", "flux1-dev.safetensors", "models/FLUX/FLUX.1-dev"),
+    ],
+    "InstantX/FLUX.1-dev-IP-Adapter": {
+        "file_list": [
+            ("InstantX/FLUX.1-dev-IP-Adapter", "ip-adapter.bin", "models/IpAdapter/InstantX/FLUX.1-dev-IP-Adapter"),
+            ("google/siglip-so400m-patch14-384", "model.safetensors", "models/IpAdapter/InstantX/FLUX.1-dev-IP-Adapter/image_encoder"),
+            ("google/siglip-so400m-patch14-384", "config.json", "models/IpAdapter/InstantX/FLUX.1-dev-IP-Adapter/image_encoder"),
+        ],
+        "load_path": [
+            "models/IpAdapter/InstantX/FLUX.1-dev-IP-Adapter/ip-adapter.bin",
+            "models/IpAdapter/InstantX/FLUX.1-dev-IP-Adapter/image_encoder",
+        ],
+    },
+    # RIFE
+    "RIFE": [
+        ("AlexWortega/RIFE", "flownet.pkl", "models/RIFE"),
+    ],
+    # CogVideo
+    "CogVideoX-5B": [
+        ("THUDM/CogVideoX-5b", "text_encoder/config.json", "models/CogVideo/CogVideoX-5b/text_encoder"),
+        ("THUDM/CogVideoX-5b", "text_encoder/model.safetensors.index.json", "models/CogVideo/CogVideoX-5b/text_encoder"),
+        ("THUDM/CogVideoX-5b", "text_encoder/model-00001-of-00002.safetensors", "models/CogVideo/CogVideoX-5b/text_encoder"),
+        ("THUDM/CogVideoX-5b", "text_encoder/model-00002-of-00002.safetensors", "models/CogVideo/CogVideoX-5b/text_encoder"),
+        ("THUDM/CogVideoX-5b", "transformer/config.json", "models/CogVideo/CogVideoX-5b/transformer"),
+        ("THUDM/CogVideoX-5b", "transformer/diffusion_pytorch_model.safetensors.index.json", "models/CogVideo/CogVideoX-5b/transformer"),
+        ("THUDM/CogVideoX-5b", "transformer/diffusion_pytorch_model-00001-of-00002.safetensors", "models/CogVideo/CogVideoX-5b/transformer"),
+        ("THUDM/CogVideoX-5b", "transformer/diffusion_pytorch_model-00002-of-00002.safetensors", "models/CogVideo/CogVideoX-5b/transformer"),
+        ("THUDM/CogVideoX-5b", "vae/diffusion_pytorch_model.safetensors", "models/CogVideo/CogVideoX-5b/vae"),
+    ],
+    # Stable Diffusion 3.5
+    "StableDiffusion3.5-large": [
+        ("stabilityai/stable-diffusion-3.5-large", "sd3.5_large.safetensors", "models/stable_diffusion_3"),
+        ("stabilityai/stable-diffusion-3.5-large", "text_encoders/clip_l.safetensors", "models/stable_diffusion_3/text_encoders"),
+        ("stabilityai/stable-diffusion-3.5-large", "text_encoders/clip_g.safetensors", "models/stable_diffusion_3/text_encoders"),
+        ("stabilityai/stable-diffusion-3.5-large", "text_encoders/t5xxl_fp16.safetensors", "models/stable_diffusion_3/text_encoders"),
+    ],
+}
+preset_models_on_modelscope = {
+    # Hunyuan DiT
+    "HunyuanDiT": [
+        ("modelscope/HunyuanDiT", "t2i/clip_text_encoder/pytorch_model.bin", "models/HunyuanDiT/t2i/clip_text_encoder"),
+        ("modelscope/HunyuanDiT", "t2i/mt5/pytorch_model.bin", "models/HunyuanDiT/t2i/mt5"),
+        ("modelscope/HunyuanDiT", "t2i/model/pytorch_model_ema.pt", "models/HunyuanDiT/t2i/model"),
+        ("modelscope/HunyuanDiT", "t2i/sdxl-vae-fp16-fix/diffusion_pytorch_model.bin", "models/HunyuanDiT/t2i/sdxl-vae-fp16-fix"),
+    ],
+    # Stable Video Diffusion
+    "stable-video-diffusion-img2vid-xt": [
+        ("AI-ModelScope/stable-video-diffusion-img2vid-xt", "svd_xt.safetensors", "models/stable_video_diffusion"),
+    ],
+    # ExVideo
+    "ExVideo-SVD-128f-v1": [
+        ("ECNU-CILab/ExVideo-SVD-128f-v1", "model.fp16.safetensors", "models/stable_video_diffusion"),
+    ],
+    "ExVideo-CogVideoX-LoRA-129f-v1": [
+        ("ECNU-CILab/ExVideo-CogVideoX-LoRA-129f-v1", "ExVideo-CogVideoX-LoRA-129f-v1.safetensors", "models/lora"),
+    ],
+    # Stable Diffusion
+    "StableDiffusion_v15": [
+        ("AI-ModelScope/stable-diffusion-v1-5", "v1-5-pruned-emaonly.safetensors", "models/stable_diffusion"),
+    ],
+    "DreamShaper_8": [
+        ("sd_lora/dreamshaper_8", "dreamshaper_8.safetensors", "models/stable_diffusion"),
+    ],
+    "AingDiffusion_v12": [
+        ("sd_lora/aingdiffusion_v12", "aingdiffusion_v12.safetensors", "models/stable_diffusion"),
+    ],
+    "Flat2DAnimerge_v45Sharp": [
+        ("sd_lora/Flat-2D-Animerge", "flat2DAnimerge_v45Sharp.safetensors", "models/stable_diffusion"),
+    ],
+    # Textual Inversion
+    "TextualInversion_VeryBadImageNegative_v1.3": [
+        ("sd_lora/verybadimagenegative_v1.3", "verybadimagenegative_v1.3.pt", "models/textual_inversion"),
+    ],
+    # Stable Diffusion XL
+    "StableDiffusionXL_v1": [
+        ("AI-ModelScope/stable-diffusion-xl-base-1.0", "sd_xl_base_1.0.safetensors", "models/stable_diffusion_xl"),
+    ],
+    "BluePencilXL_v200": [
+        ("sd_lora/bluePencilXL_v200", "bluePencilXL_v200.safetensors", "models/stable_diffusion_xl"),
+    ],
+    "StableDiffusionXL_Turbo": [
+        ("AI-ModelScope/sdxl-turbo", "sd_xl_turbo_1.0_fp16.safetensors", "models/stable_diffusion_xl_turbo"),
+    ],
+    "SDXL_lora_zyd232_ChineseInkStyle_SDXL_v1_0": [
+        ("sd_lora/zyd232_ChineseInkStyle_SDXL_v1_0", "zyd232_ChineseInkStyle_SDXL_v1_0.safetensors", "models/lora"),
+    ],
+    # Stable Diffusion 3
+    "StableDiffusion3": [
+        ("AI-ModelScope/stable-diffusion-3-medium", "sd3_medium_incl_clips_t5xxlfp16.safetensors", "models/stable_diffusion_3"),
+    ],
+    "StableDiffusion3_without_T5": [
+        ("AI-ModelScope/stable-diffusion-3-medium", "sd3_medium_incl_clips.safetensors", "models/stable_diffusion_3"),
+    ],
+    # ControlNet
+    "ControlNet_v11f1p_sd15_depth": [
+        ("AI-ModelScope/ControlNet-v1-1", "control_v11f1p_sd15_depth.pth", "models/ControlNet"),
+        ("sd_lora/Annotators", "dpt_hybrid-midas-501f0c75.pt", "models/Annotators")
+    ],
+    "ControlNet_v11p_sd15_softedge": [
+        ("AI-ModelScope/ControlNet-v1-1", "control_v11p_sd15_softedge.pth", "models/ControlNet"),
+        ("sd_lora/Annotators", "ControlNetHED.pth", "models/Annotators")
+    ],
+    "ControlNet_v11f1e_sd15_tile": [
+        ("AI-ModelScope/ControlNet-v1-1", "control_v11f1e_sd15_tile.pth", "models/ControlNet")
+    ],
+    "ControlNet_v11p_sd15_lineart": [
+        ("AI-ModelScope/ControlNet-v1-1", "control_v11p_sd15_lineart.pth", "models/ControlNet"),
+        ("sd_lora/Annotators", "sk_model.pth", "models/Annotators"),
+        ("sd_lora/Annotators", "sk_model2.pth", "models/Annotators")
+    ],
+    "ControlNet_union_sdxl_promax": [
+        ("AI-ModelScope/controlnet-union-sdxl-1.0", "diffusion_pytorch_model_promax.safetensors", "models/ControlNet/controlnet_union"),
+        ("sd_lora/Annotators", "dpt_hybrid-midas-501f0c75.pt", "models/Annotators")
+    ],
+    "Annotators:Depth": [
+        ("sd_lora/Annotators", "dpt_hybrid-midas-501f0c75.pt", "models/Annotators"),
+    ],
+    "Annotators:Softedge": [
+        ("sd_lora/Annotators", "ControlNetHED.pth", "models/Annotators"),
+    ],
+    "Annotators:Lineart": [
+        ("sd_lora/Annotators", "sk_model.pth", "models/Annotators"),
+        ("sd_lora/Annotators", "sk_model2.pth", "models/Annotators"),
+    ],
+    "Annotators:Normal": [
+        ("sd_lora/Annotators", "scannet.pt", "models/Annotators"),
+    ],
+    "Annotators:Openpose": [
+        ("sd_lora/Annotators", "body_pose_model.pth", "models/Annotators"),
+        ("sd_lora/Annotators", "facenet.pth", "models/Annotators"),
+        ("sd_lora/Annotators", "hand_pose_model.pth", "models/Annotators"),
+    ],
+    # AnimateDiff
+    "AnimateDiff_v2": [
+        ("Shanghai_AI_Laboratory/animatediff", "mm_sd_v15_v2.ckpt", "models/AnimateDiff"),
+    ],
+    "AnimateDiff_xl_beta": [
+        ("Shanghai_AI_Laboratory/animatediff", "mm_sdxl_v10_beta.ckpt", "models/AnimateDiff"),
+    ],
+    # RIFE
+    "RIFE": [
+        ("Damo_XR_Lab/cv_rife_video-frame-interpolation", "flownet.pkl", "models/RIFE"),
+    ],
+    # Qwen Prompt
+    "QwenPrompt": {
+        "file_list": [
+            ("qwen/Qwen2-1.5B-Instruct", "config.json", "models/QwenPrompt/qwen2-1.5b-instruct"),
+            ("qwen/Qwen2-1.5B-Instruct", "generation_config.json", "models/QwenPrompt/qwen2-1.5b-instruct"),
+            ("qwen/Qwen2-1.5B-Instruct", "model.safetensors", "models/QwenPrompt/qwen2-1.5b-instruct"),
+            ("qwen/Qwen2-1.5B-Instruct", "special_tokens_map.json", "models/QwenPrompt/qwen2-1.5b-instruct"),
+            ("qwen/Qwen2-1.5B-Instruct", "tokenizer.json", "models/QwenPrompt/qwen2-1.5b-instruct"),
+            ("qwen/Qwen2-1.5B-Instruct", "tokenizer_config.json", "models/QwenPrompt/qwen2-1.5b-instruct"),
+            ("qwen/Qwen2-1.5B-Instruct", "merges.txt", "models/QwenPrompt/qwen2-1.5b-instruct"),
+            ("qwen/Qwen2-1.5B-Instruct", "vocab.json", "models/QwenPrompt/qwen2-1.5b-instruct"),
+        ],
+        "load_path": [
+            "models/QwenPrompt/qwen2-1.5b-instruct",
+        ],
+    },
+    # Beautiful Prompt
+    "BeautifulPrompt": {
+        "file_list": [
+            ("AI-ModelScope/pai-bloom-1b1-text2prompt-sd", "config.json", "models/BeautifulPrompt/pai-bloom-1b1-text2prompt-sd"),
+            ("AI-ModelScope/pai-bloom-1b1-text2prompt-sd", "generation_config.json", "models/BeautifulPrompt/pai-bloom-1b1-text2prompt-sd"),
+            ("AI-ModelScope/pai-bloom-1b1-text2prompt-sd", "model.safetensors", "models/BeautifulPrompt/pai-bloom-1b1-text2prompt-sd"),
+            ("AI-ModelScope/pai-bloom-1b1-text2prompt-sd", "special_tokens_map.json", "models/BeautifulPrompt/pai-bloom-1b1-text2prompt-sd"),
+            ("AI-ModelScope/pai-bloom-1b1-text2prompt-sd", "tokenizer.json", "models/BeautifulPrompt/pai-bloom-1b1-text2prompt-sd"),
+            ("AI-ModelScope/pai-bloom-1b1-text2prompt-sd", "tokenizer_config.json", "models/BeautifulPrompt/pai-bloom-1b1-text2prompt-sd"),
+        ],
+        "load_path": [
+            "models/BeautifulPrompt/pai-bloom-1b1-text2prompt-sd",
+        ],
+    },
+    # Omost prompt
+    "OmostPrompt": {
+        "file_list": [
+            ("Omost/omost-llama-3-8b-4bits", "model-00001-of-00002.safetensors", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+            ("Omost/omost-llama-3-8b-4bits", "model-00002-of-00002.safetensors", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+            ("Omost/omost-llama-3-8b-4bits", "tokenizer.json", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+            ("Omost/omost-llama-3-8b-4bits", "tokenizer_config.json", "models/OmostPrompt/omost-llama-3-8b-4bits"),  
+            ("Omost/omost-llama-3-8b-4bits", "config.json", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+            ("Omost/omost-llama-3-8b-4bits", "generation_config.json", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+            ("Omost/omost-llama-3-8b-4bits", "model.safetensors.index.json", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+            ("Omost/omost-llama-3-8b-4bits", "special_tokens_map.json", "models/OmostPrompt/omost-llama-3-8b-4bits"),
+        ],
+        "load_path": [
+            "models/OmostPrompt/omost-llama-3-8b-4bits",
+        ],
+    },
+    # Translator
+    "opus-mt-zh-en": {
+        "file_list": [
+            ("moxying/opus-mt-zh-en", "config.json", "models/translator/opus-mt-zh-en"),
+            ("moxying/opus-mt-zh-en", "generation_config.json", "models/translator/opus-mt-zh-en"),
+            ("moxying/opus-mt-zh-en", "metadata.json", "models/translator/opus-mt-zh-en"),
+            ("moxying/opus-mt-zh-en", "pytorch_model.bin", "models/translator/opus-mt-zh-en"),
+            ("moxying/opus-mt-zh-en", "source.spm", "models/translator/opus-mt-zh-en"),
+            ("moxying/opus-mt-zh-en", "target.spm", "models/translator/opus-mt-zh-en"),
+            ("moxying/opus-mt-zh-en", "tokenizer_config.json", "models/translator/opus-mt-zh-en"),
+            ("moxying/opus-mt-zh-en", "vocab.json", "models/translator/opus-mt-zh-en"),
+        ],
+        "load_path": [
+            "models/translator/opus-mt-zh-en",
+        ],
+    },
+    # IP-Adapter
+    "IP-Adapter-SD": [
+        ("AI-ModelScope/IP-Adapter", "models/image_encoder/model.safetensors", "models/IpAdapter/stable_diffusion/image_encoder"),
+        ("AI-ModelScope/IP-Adapter", "models/ip-adapter_sd15.bin", "models/IpAdapter/stable_diffusion"),
+    ],
+    "IP-Adapter-SDXL": [
+        ("AI-ModelScope/IP-Adapter", "sdxl_models/image_encoder/model.safetensors", "models/IpAdapter/stable_diffusion_xl/image_encoder"),
+        ("AI-ModelScope/IP-Adapter", "sdxl_models/ip-adapter_sdxl.bin", "models/IpAdapter/stable_diffusion_xl"),
+    ],
+    # Kolors
+    "Kolors": {
+        "file_list": [
+            ("Kwai-Kolors/Kolors", "text_encoder/config.json", "models/kolors/Kolors/text_encoder"),
+            ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model.bin.index.json", "models/kolors/Kolors/text_encoder"),
+            ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00001-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+            ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00002-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+            ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00003-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+            ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00004-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+            ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00005-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+            ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00006-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+            ("Kwai-Kolors/Kolors", "text_encoder/pytorch_model-00007-of-00007.bin", "models/kolors/Kolors/text_encoder"),
+            ("Kwai-Kolors/Kolors", "unet/diffusion_pytorch_model.safetensors", "models/kolors/Kolors/unet"),
+            ("Kwai-Kolors/Kolors", "vae/diffusion_pytorch_model.safetensors", "models/kolors/Kolors/vae"),
+        ],
+        "load_path": [
+            "models/kolors/Kolors/text_encoder",
+            "models/kolors/Kolors/unet/diffusion_pytorch_model.safetensors",
+            "models/kolors/Kolors/vae/diffusion_pytorch_model.safetensors",
+        ],
+    },
+    "SDXL-vae-fp16-fix": [
+        ("AI-ModelScope/sdxl-vae-fp16-fix", "diffusion_pytorch_model.safetensors", "models/sdxl-vae-fp16-fix")
+    ],
+    # FLUX
+    "FLUX.1-dev": {
+        "file_list": [
+            ("AI-ModelScope/FLUX.1-dev", "text_encoder/model.safetensors", "models/FLUX/FLUX.1-dev/text_encoder"),
+            ("AI-ModelScope/FLUX.1-dev", "text_encoder_2/config.json", "models/FLUX/FLUX.1-dev/text_encoder_2"),
+            ("AI-ModelScope/FLUX.1-dev", "text_encoder_2/model-00001-of-00002.safetensors", "models/FLUX/FLUX.1-dev/text_encoder_2"),
+            ("AI-ModelScope/FLUX.1-dev", "text_encoder_2/model-00002-of-00002.safetensors", "models/FLUX/FLUX.1-dev/text_encoder_2"),
+            ("AI-ModelScope/FLUX.1-dev", "text_encoder_2/model.safetensors.index.json", "models/FLUX/FLUX.1-dev/text_encoder_2"),
+            ("AI-ModelScope/FLUX.1-dev", "ae.safetensors", "models/FLUX/FLUX.1-dev"),
+            ("AI-ModelScope/FLUX.1-dev", "flux1-dev.safetensors", "models/FLUX/FLUX.1-dev"),
+        ],
+        "load_path": [
+            "models/FLUX/FLUX.1-dev/text_encoder/model.safetensors",
+            "models/FLUX/FLUX.1-dev/text_encoder_2",
+            "models/FLUX/FLUX.1-dev/ae.safetensors",
+            "models/FLUX/FLUX.1-dev/flux1-dev.safetensors"
+        ],
+    },
+    "FLUX.1-schnell": {
+        "file_list": [
+            ("AI-ModelScope/FLUX.1-dev", "text_encoder/model.safetensors", "models/FLUX/FLUX.1-dev/text_encoder"),
+            ("AI-ModelScope/FLUX.1-dev", "text_encoder_2/config.json", "models/FLUX/FLUX.1-dev/text_encoder_2"),
+            ("AI-ModelScope/FLUX.1-dev", "text_encoder_2/model-00001-of-00002.safetensors", "models/FLUX/FLUX.1-dev/text_encoder_2"),
+            ("AI-ModelScope/FLUX.1-dev", "text_encoder_2/model-00002-of-00002.safetensors", "models/FLUX/FLUX.1-dev/text_encoder_2"),
+            ("AI-ModelScope/FLUX.1-dev", "text_encoder_2/model.safetensors.index.json", "models/FLUX/FLUX.1-dev/text_encoder_2"),
+            ("AI-ModelScope/FLUX.1-dev", "ae.safetensors", "models/FLUX/FLUX.1-dev"),
+            ("AI-ModelScope/FLUX.1-schnell", "flux1-schnell.safetensors", "models/FLUX/FLUX.1-schnell"),
+        ],
+        "load_path": [
+            "models/FLUX/FLUX.1-dev/text_encoder/model.safetensors",
+            "models/FLUX/FLUX.1-dev/text_encoder_2",
+            "models/FLUX/FLUX.1-dev/ae.safetensors",
+            "models/FLUX/FLUX.1-schnell/flux1-schnell.safetensors"
+        ],
+    },
+    "InstantX/FLUX.1-dev-Controlnet-Union-alpha": [
+        ("InstantX/FLUX.1-dev-Controlnet-Union-alpha", "diffusion_pytorch_model.safetensors", "models/ControlNet/InstantX/FLUX.1-dev-Controlnet-Union-alpha"),
+    ],
+    "jasperai/Flux.1-dev-Controlnet-Depth": [
+        ("jasperai/Flux.1-dev-Controlnet-Depth", "diffusion_pytorch_model.safetensors", "models/ControlNet/jasperai/Flux.1-dev-Controlnet-Depth"),
+    ],
+    "jasperai/Flux.1-dev-Controlnet-Surface-Normals": [
+        ("jasperai/Flux.1-dev-Controlnet-Surface-Normals", "diffusion_pytorch_model.safetensors", "models/ControlNet/jasperai/Flux.1-dev-Controlnet-Surface-Normals"),
+    ],
+    "jasperai/Flux.1-dev-Controlnet-Upscaler": [
+        ("jasperai/Flux.1-dev-Controlnet-Upscaler", "diffusion_pytorch_model.safetensors", "models/ControlNet/jasperai/Flux.1-dev-Controlnet-Upscaler"),
+    ],
+    "alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Alpha": [
+        ("alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Alpha", "diffusion_pytorch_model.safetensors", "models/ControlNet/alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Alpha"),
+    ],
+    "alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Beta": [
+        ("alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Beta", "diffusion_pytorch_model.safetensors", "models/ControlNet/alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Beta"),
+    ],
+    "Shakker-Labs/FLUX.1-dev-ControlNet-Depth": [
+        ("Shakker-Labs/FLUX.1-dev-ControlNet-Depth", "diffusion_pytorch_model.safetensors", "models/ControlNet/Shakker-Labs/FLUX.1-dev-ControlNet-Depth"),
+    ],
+    "Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro": [
+        ("Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro", "diffusion_pytorch_model.safetensors", "models/ControlNet/Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro"),
+    ],
+    "InstantX/FLUX.1-dev-IP-Adapter": {
+        "file_list": [
+            ("InstantX/FLUX.1-dev-IP-Adapter", "ip-adapter.bin", "models/IpAdapter/InstantX/FLUX.1-dev-IP-Adapter"),
+            ("AI-ModelScope/siglip-so400m-patch14-384", "model.safetensors", "models/IpAdapter/InstantX/FLUX.1-dev-IP-Adapter/image_encoder"),
+            ("AI-ModelScope/siglip-so400m-patch14-384", "config.json", "models/IpAdapter/InstantX/FLUX.1-dev-IP-Adapter/image_encoder"),
+        ],
+        "load_path": [
+            "models/IpAdapter/InstantX/FLUX.1-dev-IP-Adapter/ip-adapter.bin",
+            "models/IpAdapter/InstantX/FLUX.1-dev-IP-Adapter/image_encoder",
+        ],
+    },
+    "InfiniteYou":{
+        "file_list":[
+            ("ByteDance/InfiniteYou", "infu_flux_v1.0/aes_stage2/InfuseNetModel/diffusion_pytorch_model-00001-of-00002.safetensors", "models/InfiniteYou/InfuseNetModel"),
+            ("ByteDance/InfiniteYou", "infu_flux_v1.0/aes_stage2/InfuseNetModel/diffusion_pytorch_model-00002-of-00002.safetensors", "models/InfiniteYou/InfuseNetModel"),
+            ("ByteDance/InfiniteYou", "infu_flux_v1.0/aes_stage2/image_proj_model.bin", "models/InfiniteYou"),
+            ("ByteDance/InfiniteYou", "supports/insightface/models/antelopev2/1k3d68.onnx", "models/InfiniteYou/insightface/models/antelopev2"),
+            ("ByteDance/InfiniteYou", "supports/insightface/models/antelopev2/2d106det.onnx", "models/InfiniteYou/insightface/models/antelopev2"),
+            ("ByteDance/InfiniteYou", "supports/insightface/models/antelopev2/genderage.onnx", "models/InfiniteYou/insightface/models/antelopev2"),
+            ("ByteDance/InfiniteYou", "supports/insightface/models/antelopev2/glintr100.onnx", "models/InfiniteYou/insightface/models/antelopev2"),
+            ("ByteDance/InfiniteYou", "supports/insightface/models/antelopev2/scrfd_10g_bnkps.onnx", "models/InfiniteYou/insightface/models/antelopev2"),            
+        ],
+        "load_path":[
+            [
+                "models/InfiniteYou/InfuseNetModel/diffusion_pytorch_model-00001-of-00002.safetensors",
+                "models/InfiniteYou/InfuseNetModel/diffusion_pytorch_model-00002-of-00002.safetensors"
+            ],
+            "models/InfiniteYou/image_proj_model.bin",
+            ],
+    },
+    # ESRGAN
+    "ESRGAN_x4": [
+        ("AI-ModelScope/Real-ESRGAN", "RealESRGAN_x4.pth", "models/ESRGAN"),
+    ],
+    # RIFE
+    "RIFE": [
+        ("AI-ModelScope/RIFE", "flownet.pkl", "models/RIFE"),
+    ],
+    # Omnigen
+    "OmniGen-v1": {
+        "file_list": [
+            ("BAAI/OmniGen-v1", "vae/diffusion_pytorch_model.safetensors", "models/OmniGen/OmniGen-v1/vae"),
+            ("BAAI/OmniGen-v1", "model.safetensors", "models/OmniGen/OmniGen-v1"),
+            ("BAAI/OmniGen-v1", "config.json", "models/OmniGen/OmniGen-v1"),
+            ("BAAI/OmniGen-v1", "special_tokens_map.json", "models/OmniGen/OmniGen-v1"),
+            ("BAAI/OmniGen-v1", "tokenizer_config.json", "models/OmniGen/OmniGen-v1"),
+            ("BAAI/OmniGen-v1", "tokenizer.json", "models/OmniGen/OmniGen-v1"),
+        ],
+        "load_path": [
+            "models/OmniGen/OmniGen-v1/vae/diffusion_pytorch_model.safetensors",
+            "models/OmniGen/OmniGen-v1/model.safetensors",
+        ]
+    },
+    # CogVideo
+    "CogVideoX-5B": {
+        "file_list": [
+            ("ZhipuAI/CogVideoX-5b", "text_encoder/config.json", "models/CogVideo/CogVideoX-5b/text_encoder"),
+            ("ZhipuAI/CogVideoX-5b", "text_encoder/model.safetensors.index.json", "models/CogVideo/CogVideoX-5b/text_encoder"),
+            ("ZhipuAI/CogVideoX-5b", "text_encoder/model-00001-of-00002.safetensors", "models/CogVideo/CogVideoX-5b/text_encoder"),
+            ("ZhipuAI/CogVideoX-5b", "text_encoder/model-00002-of-00002.safetensors", "models/CogVideo/CogVideoX-5b/text_encoder"),
+            ("ZhipuAI/CogVideoX-5b", "transformer/config.json", "models/CogVideo/CogVideoX-5b/transformer"),
+            ("ZhipuAI/CogVideoX-5b", "transformer/diffusion_pytorch_model.safetensors.index.json", "models/CogVideo/CogVideoX-5b/transformer"),
+            ("ZhipuAI/CogVideoX-5b", "transformer/diffusion_pytorch_model-00001-of-00002.safetensors", "models/CogVideo/CogVideoX-5b/transformer"),
+            ("ZhipuAI/CogVideoX-5b", "transformer/diffusion_pytorch_model-00002-of-00002.safetensors", "models/CogVideo/CogVideoX-5b/transformer"),
+            ("ZhipuAI/CogVideoX-5b", "vae/diffusion_pytorch_model.safetensors", "models/CogVideo/CogVideoX-5b/vae"),
+        ],
+        "load_path": [
+            "models/CogVideo/CogVideoX-5b/text_encoder",
+            "models/CogVideo/CogVideoX-5b/transformer",
+            "models/CogVideo/CogVideoX-5b/vae/diffusion_pytorch_model.safetensors",
+        ],
+    },
+    # Stable Diffusion 3.5
+    "StableDiffusion3.5-large": [
+        ("AI-ModelScope/stable-diffusion-3.5-large", "sd3.5_large.safetensors", "models/stable_diffusion_3"),
+        ("AI-ModelScope/stable-diffusion-3.5-large", "text_encoders/clip_l.safetensors", "models/stable_diffusion_3/text_encoders"),
+        ("AI-ModelScope/stable-diffusion-3.5-large", "text_encoders/clip_g.safetensors", "models/stable_diffusion_3/text_encoders"),
+        ("AI-ModelScope/stable-diffusion-3.5-large", "text_encoders/t5xxl_fp16.safetensors", "models/stable_diffusion_3/text_encoders"),
+    ],
+    "StableDiffusion3.5-medium": [
+        ("AI-ModelScope/stable-diffusion-3.5-medium", "sd3.5_medium.safetensors", "models/stable_diffusion_3"),
+        ("AI-ModelScope/stable-diffusion-3.5-large", "text_encoders/clip_l.safetensors", "models/stable_diffusion_3/text_encoders"),
+        ("AI-ModelScope/stable-diffusion-3.5-large", "text_encoders/clip_g.safetensors", "models/stable_diffusion_3/text_encoders"),
+        ("AI-ModelScope/stable-diffusion-3.5-large", "text_encoders/t5xxl_fp16.safetensors", "models/stable_diffusion_3/text_encoders"),
+    ],
+    "StableDiffusion3.5-large-turbo": [
+        ("AI-ModelScope/stable-diffusion-3.5-large-turbo", "sd3.5_large_turbo.safetensors", "models/stable_diffusion_3"),
+        ("AI-ModelScope/stable-diffusion-3.5-large", "text_encoders/clip_l.safetensors", "models/stable_diffusion_3/text_encoders"),
+        ("AI-ModelScope/stable-diffusion-3.5-large", "text_encoders/clip_g.safetensors", "models/stable_diffusion_3/text_encoders"),
+        ("AI-ModelScope/stable-diffusion-3.5-large", "text_encoders/t5xxl_fp16.safetensors", "models/stable_diffusion_3/text_encoders"),
+    ],
+    "HunyuanVideo":{
+        "file_list": [
+            ("AI-ModelScope/clip-vit-large-patch14", "model.safetensors", "models/HunyuanVideo/text_encoder"),
+            ("DiffSynth-Studio/HunyuanVideo_MLLM_text_encoder", "model-00001-of-00004.safetensors", "models/HunyuanVideo/text_encoder_2"),
+            ("DiffSynth-Studio/HunyuanVideo_MLLM_text_encoder", "model-00002-of-00004.safetensors", "models/HunyuanVideo/text_encoder_2"),
+            ("DiffSynth-Studio/HunyuanVideo_MLLM_text_encoder", "model-00003-of-00004.safetensors", "models/HunyuanVideo/text_encoder_2"),
+            ("DiffSynth-Studio/HunyuanVideo_MLLM_text_encoder", "model-00004-of-00004.safetensors", "models/HunyuanVideo/text_encoder_2"),
+            ("DiffSynth-Studio/HunyuanVideo_MLLM_text_encoder", "config.json", "models/HunyuanVideo/text_encoder_2"),
+            ("DiffSynth-Studio/HunyuanVideo_MLLM_text_encoder", "model.safetensors.index.json", "models/HunyuanVideo/text_encoder_2"),
+            ("AI-ModelScope/HunyuanVideo", "hunyuan-video-t2v-720p/vae/pytorch_model.pt", "models/HunyuanVideo/vae"),
+            ("AI-ModelScope/HunyuanVideo", "hunyuan-video-t2v-720p/transformers/mp_rank_00_model_states.pt", "models/HunyuanVideo/transformers")
+        ],
+        "load_path": [
+            "models/HunyuanVideo/text_encoder/model.safetensors",
+            "models/HunyuanVideo/text_encoder_2",
+            "models/HunyuanVideo/vae/pytorch_model.pt",
+            "models/HunyuanVideo/transformers/mp_rank_00_model_states.pt"
+        ],
+    },
+    "HunyuanVideoI2V":{
+        "file_list": [
+            ("AI-ModelScope/clip-vit-large-patch14", "model.safetensors", "models/HunyuanVideoI2V/text_encoder"),
+            ("AI-ModelScope/llava-llama-3-8b-v1_1-transformers", "model-00001-of-00004.safetensors", "models/HunyuanVideoI2V/text_encoder_2"),
+            ("AI-ModelScope/llava-llama-3-8b-v1_1-transformers", "model-00002-of-00004.safetensors", "models/HunyuanVideoI2V/text_encoder_2"),
+            ("AI-ModelScope/llava-llama-3-8b-v1_1-transformers", "model-00003-of-00004.safetensors", "models/HunyuanVideoI2V/text_encoder_2"),
+            ("AI-ModelScope/llava-llama-3-8b-v1_1-transformers", "model-00004-of-00004.safetensors", "models/HunyuanVideoI2V/text_encoder_2"),
+            ("AI-ModelScope/llava-llama-3-8b-v1_1-transformers", "config.json", "models/HunyuanVideoI2V/text_encoder_2"),
+            ("AI-ModelScope/llava-llama-3-8b-v1_1-transformers", "model.safetensors.index.json", "models/HunyuanVideoI2V/text_encoder_2"),
+            ("AI-ModelScope/HunyuanVideo-I2V", "hunyuan-video-i2v-720p/vae/pytorch_model.pt", "models/HunyuanVideoI2V/vae"),
+            ("AI-ModelScope/HunyuanVideo-I2V", "hunyuan-video-i2v-720p/transformers/mp_rank_00_model_states.pt", "models/HunyuanVideoI2V/transformers")
+        ],
+        "load_path": [
+            "models/HunyuanVideoI2V/text_encoder/model.safetensors",
+            "models/HunyuanVideoI2V/text_encoder_2",
+            "models/HunyuanVideoI2V/vae/pytorch_model.pt",
+            "models/HunyuanVideoI2V/transformers/mp_rank_00_model_states.pt"
+        ],
+    },
+    "HunyuanVideo-fp8":{
+        "file_list": [
+            ("AI-ModelScope/clip-vit-large-patch14", "model.safetensors", "models/HunyuanVideo/text_encoder"),
+            ("DiffSynth-Studio/HunyuanVideo_MLLM_text_encoder", "model-00001-of-00004.safetensors", "models/HunyuanVideo/text_encoder_2"),
+            ("DiffSynth-Studio/HunyuanVideo_MLLM_text_encoder", "model-00002-of-00004.safetensors", "models/HunyuanVideo/text_encoder_2"),
+            ("DiffSynth-Studio/HunyuanVideo_MLLM_text_encoder", "model-00003-of-00004.safetensors", "models/HunyuanVideo/text_encoder_2"),
+            ("DiffSynth-Studio/HunyuanVideo_MLLM_text_encoder", "model-00004-of-00004.safetensors", "models/HunyuanVideo/text_encoder_2"),
+            ("DiffSynth-Studio/HunyuanVideo_MLLM_text_encoder", "config.json", "models/HunyuanVideo/text_encoder_2"),
+            ("DiffSynth-Studio/HunyuanVideo_MLLM_text_encoder", "model.safetensors.index.json", "models/HunyuanVideo/text_encoder_2"),
+            ("AI-ModelScope/HunyuanVideo", "hunyuan-video-t2v-720p/vae/pytorch_model.pt", "models/HunyuanVideo/vae"),
+            ("DiffSynth-Studio/HunyuanVideo-safetensors", "model.fp8.safetensors", "models/HunyuanVideo/transformers")
+        ],
+        "load_path": [
+            "models/HunyuanVideo/text_encoder/model.safetensors",
+            "models/HunyuanVideo/text_encoder_2",
+            "models/HunyuanVideo/vae/pytorch_model.pt",
+            "models/HunyuanVideo/transformers/model.fp8.safetensors"
+        ],
+    },
+}
+Preset_model_id: TypeAlias = Literal[
+    "HunyuanDiT",
+    "stable-video-diffusion-img2vid-xt",
+    "ExVideo-SVD-128f-v1",
+    "ExVideo-CogVideoX-LoRA-129f-v1",
+    "StableDiffusion_v15",
+    "DreamShaper_8",
+    "AingDiffusion_v12",
+    "Flat2DAnimerge_v45Sharp",
+    "TextualInversion_VeryBadImageNegative_v1.3",
+    "StableDiffusionXL_v1",
+    "BluePencilXL_v200",
+    "StableDiffusionXL_Turbo",
+    "ControlNet_v11f1p_sd15_depth",
+    "ControlNet_v11p_sd15_softedge",
+    "ControlNet_v11f1e_sd15_tile",
+    "ControlNet_v11p_sd15_lineart",
+    "AnimateDiff_v2",
+    "AnimateDiff_xl_beta",
+    "RIFE",
+    "BeautifulPrompt",
+    "opus-mt-zh-en",
+    "IP-Adapter-SD",
+    "IP-Adapter-SDXL",
+    "StableDiffusion3",
+    "StableDiffusion3_without_T5",
+    "Kolors",
+    "SDXL-vae-fp16-fix",
+    "ControlNet_union_sdxl_promax",
+    "FLUX.1-dev",
+    "FLUX.1-schnell",
+    "InstantX/FLUX.1-dev-Controlnet-Union-alpha",
+    "jasperai/Flux.1-dev-Controlnet-Depth",
+    "jasperai/Flux.1-dev-Controlnet-Surface-Normals",
+    "jasperai/Flux.1-dev-Controlnet-Upscaler",
+    "alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Alpha",
+    "alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Beta",
+    "Shakker-Labs/FLUX.1-dev-ControlNet-Depth",
+    "Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro",
+    "InstantX/FLUX.1-dev-IP-Adapter",
+    "InfiniteYou",
+    "SDXL_lora_zyd232_ChineseInkStyle_SDXL_v1_0",
+    "QwenPrompt",
+    "OmostPrompt",
+    "ESRGAN_x4",
+    "RIFE",
+    "OmniGen-v1",
+    "CogVideoX-5B",
+    "Annotators:Depth",
+    "Annotators:Softedge",
+    "Annotators:Lineart",
+    "Annotators:Normal",
+    "Annotators:Openpose",
+    "StableDiffusion3.5-large",
+    "StableDiffusion3.5-medium",
+    "HunyuanVideo",
+    "HunyuanVideo-fp8",
+    "HunyuanVideoI2V",
+]

diffsynth/data/__init__.py

@@ -0,0 +1 @@
+from .video import VideoData, save_video, save_frames

diffsynth/data/video.py

@@ -0,0 +1,244 @@
+import os
+
+import imageio
+import imageio_ffmpeg as ffmpeg
+import matplotlib.pyplot as plt
+import numpy as np
+from matplotlib import cm
+from PIL import Image
+from tqdm import tqdm
+
+
+class LowMemoryVideo:
+    def __init__(self, file_name):
+        self.reader = imageio.get_reader(file_name)
+
+    def __len__(self):
+        return self.reader.count_frames()
+
+    def __getitem__(self, item):
+        return Image.fromarray(np.array(self.reader.get_data(item))).convert("RGB")
+
+    def __del__(self):
+        self.reader.close()
+
+
+def split_file_name(file_name):
+    result = []
+    number = -1
+    for i in file_name:
+        if ord(i) >= ord("0") and ord(i) <= ord("9"):
+            if number == -1:
+                number = 0
+            number = number * 10 + ord(i) - ord("0")
+        else:
+            if number != -1:
+                result.append(number)
+                number = -1
+            result.append(i)
+    if number != -1:
+        result.append(number)
+    result = tuple(result)
+    return result
+
+
+def search_for_images(folder):
+    file_list = [
+        i for i in os.listdir(folder) if i.endswith(".jpg") or i.endswith(".png")
+    ]
+    file_list = [(split_file_name(file_name), file_name)
+                 for file_name in file_list]
+    file_list = [i[1] for i in sorted(file_list)]
+    file_list = [os.path.join(folder, i) for i in file_list]
+    return file_list
+
+
+class LowMemoryImageFolder:
+    def __init__(self, folder, file_list=None):
+        if file_list is None:
+            self.file_list = search_for_images(folder)
+        else:
+            self.file_list = [
+                os.path.join(folder, file_name) for file_name in file_list
+            ]
+
+    def __len__(self):
+        return len(self.file_list)
+
+    def __getitem__(self, item):
+        return Image.open(self.file_list[item]).convert("RGB")
+
+    def __del__(self):
+        pass
+
+
+def crop_and_resize(image, height, width):
+    image = np.array(image)
+    image_height, image_width, _ = image.shape
+    if image_height / image_width < height / width:
+        croped_width = int(image_height / height * width)
+        left = (image_width - croped_width) // 2
+        image = image[:, left: left + croped_width]
+        image = Image.fromarray(image).resize((width, height))
+    else:
+        croped_height = int(image_width / width * height)
+        left = (image_height - croped_height) // 2
+        image = image[left: left + croped_height, :]
+        image = Image.fromarray(image).resize((width, height))
+    return image
+
+
+class VideoData:
+    def __init__(
+        self, video_file=None, image_folder=None, height=None, width=None, **kwargs
+    ):
+        if video_file is not None:
+            self.data_type = "video"
+            self.data = LowMemoryVideo(video_file, **kwargs)
+        elif image_folder is not None:
+            self.data_type = "images"
+            self.data = LowMemoryImageFolder(image_folder, **kwargs)
+        else:
+            raise ValueError("Cannot open video or image folder")
+        self.length = None
+        self.set_shape(height, width)
+
+    def raw_data(self):
+        frames = []
+        for i in range(self.__len__()):
+            frames.append(self.__getitem__(i))
+        return frames
+
+    def set_length(self, length):
+        self.length = length
+
+    def set_shape(self, height, width):
+        self.height = height
+        self.width = width
+
+    def __len__(self):
+        if self.length is None:
+            return len(self.data)
+        else:
+            return self.length
+
+    def shape(self):
+        if self.height is not None and self.width is not None:
+            return self.height, self.width
+        else:
+            height, width, _ = self.__getitem__(0).shape
+            return height, width
+
+    def __getitem__(self, item):
+        frame = self.data.__getitem__(item)
+        width, height = frame.size
+        if self.height is not None and self.width is not None:
+            if self.height != height or self.width != width:
+                frame = crop_and_resize(frame, self.height, self.width)
+        return frame
+
+    def __del__(self):
+        pass
+
+    def save_images(self, folder):
+        os.makedirs(folder, exist_ok=True)
+        for i in tqdm(range(self.__len__()), desc="Saving images"):
+            frame = self.__getitem__(i)
+            frame.save(os.path.join(folder, f"{i}.png"))
+
+
+def save_video_ffmpeg(frames, save_path, fps):
+    # frames: numpy array T H W C, dtype=uint8
+    frames = np.array(frames)
+    if frames.dtype != np.uint8:
+        frames = (frames * 255).clip(0, 255).astype(np.uint8)
+    T, H, W, C = frames.shape
+    assert C in [1, 3, 4]
+
+    writer = ffmpeg.write_frames(
+        save_path,
+        (W, H),
+        fps=fps,
+        quality=9,
+        macro_block_size=None,  # 避免补边
+    )
+    for frame in frames:
+        writer.send(frame)
+
+    writer.close()
+
+
+def save_video(frames, save_path, fps, quality=9, ffmpeg_params=None, grayscale=True):
+    writer = imageio.get_writer(
+        save_path, fps=fps, quality=quality, macro_block_size=1, ffmpeg_params=ffmpeg_params
+    )
+    if not grayscale:
+        cmap = plt.get_cmap('Spectral_r')
+        lut = (cmap(np.linspace(0, 1, 256))[:, :3] * 255).astype(np.uint8)
+
+    for frame in frames:
+        frame = np.array(frame)
+
+        if not grayscale:
+            if frame.ndim == 3:
+                if frame.shape[-1] >= 3:
+                    frame = frame[..., 0]
+                elif frame.shape[-1] == 1:
+                    frame = frame[..., 0]
+
+            if frame.dtype in [np.float32, np.float64]:
+                indices = (frame * 255).clip(0, 255).astype(np.uint8)
+            else:
+                indices = frame.clip(0, 255).astype(np.uint8)
+
+            frame_out = lut[indices]
+
+        else:
+
+            if frame.dtype in [np.float32, np.float64]:
+                frame_out = (frame * 255).clip(0, 255).astype(np.uint8)
+            else:
+                frame_out = frame.astype(np.uint8)
+
+            if frame_out.ndim == 3 and frame_out.shape[-1] == 1:
+                frame_out = frame_out[..., 0]
+
+        writer.append_data(frame_out)
+
+    writer.close()
+    # writer = imageio.get_writer(
+    #     save_path, fps=fps, quality=quality, ffmpeg_params=ffmpeg_params
+    # )
+    # import matplotlib.pyplot as plt
+    # from matplotlib import cm
+
+    # cmap = plt.get_cmap('Spectral_r') if not grayscale else None
+
+    # for frame in frames:
+    #     frame = np.array(frame)
+
+    #     if not grayscale:
+    #         if frame.ndim == 3 and frame.shape[-1] >= 3:
+    #             frame = frame[..., 0]
+
+    #         if frame.dtype == np.uint8:
+    #             frame = frame.astype(np.float32) / 255.0
+
+    #         rgba_frame = cmap(frame)
+    #         frame = (rgba_frame[..., :3] * 255).clip(0, 255).astype(np.uint8)
+
+    #     else:
+    #         if frame.dtype == np.float32 or frame.dtype == np.float64:
+    #             frame = (frame * 255).clip(0, 255).astype(np.uint8)
+
+    #     writer.append_data(frame)
+
+    # writer.close()
+
+
+def save_frames(frames, save_path):
+    os.makedirs(save_path, exist_ok=True)
+    for i, frame in enumerate(tqdm(frames, desc="Saving images")):
+        frame.save(os.path.join(save_path, f"{i}.png"))
+        frame.save(os.path.join(save_path, f"{i}.png"))
+        frame.save(os.path.join(save_path, f"{i}.png"))

diffsynth/distributed/xdit_context_parallel.py

@@ -0,0 +1,129 @@
+import torch
+from typing import Optional
+from einops import rearrange
+from xfuser.core.distributed import (get_sequence_parallel_rank,
+                                     get_sequence_parallel_world_size,
+                                     get_sp_group)
+from xfuser.core.long_ctx_attention import xFuserLongContextAttention
+
+def sinusoidal_embedding_1d(dim, position):
+    sinusoid = torch.outer(position.type(torch.float64), torch.pow(
+        10000, -torch.arange(dim//2, dtype=torch.float64, device=position.device).div(dim//2)))
+    x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
+    return x.to(position.dtype)
+
+def pad_freqs(original_tensor, target_len):
+    seq_len, s1, s2 = original_tensor.shape
+    pad_size = target_len - seq_len
+    padding_tensor = torch.ones(
+        pad_size,
+        s1,
+        s2,
+        dtype=original_tensor.dtype,
+        device=original_tensor.device)
+    padded_tensor = torch.cat([original_tensor, padding_tensor], dim=0)
+    return padded_tensor
+    
+def rope_apply(x, freqs, num_heads):
+    x = rearrange(x, "b s (n d) -> b s n d", n=num_heads)
+    s_per_rank = x.shape[1]
+
+    x_out = torch.view_as_complex(x.to(torch.float64).reshape(
+        x.shape[0], x.shape[1], x.shape[2], -1, 2))
+
+    sp_size = get_sequence_parallel_world_size()
+    sp_rank = get_sequence_parallel_rank()
+    freqs = pad_freqs(freqs, s_per_rank * sp_size)
+    freqs_rank = freqs[(sp_rank * s_per_rank):((sp_rank + 1) * s_per_rank), :, :]
+
+    x_out = torch.view_as_real(x_out * freqs_rank).flatten(2)
+    return x_out.to(x.dtype)
+
+def usp_dit_forward(self,
+            x: torch.Tensor,
+            timestep: torch.Tensor,
+            context: torch.Tensor,
+            clip_feature: Optional[torch.Tensor] = None,
+            y: Optional[torch.Tensor] = None,
+            use_gradient_checkpointing: bool = False,
+            use_gradient_checkpointing_offload: bool = False,
+            **kwargs,
+            ):
+    t = self.time_embedding(
+        sinusoidal_embedding_1d(self.freq_dim, timestep))
+    t_mod = self.time_projection(t).unflatten(1, (6, self.dim))
+    context = self.text_embedding(context)
+    
+    if self.has_image_input:
+        x = torch.cat([x, y], dim=1)  # (b, c_x + c_y, f, h, w)
+        clip_embdding = self.img_emb(clip_feature)
+        context = torch.cat([clip_embdding, context], dim=1)
+    
+    x, (f, h, w) = self.patchify(x)
+    
+    freqs = torch.cat([
+        self.freqs[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1),
+        self.freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1),
+        self.freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1)
+    ], dim=-1).reshape(f * h * w, 1, -1).to(x.device)
+    
+    def create_custom_forward(module):
+        def custom_forward(*inputs):
+            return module(*inputs)
+        return custom_forward
+
+    # Context Parallel
+    x = torch.chunk(
+        x, get_sequence_parallel_world_size(),
+        dim=1)[get_sequence_parallel_rank()]
+
+    for block in self.blocks:
+        if self.training and use_gradient_checkpointing:
+            if use_gradient_checkpointing_offload:
+                with torch.autograd.graph.save_on_cpu():
+                    x = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(block),
+                        x, context, t_mod, freqs,
+                        use_reentrant=False,
+                    )
+            else:
+                x = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    x, context, t_mod, freqs,
+                    use_reentrant=False,
+                )
+        else:
+            x = block(x, context, t_mod, freqs)
+
+    x = self.head(x, t)
+
+    # Context Parallel
+    x = get_sp_group().all_gather(x, dim=1)
+
+    # unpatchify
+    x = self.unpatchify(x, (f, h, w))
+    return x
+
+
+def usp_attn_forward(self, x, freqs):
+    q = self.norm_q(self.q(x))
+    k = self.norm_k(self.k(x))
+    v = self.v(x)
+
+    q = rope_apply(q, freqs, self.num_heads)
+    k = rope_apply(k, freqs, self.num_heads)
+    q = rearrange(q, "b s (n d) -> b s n d", n=self.num_heads)
+    k = rearrange(k, "b s (n d) -> b s n d", n=self.num_heads)
+    v = rearrange(v, "b s (n d) -> b s n d", n=self.num_heads)
+
+    x = xFuserLongContextAttention()(
+        None,
+        query=q,
+        key=k,
+        value=v,
+    )
+    x = x.flatten(2)
+
+    del q, k, v
+    torch.cuda.empty_cache()
+    return self.o(x)

diffsynth/models/__init__.py

@@ -0,0 +1 @@
+from .model_manager import *

diffsynth/models/downloader.py

@@ -0,0 +1,116 @@
+import os
+import shutil
+from typing import List
+
+from huggingface_hub import hf_hub_download
+from modelscope import snapshot_download
+from typing_extensions import Literal, TypeAlias
+
+from ..configs.model_config import (Preset_model_id,
+                                    preset_models_on_huggingface,
+                                    preset_models_on_modelscope)
+
+
+def download_from_modelscope(model_id, origin_file_path, local_dir):
+    os.makedirs(local_dir, exist_ok=True)
+    file_name = os.path.basename(origin_file_path)
+    if file_name in os.listdir(local_dir):
+        print(f"    {file_name} has been already in {local_dir}.")
+    else:
+        print(f"    Start downloading {os.path.join(local_dir, file_name)}")
+        snapshot_download(model_id, allow_file_pattern=origin_file_path, local_dir=local_dir)
+        downloaded_file_path = os.path.join(local_dir, origin_file_path)
+        target_file_path = os.path.join(local_dir, os.path.split(origin_file_path)[-1])
+        if downloaded_file_path != target_file_path:
+            shutil.move(downloaded_file_path, target_file_path)
+            shutil.rmtree(os.path.join(local_dir, origin_file_path.split("/")[0]))
+
+
+def download_from_huggingface(model_id, origin_file_path, local_dir):
+    os.makedirs(local_dir, exist_ok=True)
+    file_name = os.path.basename(origin_file_path)
+    if file_name in os.listdir(local_dir):
+        print(f"    {file_name} has been already in {local_dir}.")
+    else:
+        print(f"    Start downloading {os.path.join(local_dir, file_name)}")
+        hf_hub_download(model_id, origin_file_path, local_dir=local_dir)
+        downloaded_file_path = os.path.join(local_dir, origin_file_path)
+        target_file_path = os.path.join(local_dir, file_name)
+        if downloaded_file_path != target_file_path:
+            shutil.move(downloaded_file_path, target_file_path)
+            shutil.rmtree(os.path.join(local_dir, origin_file_path.split("/")[0]))
+
+
+Preset_model_website: TypeAlias = Literal[
+    "HuggingFace",
+    "ModelScope",
+]
+website_to_preset_models = {
+    "HuggingFace": preset_models_on_huggingface,
+    "ModelScope": preset_models_on_modelscope,
+}
+website_to_download_fn = {
+    "HuggingFace": download_from_huggingface,
+    "ModelScope": download_from_modelscope,
+}
+
+
+def download_customized_models(
+    model_id,
+    origin_file_path,
+    local_dir,
+    downloading_priority: List[Preset_model_website] = ["ModelScope", "HuggingFace"],
+):
+    downloaded_files = []
+    for website in downloading_priority:
+        # Check if the file is downloaded.
+        file_to_download = os.path.join(local_dir, os.path.basename(origin_file_path))
+        if file_to_download in downloaded_files:
+            continue
+        # Download
+        website_to_download_fn[website](model_id, origin_file_path, local_dir)
+        if os.path.basename(origin_file_path) in os.listdir(local_dir):
+            downloaded_files.append(file_to_download)
+    return downloaded_files
+
+
+def download_models(
+    model_id_list: List[Preset_model_id] = [],
+    downloading_priority: List[Preset_model_website] = ["ModelScope", "HuggingFace"],
+):
+    print(f"Downloading models: {model_id_list}")
+    downloaded_files = []
+    load_files = []
+
+    for model_id in model_id_list:
+        for website in downloading_priority:
+            if model_id in website_to_preset_models[website]:
+                
+                # Parse model metadata
+                model_metadata = website_to_preset_models[website][model_id]
+                if isinstance(model_metadata, list):
+                    file_data = model_metadata
+                else:
+                    file_data = model_metadata.get("file_list", [])
+
+                # Try downloading the model from this website.
+                model_files = []
+                for model_id, origin_file_path, local_dir in file_data:
+                    # Check if the file is downloaded.
+                    file_to_download = os.path.join(local_dir, os.path.basename(origin_file_path))
+                    if file_to_download in downloaded_files:
+                        continue
+                    # Download
+                    website_to_download_fn[website](model_id, origin_file_path, local_dir)
+                    if os.path.basename(origin_file_path) in os.listdir(local_dir):
+                        downloaded_files.append(file_to_download)
+                        model_files.append(file_to_download)
+                
+                # If the model is successfully downloaded, break.
+                if len(model_files) > 0:
+                    if isinstance(model_metadata, dict) and "load_path" in model_metadata:
+                        model_files = model_metadata["load_path"]
+                    load_files.extend(model_files)
+                    break
+                
+    return load_files

diffsynth/models/model_manager.py

@@ -0,0 +1,416 @@
+import importlib
+import json
+import os
+from typing import List
+
+import torch
+
+from ..configs.model_config import (huggingface_model_loader_configs,
+                                    model_loader_configs,
+                                    patch_model_loader_configs)
+from .downloader import (Preset_model_id, Preset_model_website,
+                         download_customized_models, download_models)
+from .utils import (hash_state_dict_keys, init_weights_on_device,
+                    load_state_dict, split_state_dict_with_prefix)
+
+
+def load_model_from_single_file(state_dict, model_names, model_classes, model_resource, torch_dtype, device):
+    loaded_model_names, loaded_models = [], []
+    for model_name, model_class in zip(model_names, model_classes):
+        print(f"    model_name: {model_name} model_class: {model_class.__name__}")
+        state_dict_converter = model_class.state_dict_converter()
+        if model_resource == "civitai":
+            state_dict_results = state_dict_converter.from_civitai(state_dict)
+        elif model_resource == "diffusers":
+            state_dict_results = state_dict_converter.from_diffusers(state_dict)
+        if isinstance(state_dict_results, tuple):
+            model_state_dict, extra_kwargs = state_dict_results
+            print(f"        This model is initialized with extra kwargs: {extra_kwargs}")
+        else:
+            model_state_dict, extra_kwargs = state_dict_results, {}
+        torch_dtype = torch.float32 if extra_kwargs.get("upcast_to_float32", False) else torch_dtype
+        with init_weights_on_device():
+            model = model_class(**extra_kwargs)
+        if hasattr(model, "eval"):
+            model = model.eval()
+        model.load_state_dict(model_state_dict, assign=True)
+        model = model.to(dtype=torch_dtype, device=device)
+        loaded_model_names.append(model_name)
+        loaded_models.append(model)
+    return loaded_model_names, loaded_models
+
+
+def load_model_from_huggingface_folder(file_path, model_names, model_classes, torch_dtype, device):
+    loaded_model_names, loaded_models = [], []
+    for model_name, model_class in zip(model_names, model_classes):
+        if torch_dtype in [torch.float32, torch.float16, torch.bfloat16]:
+            model = model_class.from_pretrained(file_path, torch_dtype=torch_dtype).eval()
+        else:
+            model = model_class.from_pretrained(file_path).eval().to(dtype=torch_dtype)
+        if torch_dtype == torch.float16 and hasattr(model, "half"):
+            model = model.half()
+        try:
+            model = model.to(device=device)
+        except:
+            pass
+        loaded_model_names.append(model_name)
+        loaded_models.append(model)
+    return loaded_model_names, loaded_models
+
+
+def load_single_patch_model_from_single_file(state_dict, model_name, model_class, base_model, extra_kwargs, torch_dtype, device):
+    print(f"    model_name: {model_name} model_class: {model_class.__name__} extra_kwargs: {extra_kwargs}")
+    base_state_dict = base_model.state_dict()
+    base_model.to("cpu")
+    del base_model
+    model = model_class(**extra_kwargs)
+    model.load_state_dict(base_state_dict, strict=False)
+    model.load_state_dict(state_dict, strict=False)
+    model.to(dtype=torch_dtype, device=device)
+    return model
+
+
+def load_patch_model_from_single_file(state_dict, model_names, model_classes, extra_kwargs, model_manager, torch_dtype, device):
+    loaded_model_names, loaded_models = [], []
+    for model_name, model_class in zip(model_names, model_classes):
+        while True:
+            for model_id in range(len(model_manager.model)):
+                base_model_name = model_manager.model_name[model_id]
+                if base_model_name == model_name:
+                    base_model_path = model_manager.model_path[model_id]
+                    base_model = model_manager.model[model_id]
+                    print(f"    Adding patch model to {base_model_name} ({base_model_path})")
+                    patched_model = load_single_patch_model_from_single_file(
+                        state_dict, model_name, model_class, base_model, extra_kwargs, torch_dtype, device)
+                    loaded_model_names.append(base_model_name)
+                    loaded_models.append(patched_model)
+                    model_manager.model.pop(model_id)
+                    model_manager.model_path.pop(model_id)
+                    model_manager.model_name.pop(model_id)
+                    break
+            else:
+                break
+    return loaded_model_names, loaded_models
+
+
+
+class ModelDetectorTemplate:
+    def __init__(self):
+        pass
+
+    def match(self, file_path="", state_dict={}):
+        return False
+    
+    def load(self, file_path="", state_dict={}, device="cuda", torch_dtype=torch.float16, **kwargs):
+        return [], []
+    
+
+
+class ModelDetectorFromSingleFile:
+    def __init__(self, model_loader_configs=[]):
+        self.keys_hash_with_shape_dict = {}
+        self.keys_hash_dict = {}
+        for metadata in model_loader_configs:
+            self.add_model_metadata(*metadata)
+
+
+    def add_model_metadata(self, keys_hash, keys_hash_with_shape, model_names, model_classes, model_resource):
+        self.keys_hash_with_shape_dict[keys_hash_with_shape] = (model_names, model_classes, model_resource)
+        if keys_hash is not None:
+            self.keys_hash_dict[keys_hash] = (model_names, model_classes, model_resource)
+
+
+    def match(self, file_path="", state_dict={}):
+        if isinstance(file_path, str) and os.path.isdir(file_path):
+            return False
+        if len(state_dict) == 0:
+            state_dict = load_state_dict(file_path)
+        keys_hash_with_shape = hash_state_dict_keys(state_dict, with_shape=True)
+        if keys_hash_with_shape in self.keys_hash_with_shape_dict:
+            return True
+        keys_hash = hash_state_dict_keys(state_dict, with_shape=False)
+        if keys_hash in self.keys_hash_dict:
+            return True
+        return False
+
+
+    def load(self, file_path="", state_dict={}, device="cuda", torch_dtype=torch.float16, **kwargs):
+        if len(state_dict) == 0:
+            state_dict = load_state_dict(file_path)
+
+        # Load models with strict matching
+        keys_hash_with_shape = hash_state_dict_keys(state_dict, with_shape=True)
+        if keys_hash_with_shape in self.keys_hash_with_shape_dict:
+            model_names, model_classes, model_resource = self.keys_hash_with_shape_dict[keys_hash_with_shape]
+            loaded_model_names, loaded_models = load_model_from_single_file(state_dict, model_names, model_classes, model_resource, torch_dtype, device)
+            return loaded_model_names, loaded_models
+
+        # Load models without strict matching
+        # (the shape of parameters may be inconsistent, and the state_dict_converter will modify the model architecture)
+        keys_hash = hash_state_dict_keys(state_dict, with_shape=False)
+        if keys_hash in self.keys_hash_dict:
+            model_names, model_classes, model_resource = self.keys_hash_dict[keys_hash]
+            loaded_model_names, loaded_models = load_model_from_single_file(state_dict, model_names, model_classes, model_resource, torch_dtype, device)
+            return loaded_model_names, loaded_models
+
+        return loaded_model_names, loaded_models
+
+
+
+class ModelDetectorFromSplitedSingleFile(ModelDetectorFromSingleFile):
+    def __init__(self, model_loader_configs=[]):
+        super().__init__(model_loader_configs)
+
+
+    def match(self, file_path="", state_dict={}):
+        if isinstance(file_path, str) and os.path.isdir(file_path):
+            return False
+        if len(state_dict) == 0:
+            state_dict = load_state_dict(file_path)
+        splited_state_dict = split_state_dict_with_prefix(state_dict)
+        for sub_state_dict in splited_state_dict:
+            if super().match(file_path, sub_state_dict):
+                return True
+        return False
+
+
+    def load(self, file_path="", state_dict={}, device="cuda", torch_dtype=torch.float16, **kwargs):
+        # Split the state_dict and load from each component
+        splited_state_dict = split_state_dict_with_prefix(state_dict)
+        valid_state_dict = {}
+        for sub_state_dict in splited_state_dict:
+            if super().match(file_path, sub_state_dict):
+                valid_state_dict.update(sub_state_dict)
+        if super().match(file_path, valid_state_dict):
+            loaded_model_names, loaded_models = super().load(file_path, valid_state_dict, device, torch_dtype)
+        else:
+            loaded_model_names, loaded_models = [], []
+            for sub_state_dict in splited_state_dict:
+                if super().match(file_path, sub_state_dict):
+                    loaded_model_names_, loaded_models_ = super().load(file_path, valid_state_dict, device, torch_dtype)
+                    loaded_model_names += loaded_model_names_
+                    loaded_models += loaded_models_
+        return loaded_model_names, loaded_models
+    
+
+
+class ModelDetectorFromHuggingfaceFolder:
+    def __init__(self, model_loader_configs=[]):
+        self.architecture_dict = {}
+        for metadata in model_loader_configs:
+            self.add_model_metadata(*metadata)
+
+
+    def add_model_metadata(self, architecture, huggingface_lib, model_name, redirected_architecture):
+        self.architecture_dict[architecture] = (huggingface_lib, model_name, redirected_architecture)
+
+
+    def match(self, file_path="", state_dict={}):
+        if not isinstance(file_path, str) or os.path.isfile(file_path):
+            return False
+        file_list = os.listdir(file_path)
+        if "config.json" not in file_list:
+            return False
+        with open(os.path.join(file_path, "config.json"), "r") as f:
+            config = json.load(f)
+        if "architectures" not in config and "_class_name" not in config:
+            return False
+        return True
+
+
+    def load(self, file_path="", state_dict={}, device="cuda", torch_dtype=torch.float16, **kwargs):
+        with open(os.path.join(file_path, "config.json"), "r") as f:
+            config = json.load(f)
+        loaded_model_names, loaded_models = [], []
+        architectures = config["architectures"] if "architectures" in config else [config["_class_name"]]
+        for architecture in architectures:
+            huggingface_lib, model_name, redirected_architecture = self.architecture_dict[architecture]
+            if redirected_architecture is not None:
+                architecture = redirected_architecture
+            model_class = importlib.import_module(huggingface_lib).__getattribute__(architecture)
+            loaded_model_names_, loaded_models_ = load_model_from_huggingface_folder(file_path, [model_name], [model_class], torch_dtype, device)
+            loaded_model_names += loaded_model_names_
+            loaded_models += loaded_models_
+        return loaded_model_names, loaded_models
+    
+
+
+class ModelDetectorFromPatchedSingleFile:
+    def __init__(self, model_loader_configs=[]):
+        self.keys_hash_with_shape_dict = {}
+        for metadata in model_loader_configs:
+            self.add_model_metadata(*metadata)
+
+
+    def add_model_metadata(self, keys_hash_with_shape, model_name, model_class, extra_kwargs):
+        self.keys_hash_with_shape_dict[keys_hash_with_shape] = (model_name, model_class, extra_kwargs)
+
+
+    def match(self, file_path="", state_dict={}):
+        if not isinstance(file_path, str) or os.path.isdir(file_path):
+            return False
+        if len(state_dict) == 0:
+            state_dict = load_state_dict(file_path)
+        keys_hash_with_shape = hash_state_dict_keys(state_dict, with_shape=True)
+        if keys_hash_with_shape in self.keys_hash_with_shape_dict:
+            return True
+        return False
+
+
+    def load(self, file_path="", state_dict={}, device="cuda", torch_dtype=torch.float16, model_manager=None, **kwargs):
+        if len(state_dict) == 0:
+            state_dict = load_state_dict(file_path)
+
+        # Load models with strict matching
+        loaded_model_names, loaded_models = [], []
+        keys_hash_with_shape = hash_state_dict_keys(state_dict, with_shape=True)
+        if keys_hash_with_shape in self.keys_hash_with_shape_dict:
+            model_names, model_classes, extra_kwargs = self.keys_hash_with_shape_dict[keys_hash_with_shape]
+            loaded_model_names_, loaded_models_ = load_patch_model_from_single_file(
+                state_dict, model_names, model_classes, extra_kwargs, model_manager, torch_dtype, device)
+            loaded_model_names += loaded_model_names_
+            loaded_models += loaded_models_
+        return loaded_model_names, loaded_models
+
+
+
+class ModelManager:
+    def __init__(
+        self,
+        torch_dtype=torch.float16,
+        device="cuda",
+        model_id_list: List[Preset_model_id] = [],
+        downloading_priority: List[Preset_model_website] = [ "HuggingFace","ModelScope"],
+        file_path_list: List[str] = [],
+    ):
+        self.torch_dtype = torch_dtype
+        self.device = device
+        self.model = []
+        self.model_path = []
+        self.model_name = []
+        downloaded_files = download_models(model_id_list, downloading_priority) if len(model_id_list) > 0 else []
+        self.model_detector = [
+            ModelDetectorFromSingleFile(model_loader_configs),
+            ModelDetectorFromSplitedSingleFile(model_loader_configs),
+            ModelDetectorFromHuggingfaceFolder(huggingface_model_loader_configs),
+            ModelDetectorFromPatchedSingleFile(patch_model_loader_configs),
+        ]
+        self.load_models(downloaded_files + file_path_list)
+
+
+    def load_model_from_single_file(self, file_path="", state_dict={}, model_names=[], model_classes=[], model_resource=None):
+        # print(f"Loading models from file: {file_path}")
+        if len(state_dict) == 0:
+            state_dict = load_state_dict(file_path)
+        model_names, models = load_model_from_single_file(state_dict, model_names, model_classes, model_resource, self.torch_dtype, self.device)
+        for model_name, model in zip(model_names, models):
+            self.model.append(model)
+            self.model_path.append(file_path)
+            self.model_name.append(model_name)
+        print(f"    The following models are loaded: {model_names}.")
+
+
+    def load_model_from_huggingface_folder(self, file_path="", model_names=[], model_classes=[]):
+        # print(f"Loading models from folder: {file_path}")
+        model_names, models = load_model_from_huggingface_folder(file_path, model_names, model_classes, self.torch_dtype, self.device)
+        for model_name, model in zip(model_names, models):
+            self.model.append(model)
+            self.model_path.append(file_path)
+            self.model_name.append(model_name)
+        print(f"    The following models are loaded: {model_names}.")
+
+
+    def load_patch_model_from_single_file(self, file_path="", state_dict={}, model_names=[], model_classes=[], extra_kwargs={}):
+        print(f"Loading patch models from file: {file_path}")
+        model_names, models = load_patch_model_from_single_file(
+            state_dict, model_names, model_classes, extra_kwargs, self, self.torch_dtype, self.device)
+        for model_name, model in zip(model_names, models):
+            self.model.append(model)
+            self.model_path.append(file_path)
+            self.model_name.append(model_name)
+        print(f"    The following patched models are loaded: {model_names}.")
+
+
+    def load_lora(self, file_path="", state_dict={}, lora_alpha=1.0):
+        if isinstance(file_path, list):
+            for file_path_ in file_path:
+                self.load_lora(file_path_, state_dict=state_dict, lora_alpha=lora_alpha)
+        else:
+            print(f"Loading LoRA models from file: {file_path}")
+            is_loaded = False
+            if len(state_dict) == 0:
+                state_dict = load_state_dict(file_path)
+            for model_name, model, model_path in zip(self.model_name, self.model, self.model_path):
+                for lora in get_lora_loaders():
+                    match_results = lora.match(model, state_dict)
+                    if match_results is not None:
+                        print(f"    Adding LoRA to {model_name} ({model_path}).")
+                        lora_prefix, model_resource = match_results
+                        lora.load(model, state_dict, lora_prefix, alpha=lora_alpha, model_resource=model_resource)
+                        is_loaded = True
+                        break
+            if not is_loaded:
+                print(f"    Cannot load LoRA: {file_path}")
+
+
+    def load_model(self, file_path, model_names=None, device=None, torch_dtype=None):
+        # print(f"Loading models from: {file_path}")
+        if device is None: device = self.device
+        if torch_dtype is None: torch_dtype = self.torch_dtype
+        if isinstance(file_path, list):
+            state_dict = {}
+            for path in file_path:
+                state_dict.update(load_state_dict(path))
+        elif os.path.isfile(file_path):
+            state_dict = load_state_dict(file_path)
+        else:
+            state_dict = None
+        for model_detector in self.model_detector:
+            if model_detector.match(file_path, state_dict):
+                model_names, models = model_detector.load(
+                    file_path, state_dict,
+                    device=device, torch_dtype=torch_dtype,
+                    allowed_model_names=model_names, model_manager=self
+                )
+                for model_name, model in zip(model_names, models):
+                    self.model.append(model)
+                    self.model_path.append(file_path)
+                    self.model_name.append(model_name)
+                print(f"    The following models are loaded: {model_names}.")
+                break
+        else:
+            print(f"    We cannot detect the model type. No models are loaded.")
+        
+
+    def load_models(self, file_path_list, model_names=None, device=None, torch_dtype=None):
+        for file_path in file_path_list:
+            self.load_model(file_path, model_names, device=device, torch_dtype=torch_dtype)
+
+    
+    def fetch_model(self, model_name, file_path=None, require_model_path=False):
+        fetched_models = []
+        fetched_model_paths = []
+        for model, model_path, model_name_ in zip(self.model, self.model_path, self.model_name):
+            if file_path is not None and file_path != model_path:
+                continue
+            if model_name == model_name_:
+                fetched_models.append(model)
+                fetched_model_paths.append(model_path)
+        if len(fetched_models) == 0:
+            print(f"No {model_name} models available.")
+            return None
+        if len(fetched_models) == 1:
+            pass 
+            # print(f"Using {model_name} from {fetched_model_paths[0]}.")
+        else:
+            print(f"More than one {model_name} models are loaded in model manager: {fetched_model_paths}. Using {model_name} from {fetched_model_paths[0]}.")
+        if require_model_path:
+            return fetched_models[0], fetched_model_paths[0]
+        else:
+            return fetched_models[0]
+        
+
+    def to(self, device):
+        for model in self.model:
+            model.to(device)
+

diffsynth/models/tiler.py

@@ -0,0 +1,234 @@
+import torch
+from einops import rearrange, repeat
+
+
+class TileWorker:
+    def __init__(self):
+        pass
+
+
+    def mask(self, height, width, border_width):
+        # Create a mask with shape (height, width).
+        # The centre area is filled with 1, and the border line is filled with values in range (0, 1].
+        x = torch.arange(height).repeat(width, 1).T
+        y = torch.arange(width).repeat(height, 1)
+        mask = torch.stack([x + 1, height - x, y + 1, width - y]).min(dim=0).values
+        mask = (mask / border_width).clip(0, 1)
+        return mask
+
+
+    def tile(self, model_input, tile_size, tile_stride, tile_device, tile_dtype):
+        # Convert a tensor (b, c, h, w) to (b, c, tile_size, tile_size, tile_num)
+        batch_size, channel, _, _ = model_input.shape
+        model_input = model_input.to(device=tile_device, dtype=tile_dtype)
+        unfold_operator = torch.nn.Unfold(
+            kernel_size=(tile_size, tile_size),
+            stride=(tile_stride, tile_stride)
+        )
+        model_input = unfold_operator(model_input)
+        model_input = model_input.view((batch_size, channel, tile_size, tile_size, -1))
+
+        return model_input
+
+
+    def tiled_inference(self, forward_fn, model_input, tile_batch_size, inference_device, inference_dtype, tile_device, tile_dtype):
+        # Call y=forward_fn(x) for each tile
+        tile_num = model_input.shape[-1]
+        model_output_stack = []
+
+        for tile_id in range(0, tile_num, tile_batch_size):
+
+            # process input
+            tile_id_ = min(tile_id + tile_batch_size, tile_num)
+            x = model_input[:, :, :, :, tile_id: tile_id_]
+            x = x.to(device=inference_device, dtype=inference_dtype)
+            x = rearrange(x, "b c h w n -> (n b) c h w")
+
+            # process output
+            y = forward_fn(x)
+            y = rearrange(y, "(n b) c h w -> b c h w n", n=tile_id_-tile_id)
+            y = y.to(device=tile_device, dtype=tile_dtype)
+            model_output_stack.append(y)
+
+        model_output = torch.concat(model_output_stack, dim=-1)
+        return model_output
+
+
+    def io_scale(self, model_output, tile_size):
+        # Determine the size modification happened in forward_fn
+        # We only consider the same scale on height and width.
+        io_scale = model_output.shape[2] / tile_size
+        return io_scale
+    
+
+    def untile(self, model_output, height, width, tile_size, tile_stride, border_width, tile_device, tile_dtype):
+        # The reversed function of tile
+        mask = self.mask(tile_size, tile_size, border_width)
+        mask = mask.to(device=tile_device, dtype=tile_dtype)
+        mask = rearrange(mask, "h w -> 1 1 h w 1")
+        model_output = model_output * mask
+
+        fold_operator = torch.nn.Fold(
+            output_size=(height, width),
+            kernel_size=(tile_size, tile_size),
+            stride=(tile_stride, tile_stride)
+        )
+        mask = repeat(mask[0, 0, :, :, 0], "h w -> 1 (h w) n", n=model_output.shape[-1])
+        model_output = rearrange(model_output, "b c h w n -> b (c h w) n")
+        model_output = fold_operator(model_output) / fold_operator(mask)
+
+        return model_output
+
+
+    def tiled_forward(self, forward_fn, model_input, tile_size, tile_stride, tile_batch_size=1, tile_device="cpu", tile_dtype=torch.float32, border_width=None):
+        # Prepare
+        inference_device, inference_dtype = model_input.device, model_input.dtype
+        height, width = model_input.shape[2], model_input.shape[3]
+        border_width = int(tile_stride*0.5) if border_width is None else border_width
+
+        # tile
+        model_input = self.tile(model_input, tile_size, tile_stride, tile_device, tile_dtype)
+
+        # inference
+        model_output = self.tiled_inference(forward_fn, model_input, tile_batch_size, inference_device, inference_dtype, tile_device, tile_dtype)
+
+        # resize
+        io_scale = self.io_scale(model_output, tile_size)
+        height, width = int(height*io_scale), int(width*io_scale)
+        tile_size, tile_stride = int(tile_size*io_scale), int(tile_stride*io_scale)
+        border_width = int(border_width*io_scale)
+
+        # untile
+        model_output = self.untile(model_output, height, width, tile_size, tile_stride, border_width, tile_device, tile_dtype)
+        
+        # Done!
+        model_output = model_output.to(device=inference_device, dtype=inference_dtype)
+        return model_output
+    
+
+
+class FastTileWorker:
+    def __init__(self):
+        pass
+
+
+    def build_mask(self, data, is_bound):
+        _, _, H, W = data.shape
+        h = repeat(torch.arange(H), "H -> H W", H=H, W=W)
+        w = repeat(torch.arange(W), "W -> H W", H=H, W=W)
+        border_width = (H + W) // 4
+        pad = torch.ones_like(h) * border_width
+        mask = torch.stack([
+            pad if is_bound[0] else h + 1,
+            pad if is_bound[1] else H - h,
+            pad if is_bound[2] else w + 1,
+            pad if is_bound[3] else W - w
+        ]).min(dim=0).values
+        mask = mask.clip(1, border_width)
+        mask = (mask / border_width).to(dtype=data.dtype, device=data.device)
+        mask = rearrange(mask, "H W -> 1 H W")
+        return mask
+
+
+    def tiled_forward(self, forward_fn, model_input, tile_size, tile_stride, tile_device="cpu", tile_dtype=torch.float32, border_width=None):
+        # Prepare
+        B, C, H, W = model_input.shape
+        border_width = int(tile_stride*0.5) if border_width is None else border_width
+        weight = torch.zeros((1, 1, H, W), dtype=tile_dtype, device=tile_device)
+        values = torch.zeros((B, C, H, W), dtype=tile_dtype, device=tile_device)
+
+        # Split tasks
+        tasks = []
+        for h in range(0, H, tile_stride):
+            for w in range(0, W, tile_stride):
+                if (h-tile_stride >= 0 and h-tile_stride+tile_size >= H) or (w-tile_stride >= 0 and w-tile_stride+tile_size >= W):
+                    continue
+                h_, w_ = h + tile_size, w + tile_size
+                if h_ > H: h, h_ = H - tile_size, H
+                if w_ > W: w, w_ = W - tile_size, W
+                tasks.append((h, h_, w, w_))
+        
+        # Run
+        for hl, hr, wl, wr in tasks:
+            # Forward
+            hidden_states_batch = forward_fn(hl, hr, wl, wr).to(dtype=tile_dtype, device=tile_device)
+
+            mask = self.build_mask(hidden_states_batch, is_bound=(hl==0, hr>=H, wl==0, wr>=W))
+            values[:, :, hl:hr, wl:wr] += hidden_states_batch * mask
+            weight[:, :, hl:hr, wl:wr] += mask
+        values /= weight
+        return values
+
+
+
+class TileWorker2Dto3D:
+    """
+    Process 3D tensors, but only enable TileWorker on 2D.
+    """
+    def __init__(self):
+        pass
+
+
+    def build_mask(self, T, H, W, dtype, device, is_bound, border_width):
+        t = repeat(torch.arange(T), "T -> T H W", T=T, H=H, W=W)
+        h = repeat(torch.arange(H), "H -> T H W", T=T, H=H, W=W)
+        w = repeat(torch.arange(W), "W -> T H W", T=T, H=H, W=W)
+        border_width = (H + W) // 4 if border_width is None else border_width
+        pad = torch.ones_like(h) * border_width
+        mask = torch.stack([
+            pad if is_bound[0] else t + 1,
+            pad if is_bound[1] else T - t,
+            pad if is_bound[2] else h + 1,
+            pad if is_bound[3] else H - h,
+            pad if is_bound[4] else w + 1,
+            pad if is_bound[5] else W - w
+        ]).min(dim=0).values
+        mask = mask.clip(1, border_width)
+        mask = (mask / border_width).to(dtype=dtype, device=device)
+        mask = rearrange(mask, "T H W -> 1 1 T H W")
+        return mask
+
+
+    def tiled_forward(
+        self,
+        forward_fn,
+        model_input,
+        tile_size, tile_stride,
+        tile_device="cpu", tile_dtype=torch.float32,
+        computation_device="cuda", computation_dtype=torch.float32,
+        border_width=None, scales=[1, 1, 1, 1],
+        progress_bar=lambda x:x
+    ):
+        B, C, T, H, W = model_input.shape
+        scale_C, scale_T, scale_H, scale_W = scales
+        tile_size_H, tile_size_W = tile_size
+        tile_stride_H, tile_stride_W = tile_stride
+
+        value = torch.zeros((B, int(C*scale_C), int(T*scale_T), int(H*scale_H), int(W*scale_W)), dtype=tile_dtype, device=tile_device)
+        weight = torch.zeros((1, 1, int(T*scale_T), int(H*scale_H), int(W*scale_W)), dtype=tile_dtype, device=tile_device)
+
+        # Split tasks
+        tasks = []
+        for h in range(0, H, tile_stride_H):
+            for w in range(0, W, tile_stride_W):
+                if (h-tile_stride_H >= 0 and h-tile_stride_H+tile_size_H >= H) or (w-tile_stride_W >= 0 and w-tile_stride_W+tile_size_W >= W):
+                    continue
+                h_, w_ = h + tile_size_H, w + tile_size_W
+                if h_ > H: h, h_ = max(H - tile_size_H, 0), H
+                if w_ > W: w, w_ = max(W - tile_size_W, 0), W
+                tasks.append((h, h_, w, w_))
+
+        # Run
+        for hl, hr, wl, wr in progress_bar(tasks):
+            mask = self.build_mask(
+                int(T*scale_T), int((hr-hl)*scale_H), int((wr-wl)*scale_W),
+                tile_dtype, tile_device,
+                is_bound=(True, True, hl==0, hr>=H, wl==0, wr>=W),
+                border_width=border_width
+            )
+            grid_input = model_input[:, :, :, hl:hr, wl:wr].to(dtype=computation_dtype, device=computation_device)
+            grid_output = forward_fn(grid_input).to(dtype=tile_dtype, device=tile_device)
+            value[:, :, :, int(hl*scale_H):int(hr*scale_H), int(wl*scale_W):int(wr*scale_W)] += grid_output * mask
+            weight[:, :, :, int(hl*scale_H):int(hr*scale_H), int(wl*scale_W):int(wr*scale_W)] += mask
+        value = value / weight
+        return value

diffsynth/models/utils.py

@@ -0,0 +1,185 @@
+import hashlib
+import os
+from contextlib import contextmanager
+
+import torch
+from safetensors import safe_open
+
+
+@contextmanager
+def init_weights_on_device(device = torch.device("meta"), include_buffers :bool = False):
+    
+    old_register_parameter = torch.nn.Module.register_parameter
+    if include_buffers:
+        old_register_buffer = torch.nn.Module.register_buffer
+    
+    def register_empty_parameter(module, name, param):
+        old_register_parameter(module, name, param)
+        if param is not None:
+            param_cls = type(module._parameters[name])
+            kwargs = module._parameters[name].__dict__
+            kwargs["requires_grad"] = param.requires_grad
+            module._parameters[name] = param_cls(module._parameters[name].to(device), **kwargs)
+
+    def register_empty_buffer(module, name, buffer, persistent=True):
+        old_register_buffer(module, name, buffer, persistent=persistent)
+        if buffer is not None:
+            module._buffers[name] = module._buffers[name].to(device)
+            
+    def patch_tensor_constructor(fn):
+        def wrapper(*args, **kwargs):
+            kwargs["device"] = device
+            return fn(*args, **kwargs)
+
+        return wrapper
+    
+    if include_buffers:
+        tensor_constructors_to_patch = {
+            torch_function_name: getattr(torch, torch_function_name)
+            for torch_function_name in ["empty", "zeros", "ones", "full"]
+        }
+    else:
+        tensor_constructors_to_patch = {}
+    
+    try:
+        torch.nn.Module.register_parameter = register_empty_parameter
+        if include_buffers:
+            torch.nn.Module.register_buffer = register_empty_buffer
+        for torch_function_name in tensor_constructors_to_patch.keys():
+            setattr(torch, torch_function_name, patch_tensor_constructor(getattr(torch, torch_function_name)))
+        yield
+    finally:
+        torch.nn.Module.register_parameter = old_register_parameter
+        if include_buffers:
+            torch.nn.Module.register_buffer = old_register_buffer
+        for torch_function_name, old_torch_function in tensor_constructors_to_patch.items():
+            setattr(torch, torch_function_name, old_torch_function)
+
+def load_state_dict_from_folder(file_path, torch_dtype=None):
+    state_dict = {}
+    for file_name in os.listdir(file_path):
+        if "." in file_name and file_name.split(".")[-1] in [
+            "safetensors", "bin", "ckpt", "pth", "pt"
+        ]:
+            state_dict.update(load_state_dict(os.path.join(file_path, file_name), torch_dtype=torch_dtype))
+    return state_dict
+
+
+def load_state_dict(file_path, torch_dtype=None, device="cpu"):
+    if file_path.endswith(".safetensors"):
+        return load_state_dict_from_safetensors(file_path, torch_dtype=torch_dtype, device=device)
+    else:
+        return load_state_dict_from_bin(file_path, torch_dtype=torch_dtype, device=device)
+
+
+def load_state_dict_from_safetensors(file_path, torch_dtype=None, device="cpu"):
+    state_dict = {}
+    with safe_open(file_path, framework="pt", device=device) as f:
+        for k in f.keys():
+            state_dict[k] = f.get_tensor(k)
+            if torch_dtype is not None:
+                state_dict[k] = state_dict[k].to(torch_dtype)
+    return state_dict
+
+
+def load_state_dict_from_bin(file_path, torch_dtype=None, device="cpu"):
+    state_dict = torch.load(file_path, map_location=device, weights_only=True)
+    if torch_dtype is not None:
+        for i in state_dict:
+            if isinstance(state_dict[i], torch.Tensor):
+                state_dict[i] = state_dict[i].to(torch_dtype)
+    return state_dict
+
+
+def search_for_embeddings(state_dict):
+    embeddings = []
+    for k in state_dict:
+        if isinstance(state_dict[k], torch.Tensor):
+            embeddings.append(state_dict[k])
+        elif isinstance(state_dict[k], dict):
+            embeddings += search_for_embeddings(state_dict[k])
+    return embeddings
+
+
+def search_parameter(param, state_dict):
+    for name, param_ in state_dict.items():
+        if param.numel() == param_.numel():
+            if param.shape == param_.shape:
+                if torch.dist(param, param_) < 1e-3:
+                    return name
+            else:
+                if torch.dist(param.flatten(), param_.flatten()) < 1e-3:
+                    return name
+    return None
+
+
+def build_rename_dict(source_state_dict, target_state_dict, split_qkv=False):
+    matched_keys = set()
+    with torch.no_grad():
+        for name in source_state_dict:
+            rename = search_parameter(source_state_dict[name], target_state_dict)
+            if rename is not None:
+                print(f'"{name}": "{rename}",')
+                matched_keys.add(rename)
+            elif split_qkv and len(source_state_dict[name].shape)>=1 and source_state_dict[name].shape[0]%3==0:
+                length = source_state_dict[name].shape[0] // 3
+                rename = []
+                for i in range(3):
+                    rename.append(search_parameter(source_state_dict[name][i*length: i*length+length], target_state_dict))
+                if None not in rename:
+                    print(f'"{name}": {rename},')
+                    for rename_ in rename:
+                        matched_keys.add(rename_)
+    for name in target_state_dict:
+        if name not in matched_keys:
+            print("Cannot find", name, target_state_dict[name].shape)
+
+
+def search_for_files(folder, extensions):
+    files = []
+    if os.path.isdir(folder):
+        for file in sorted(os.listdir(folder)):
+            files += search_for_files(os.path.join(folder, file), extensions)
+    elif os.path.isfile(folder):
+        for extension in extensions:
+            if folder.endswith(extension):
+                files.append(folder)
+                break
+    return files
+
+
+def convert_state_dict_keys_to_single_str(state_dict, with_shape=True):
+    keys = []
+    for key, value in state_dict.items():
+        if isinstance(key, str):
+            if isinstance(value, torch.Tensor):
+                if with_shape:
+                    shape = "_".join(map(str, list(value.shape)))
+                    keys.append(key + ":" + shape)
+                keys.append(key)
+            elif isinstance(value, dict):
+                keys.append(key + "|" + convert_state_dict_keys_to_single_str(value, with_shape=with_shape))
+    keys.sort()
+    keys_str = ",".join(keys)
+    return keys_str
+
+
+def split_state_dict_with_prefix(state_dict):
+    keys = sorted([key for key in state_dict if isinstance(key, str)])
+    prefix_dict = {}
+    for key in  keys:
+        prefix = key if "." not in key else key.split(".")[0]
+        if prefix not in prefix_dict:
+            prefix_dict[prefix] = []
+        prefix_dict[prefix].append(key)
+    state_dicts = []
+    for prefix, keys in prefix_dict.items():
+        sub_state_dict = {key: state_dict[key] for key in keys}
+        state_dicts.append(sub_state_dict)
+    return state_dicts
+
+
+def hash_state_dict_keys(state_dict, with_shape=True):
+    keys_str = convert_state_dict_keys_to_single_str(state_dict, with_shape=with_shape)
+    keys_str = keys_str.encode(encoding="UTF-8")
+    return hashlib.md5(keys_str).hexdigest()  

diffsynth/models/wan_video_camera_controller.py

@@ -0,0 +1,221 @@
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import rearrange
+import os
+from typing_extensions import Literal
+
+
+class SimpleAdapter(nn.Module):
+    def __init__(self, in_dim, out_dim, kernel_size, stride, num_residual_blocks=1):
+        super(SimpleAdapter, self).__init__()
+
+        # Pixel Unshuffle: reduce spatial dimensions by a factor of 8
+        self.pixel_unshuffle = nn.PixelUnshuffle(downscale_factor=8)
+
+        # Convolution: reduce spatial dimensions by a factor
+        #  of 2 (without overlap)
+        self.conv = nn.Conv2d(in_dim * 64, out_dim,
+                              kernel_size=kernel_size, stride=stride, padding=0)
+
+        # Residual blocks for feature extraction
+        self.residual_blocks = nn.Sequential(
+            *[ResidualBlock(out_dim) for _ in range(num_residual_blocks)]
+        )
+
+    def forward(self, x):
+        # Reshape to merge the frame dimension into batch
+        bs, c, f, h, w = x.size()
+        x = x.permute(0, 2, 1, 3, 4).contiguous().view(bs * f, c, h, w)
+
+        # Pixel Unshuffle operation
+        x_unshuffled = self.pixel_unshuffle(x)
+
+        # Convolution operation
+        x_conv = self.conv(x_unshuffled)
+
+        # Feature extraction with residual blocks
+        out = self.residual_blocks(x_conv)
+
+        # Reshape to restore original bf dimension
+        out = out.view(bs, f, out.size(1), out.size(2), out.size(3))
+
+        # Permute dimensions to reorder (if needed), e.g., swap channels and feature frames
+        out = out.permute(0, 2, 1, 3, 4)
+
+        return out
+
+    def process_camera_coordinates(
+        self,
+        direction: Literal["Left", "Right", "Up", "Down", "LeftUp", "LeftDown", "RightUp", "RightDown"],
+        length: int,
+        height: int,
+        width: int,
+        speed: float = 1/54,
+        origin=(0, 0.532139961, 0.946026558, 0.5, 0.5,
+                0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0)
+    ):
+        if origin is None:
+            origin = (0, 0.532139961, 0.946026558, 0.5, 0.5,
+                      0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0)
+        print(
+            f"Generating camera coordinates with direction: {direction}, length: {length}, speed: {speed}, origin: {origin}")
+        coordinates = generate_camera_coordinates(
+            direction, length, speed, origin)
+        print(f"Generated {len(coordinates)} camera coordinates.")
+        plucker_embedding = process_pose_file(coordinates, width, height)
+        print(
+            f"Processed camera coordinates into plucker embedding with shape: {plucker_embedding.shape}")
+        return plucker_embedding
+
+
+class ResidualBlock(nn.Module):
+    def __init__(self, dim):
+        super(ResidualBlock, self).__init__()
+        self.conv1 = nn.Conv2d(dim, dim, kernel_size=3, padding=1)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = nn.Conv2d(dim, dim, kernel_size=3, padding=1)
+
+    def forward(self, x):
+        residual = x
+        out = self.relu(self.conv1(x))
+        out = self.conv2(out)
+        out += residual
+        return out
+
+
+class Camera(object):
+    """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py
+    """
+
+    def __init__(self, entry):
+        fx, fy, cx, cy = entry[1:5]
+        self.fx = fx
+        self.fy = fy
+        self.cx = cx
+        self.cy = cy
+        w2c_mat = np.array(entry[7:]).reshape(3, 4)
+        w2c_mat_4x4 = np.eye(4)
+        w2c_mat_4x4[:3, :] = w2c_mat
+        self.w2c_mat = w2c_mat_4x4
+        self.c2w_mat = np.linalg.inv(w2c_mat_4x4)
+
+
+def get_relative_pose(cam_params):
+    """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py
+    """
+    abs_w2cs = [cam_param.w2c_mat for cam_param in cam_params]
+    abs_c2ws = [cam_param.c2w_mat for cam_param in cam_params]
+    cam_to_origin = 0
+    target_cam_c2w = np.array([
+        [1, 0, 0, 0],
+        [0, 1, 0, -cam_to_origin],
+        [0, 0, 1, 0],
+        [0, 0, 0, 1]
+    ])
+    abs2rel = target_cam_c2w @ abs_w2cs[0]
+    ret_poses = [target_cam_c2w, ] + \
+        [abs2rel @ abs_c2w for abs_c2w in abs_c2ws[1:]]
+    ret_poses = np.array(ret_poses, dtype=np.float32)
+    return ret_poses
+
+
+def custom_meshgrid(*args):
+    # torch>=2.0.0 only
+    return torch.meshgrid(*args, indexing='ij')
+
+
+def ray_condition(K, c2w, H, W, device):
+    """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py
+    """
+    # c2w: B, V, 4, 4
+    # K: B, V, 4
+
+    B = K.shape[0]
+
+    j, i = custom_meshgrid(
+        torch.linspace(0, H - 1, H, device=device, dtype=c2w.dtype),
+        torch.linspace(0, W - 1, W, device=device, dtype=c2w.dtype),
+    )
+    i = i.reshape([1, 1, H * W]).expand([B, 1, H * W]) + 0.5  # [B, HxW]
+    j = j.reshape([1, 1, H * W]).expand([B, 1, H * W]) + 0.5  # [B, HxW]
+
+    fx, fy, cx, cy = K.chunk(4, dim=-1)  # B,V, 1
+
+    zs = torch.ones_like(i)  # [B, HxW]
+    xs = (i - cx) / fx * zs
+    ys = (j - cy) / fy * zs
+    zs = zs.expand_as(ys)
+
+    directions = torch.stack((xs, ys, zs), dim=-1)  # B, V, HW, 3
+    directions = directions / \
+        directions.norm(dim=-1, keepdim=True)  # B, V, HW, 3
+
+    rays_d = directions @ c2w[..., :3, :3].transpose(-1, -2)  # B, V, 3, HW
+    rays_o = c2w[..., :3, 3]  # B, V, 3
+    rays_o = rays_o[:, :, None].expand_as(rays_d)  # B, V, 3, HW
+    # c2w @ dirctions
+    rays_dxo = torch.linalg.cross(rays_o, rays_d)
+    plucker = torch.cat([rays_dxo, rays_d], dim=-1)
+    plucker = plucker.reshape(B, c2w.shape[1], H, W, 6)  # B, V, H, W, 6
+    # plucker = plucker.permute(0, 1, 4, 2, 3)
+    return plucker
+
+
+def process_pose_file(cam_params, width=672, height=384, original_pose_width=1280, original_pose_height=720, device='cpu', return_poses=False):
+    if return_poses:
+        return cam_params
+    else:
+        cam_params = [Camera(cam_param) for cam_param in cam_params]
+
+        sample_wh_ratio = width / height
+        # Assuming placeholder ratios, change as needed
+        pose_wh_ratio = original_pose_width / original_pose_height
+
+        if pose_wh_ratio > sample_wh_ratio:
+            resized_ori_w = height * pose_wh_ratio
+            for cam_param in cam_params:
+                cam_param.fx = resized_ori_w * cam_param.fx / width
+        else:
+            resized_ori_h = width / pose_wh_ratio
+            for cam_param in cam_params:
+                cam_param.fy = resized_ori_h * cam_param.fy / height
+
+        intrinsic = np.asarray([[cam_param.fx * width,
+                                cam_param.fy * height,
+                                cam_param.cx * width,
+                                cam_param.cy * height]
+                                for cam_param in cam_params], dtype=np.float32)
+
+        K = torch.as_tensor(intrinsic)[None]  # [1, 1, 4]
+        # Assuming this function is defined elsewhere
+        c2ws = get_relative_pose(cam_params)
+        c2ws = torch.as_tensor(c2ws)[None]  # [1, n_frame, 4, 4]
+        plucker_embedding = ray_condition(K, c2ws, height, width, device=device)[
+            0].permute(0, 3, 1, 2).contiguous()  # V, 6, H, W
+        plucker_embedding = plucker_embedding[None]
+        plucker_embedding = rearrange(
+            plucker_embedding, "b f c h w -> b f h w c")[0]
+        return plucker_embedding
+
+
+def generate_camera_coordinates(
+    direction: Literal["Left", "Right", "Up", "Down", "LeftUp", "LeftDown", "RightUp", "RightDown"],
+    length: int,
+    speed: float = 1/54,
+    origin=(0, 0.532139961, 0.946026558, 0.5, 0.5,
+            0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0)
+):
+    coordinates = [list(origin)]
+    while len(coordinates) < length:
+        coor = coordinates[-1].copy()
+        if "Left" in direction:
+            coor[9] += speed
+        if "Right" in direction:
+            coor[9] -= speed
+        if "Up" in direction:
+            coor[13] += speed
+        if "Down" in direction:
+            coor[13] -= speed
+        coordinates.append(coor)
+    return coordinates

diffsynth/models/wan_video_dit.py

@@ -0,0 +1,974 @@
+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from diffusers.models.lora import LoRALinearLayer
+from einops import rearrange
+
+from .utils import hash_state_dict_keys
+from .wan_video_camera_controller import SimpleAdapter
+
+try:
+    import flash_attn_interface
+
+    FLASH_ATTN_3_AVAILABLE = True
+except ModuleNotFoundError:
+    FLASH_ATTN_3_AVAILABLE = False
+
+try:
+    import flash_attn
+
+    FLASH_ATTN_2_AVAILABLE = True
+except ModuleNotFoundError:
+    FLASH_ATTN_2_AVAILABLE = False
+
+try:
+    from sageattention import sageattn
+
+    SAGE_ATTN_AVAILABLE = True
+    print(f"========= Using sage attention, please note that this is for inference speed up only!==========")
+except ModuleNotFoundError:
+    SAGE_ATTN_AVAILABLE = False
+
+
+def flash_attention(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    num_heads: int,
+    compatibility_mode=False,
+):
+    if compatibility_mode:
+        q = rearrange(q, "b s (n d) -> b n s d", n=num_heads)
+        k = rearrange(k, "b s (n d) -> b n s d", n=num_heads)
+        v = rearrange(v, "b s (n d) -> b n s d", n=num_heads)
+        x = F.scaled_dot_product_attention(q, k, v)
+        x = rearrange(x, "b n s d -> b s (n d)", n=num_heads)
+    elif FLASH_ATTN_3_AVAILABLE:
+        q = rearrange(q, "b s (n d) -> b s n d", n=num_heads)
+        k = rearrange(k, "b s (n d) -> b s n d", n=num_heads)
+        v = rearrange(v, "b s (n d) -> b s n d", n=num_heads)
+        x = flash_attn_interface.flash_attn_func(q, k, v)
+        if isinstance(x, tuple):
+            x = x[0]
+        x = rearrange(x, "b s n d -> b s (n d)", n=num_heads)
+    elif FLASH_ATTN_2_AVAILABLE:
+        q = rearrange(q, "b s (n d) -> b s n d", n=num_heads)
+        k = rearrange(k, "b s (n d) -> b s n d", n=num_heads)
+        v = rearrange(v, "b s (n d) -> b s n d", n=num_heads)
+        x = flash_attn.flash_attn_func(q, k, v)
+        x = rearrange(x, "b s n d -> b s (n d)", n=num_heads)
+    elif SAGE_ATTN_AVAILABLE:
+        q = rearrange(q, "b s (n d) -> b n s d", n=num_heads)
+        k = rearrange(k, "b s (n d) -> b n s d", n=num_heads)
+        v = rearrange(v, "b s (n d) -> b n s d", n=num_heads)
+        x = sageattn(q, k, v)
+        x = rearrange(x, "b n s d -> b s (n d)", n=num_heads)
+    else:
+        q = rearrange(q, "b s (n d) -> b n s d", n=num_heads)
+        k = rearrange(k, "b s (n d) -> b n s d", n=num_heads)
+        v = rearrange(v, "b s (n d) -> b n s d", n=num_heads)
+        x = F.scaled_dot_product_attention(q, k, v)
+        x = rearrange(x, "b n s d -> b s (n d)", n=num_heads)
+    return x
+
+
+def modulate(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor):
+    return x * (1 + scale) + shift
+
+
+def sinusoidal_embedding_1d(dim, position):
+    sinusoid = torch.outer(
+        position.type(torch.float64),
+        torch.pow(
+            10000,
+            -torch.arange(dim // 2, dtype=torch.float64, device=position.device).div(
+                dim // 2
+            ),
+        ),
+    )
+    x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
+    return x.to(position.dtype)
+
+
+def precompute_freqs_cis_3d(dim: int, end: int = 1024, theta: float = 10000.0):
+    # 3d rope precompute
+    f_freqs_cis = precompute_freqs_cis(dim - 2 * (dim // 3), end, theta)
+    h_freqs_cis = precompute_freqs_cis(dim // 3, end, theta)
+    w_freqs_cis = precompute_freqs_cis(dim // 3, end, theta)
+    return f_freqs_cis, h_freqs_cis, w_freqs_cis
+
+
+def precompute_freqs_cis(dim: int, end: int = 1024, theta: float = 10000.0):
+    # 1d rope precompute
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)
+                   [: (dim // 2)].double() / dim))
+    freqs = torch.outer(torch.arange(end, device=freqs.device), freqs)
+    freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64
+    return freqs_cis
+
+
+def rope_apply(x, freqs, num_heads):
+    x = rearrange(x, "b s (n d) -> b s n d", n=num_heads)
+    x_out = torch.view_as_complex(
+        x.to(torch.float64).reshape(x.shape[0], x.shape[1], x.shape[2], -1, 2)
+    )
+    x_out = torch.view_as_real(x_out * freqs).flatten(2)
+    return x_out.to(x.dtype)
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim, eps=1e-5):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+
+    def norm(self, x):
+        return x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps)
+
+    def forward(self, x):
+        dtype = x.dtype
+        # print(f"x device: {x.device}, weight device: {self.weight.device}")
+        return self.norm(x.float()).to(dtype) * self.weight
+
+
+class AttentionModule(nn.Module):
+    def __init__(self, num_heads):
+        super().__init__()
+        self.num_heads = num_heads
+
+    def forward(self, q, k, v):
+        x = flash_attention(q=q, k=k, v=v, num_heads=self.num_heads)
+        return x
+
+
+class SelfAttention(nn.Module):
+    def __init__(self, dim: int, num_heads: int, eps: float = 1e-6):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.q = nn.Linear(dim, dim)
+        self.k = nn.Linear(dim, dim)
+        self.v = nn.Linear(dim, dim)
+        self.o = nn.Linear(dim, dim)
+        self.norm_q = RMSNorm(dim, eps=eps)
+        self.norm_k = RMSNorm(dim, eps=eps)
+
+        self.attn = AttentionModule(self.num_heads)
+
+    def forward(self, x, freqs):
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(x))
+        v = self.v(x)
+        q = rope_apply(q, freqs, self.num_heads)
+        k = rope_apply(k, freqs, self.num_heads)
+        x = self.attn(q, k, v)
+        return self.o(x)
+
+
+class CrossAttention(nn.Module):
+    def __init__(
+        self, dim: int, num_heads: int, eps: float = 1e-6, has_image_input: bool = False
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.q = nn.Linear(dim, dim)
+        self.k = nn.Linear(dim, dim)
+        self.v = nn.Linear(dim, dim)
+        self.o = nn.Linear(dim, dim)
+        self.norm_q = RMSNorm(dim, eps=eps)
+        self.norm_k = RMSNorm(dim, eps=eps)
+        self.has_image_input = has_image_input
+        if has_image_input:
+            self.k_img = nn.Linear(dim, dim)
+            self.v_img = nn.Linear(dim, dim)
+            self.norm_k_img = RMSNorm(dim, eps=eps)
+
+        self.attn = AttentionModule(self.num_heads)
+
+    def forward(self, x: torch.Tensor, y: torch.Tensor):
+        if self.has_image_input:
+            img = y[:, :257]
+            ctx = y[:, 257:]
+        else:
+            ctx = y
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(ctx))
+        v = self.v(ctx)
+        x = self.attn(q, k, v)
+        if self.has_image_input:
+            k_img = self.norm_k_img(self.k_img(img))
+            v_img = self.v_img(img)
+            y = flash_attention(q, k_img, v_img, num_heads=self.num_heads)
+            x = x + y
+        return self.o(x)
+
+
+class SelfAttentionSeparate(nn.Module):
+    def __init__(self, dim: int, num_heads: int, eps: float = 1e-6, rank=64):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.q = nn.Linear(dim, dim)
+        self.k = nn.Linear(dim, dim)
+        self.v = nn.Linear(dim, dim)
+        self.o = nn.Linear(dim, dim)
+        if rank > 0:
+            # LoraLinear
+            self.q_zl_before = LoRALinearLayer(dim, dim, rank=rank)
+            self.k_zl_before = LoRALinearLayer(dim, dim, rank=rank)
+            self.v_zl_before = LoRALinearLayer(dim, dim, rank=rank)
+
+            self.q_zl_after = LoRALinearLayer(dim, dim, rank=rank)
+            self.k_zl_after = LoRALinearLayer(dim, dim, rank=rank)
+            self.v_zl_after = LoRALinearLayer(dim, dim, rank=rank)
+        else:
+            # Normal Linear
+            self.q_zl_before = nn.Linear(dim, dim)
+            self.k_zl_before = nn.Linear(dim, dim)
+            self.v_zl_before = nn.Linear(dim, dim)
+
+            self.q_zl_after = nn.Linear(dim, dim)
+            self.k_zl_after = nn.Linear(dim, dim)
+            self.v_zl_after = nn.Linear(dim, dim)
+
+        self.norm_q = RMSNorm(dim, eps=eps)
+        self.norm_k = RMSNorm(dim, eps=eps)
+
+        self.attn = AttentionModule(self.num_heads)
+        self.zero_init_linear()
+
+    def zero_init_linear(self):
+        layers_to_handle = [
+            self.q_zl_before,
+            self.k_zl_before,
+            self.v_zl_before,
+            self.q_zl_after,
+            self.k_zl_after,
+            self.v_zl_after,
+        ]
+        for _layer in layers_to_handle:
+            if isinstance(_layer, nn.Linear):
+                nn.init.zeros_(_layer.weight)
+                if _layer.bias is not None:
+                    nn.init.zeros_(_layer.bias)
+
+    def forward(self, x, freqs, camera_pose_embedding=None):
+        if camera_pose_embedding is not None:
+
+            q = self.norm_q(self.q(x))
+            k = self.norm_k(self.k(x))
+            v = self.v(x)
+        # TODO uncomment
+        # ----------------------------------------------------------------
+        else:
+            q = self.norm_q(
+                self.q(x + self.q_zl_before(camera_pose_embedding))
+                + self.q_zl_after(camera_pose_embedding)
+            )
+            k = self.norm_k(
+                self.k(x + self.k_zl_before(camera_pose_embedding))
+                + self.k_zl_after(camera_pose_embedding)
+            )
+            v = self.v(x + self.v_zl_before(camera_pose_embedding)) + self.v_zl_after(
+                camera_pose_embedding
+            )
+        # --------------------------------------------------------------------
+
+        q = rope_apply(q, freqs, self.num_heads)
+        k = rope_apply(k, freqs, self.num_heads)
+        x = self.attn(q, k, v)
+        return self.o(x)
+
+
+class CrossAttentionSeparate(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        eps: float = 1e-6,
+        has_image_input: bool = False,
+        rank=64,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.q = nn.Linear(dim, dim)
+        self.k = nn.Linear(dim, dim)
+        self.v = nn.Linear(dim, dim)
+        self.o = nn.Linear(dim, dim)
+
+        if rank > 0:
+            # LoraLinear
+            self.q_zl_before = LoRALinearLayer(dim, dim, rank=rank)
+            self.q_zl_after = LoRALinearLayer(dim, dim, rank=rank)
+        else:
+            # Normal linear
+            self.q_zl_before = nn.Linear(dim, dim, bias=False)
+            self.q_zl_after = nn.Linear(dim, dim, bias=False)
+
+        self.norm_q = RMSNorm(dim, eps=eps)
+        self.norm_k = RMSNorm(dim, eps=eps)
+        self.has_image_input = has_image_input
+        if has_image_input:
+            self.k_img = nn.Linear(dim, dim)
+            self.v_img = nn.Linear(dim, dim)
+            self.norm_k_img = RMSNorm(dim, eps=eps)
+
+        self.attn = AttentionModule(self.num_heads)
+        self.zero_init_linear()
+
+    def zero_init_linear(self):
+        layers_to_handle = [
+            self.q_zl_before,
+            self.q_zl_after,
+        ]
+        for _layer in layers_to_handle:
+            if isinstance(_layer, nn.Linear):
+                nn.init.zeros_(_layer.weight)
+                if _layer.bias is not None:
+                    nn.init.zeros_(_layer.bias)
+
+    def forward(self, x: torch.Tensor, y: torch.Tensor, camera_pose_embedding=None):
+        if self.has_image_input:
+            img = y[:, :257]
+            ctx = y[:, 257:]
+        else:
+            ctx = y
+        # q = self.norm_q(self.q(x))
+        # k = self.norm_k(self.k(ctx))
+        # v = self.v(ctx)
+        # TODO uncomment
+        # -------------------------------------------------
+        q = self.norm_q(
+            self.q(x + self.q_zl_before(camera_pose_embedding))
+            + self.q_zl_after(camera_pose_embedding)
+        )
+        k = self.norm_k(self.k(ctx))
+        v = self.v(ctx)
+        # -------------------------------------------------
+
+        x = self.attn(q, k, v)
+        if self.has_image_input:
+            k_img = self.norm_k_img(self.k_img(img))
+            v_img = self.v_img(img)
+            y = flash_attention(q, k_img, v_img, num_heads=self.num_heads)
+            x = x + y
+        return self.o(x)
+
+
+class GateModule(nn.Module):
+    def __init__(
+        self,
+    ):
+        super().__init__()
+
+    def forward(self, x, gate, residual):
+        return x + gate * residual
+
+
+class DiTBlock(nn.Module):
+    def __init__(
+        self,
+        has_image_input: bool,
+        dim: int,
+        num_heads: int,
+        ffn_dim: int,
+        eps: float = 1e-6,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.ffn_dim = ffn_dim
+
+        self.self_attn = SelfAttention(dim, num_heads, eps)
+        self.cross_attn = CrossAttention(
+            dim, num_heads, eps, has_image_input=has_image_input
+        )
+        self.norm1 = nn.LayerNorm(dim, eps=eps, elementwise_affine=False)
+        self.norm2 = nn.LayerNorm(dim, eps=eps, elementwise_affine=False)
+        self.norm3 = nn.LayerNorm(dim, eps=eps)
+        self.ffn = nn.Sequential(
+            nn.Linear(dim, ffn_dim),
+            nn.GELU(approximate="tanh"),
+            nn.Linear(ffn_dim, dim),
+        )
+        self.modulation = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
+        self.gate = GateModule()
+
+    # @torch.compile(mode='max-autotune')
+    def forward(self, x, context, t_mod, freqs, camera_pose_embedding=None):
+        # msa: multi-head self-attention  mlp: multi-layer perceptron
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+            self.modulation.to(dtype=t_mod.dtype, device=t_mod.device) + t_mod
+        ).chunk(6, dim=1)
+        input_x = modulate(self.norm1(x), shift_msa, scale_msa)
+        x = self.gate(x, gate_msa, self.self_attn(input_x, freqs))
+        x = x + self.cross_attn(self.norm3(x), context)
+        input_x = modulate(self.norm2(x), shift_mlp, scale_mlp)
+        x = self.gate(x, gate_mlp, self.ffn(input_x))
+        return x
+
+
+class CameraDiTBlock(nn.Module):
+    def __init__(
+        self,
+        has_image_input: bool,
+        dim: int,
+        num_heads: int,
+        ffn_dim: int,
+        eps: float = 1e-6,
+        camera_lora_rank=64,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.ffn_dim = ffn_dim
+
+        self.self_attn = SelfAttentionSeparate(
+            dim, num_heads, eps, rank=camera_lora_rank
+        )
+        self.cross_attn = CrossAttentionSeparate(
+            dim, num_heads, eps, has_image_input=has_image_input, rank=camera_lora_rank
+        )
+        self.norm1 = nn.LayerNorm(dim, eps=eps, elementwise_affine=False)
+        self.norm2 = nn.LayerNorm(dim, eps=eps, elementwise_affine=False)
+        self.norm3 = nn.LayerNorm(dim, eps=eps)
+        self.ffn = nn.Sequential(
+            nn.Linear(dim, ffn_dim),
+            nn.GELU(approximate="tanh"),
+            nn.Linear(ffn_dim, dim),
+        )
+        self.modulation = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
+        self.gate = GateModule()
+
+    def forward(self, x, context, t_mod, freqs, camera_pose_embedding=None):
+        # msa: multi-head self-attention  mlp: multi-layer perceptron
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+            self.modulation.to(dtype=t_mod.dtype, device=t_mod.device) + t_mod
+        ).chunk(6, dim=1)
+        input_x = modulate(self.norm1(x), shift_msa, scale_msa)
+        # x = self.gate(x, gate_msa, self.self_attn(input_x, freqs))
+        x = self.gate(
+            x, gate_msa, self.self_attn(input_x, freqs, camera_pose_embedding)
+        )
+
+        x = x + self.cross_attn(self.norm3(x), context, camera_pose_embedding)
+        input_x = modulate(self.norm2(x), shift_mlp, scale_mlp)
+        x = self.gate(x, gate_mlp, self.ffn(input_x))
+        return x
+
+
+class MLP(torch.nn.Module):
+    def __init__(self, in_dim, out_dim, has_pos_emb=False):
+        super().__init__()
+        self.proj = torch.nn.Sequential(
+            nn.LayerNorm(in_dim),
+            nn.Linear(in_dim, in_dim),
+            nn.GELU(),
+            nn.Linear(in_dim, out_dim),
+            nn.LayerNorm(out_dim),
+        )
+        self.has_pos_emb = has_pos_emb
+        if has_pos_emb:
+            self.emb_pos = torch.nn.Parameter(torch.zeros((1, 514, 1280)))
+
+    def forward(self, x):
+        if self.has_pos_emb:
+            x = x + self.emb_pos.to(dtype=x.dtype, device=x.device)
+        return self.proj(x)
+
+
+class Head(nn.Module):
+    def __init__(
+        self, dim: int, out_dim: int, patch_size: Tuple[int, int, int], eps: float
+    ):
+        super().__init__()
+        self.dim = dim
+        self.patch_size = patch_size
+        self.norm = nn.LayerNorm(dim, eps=eps, elementwise_affine=False)
+        self.head = nn.Linear(dim, out_dim * math.prod(patch_size))
+        self.modulation = nn.Parameter(torch.randn(1, 2, dim) / dim**0.5)
+
+    def forward(self, x, t_mod):
+        shift, scale = (
+            self.modulation.to(dtype=t_mod.dtype, device=t_mod.device) + t_mod
+        ).chunk(2, dim=1)
+        x = self.head(self.norm(x) * (1 + scale) + shift)
+        return x
+
+
+class WanModel(torch.nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        in_dim: int,
+        ffn_dim: int,
+        out_dim: int,
+        text_dim: int,
+        freq_dim: int,
+        eps: float,
+        patch_size: Tuple[int, int, int],
+        num_heads: int,
+        num_layers: int,
+        has_image_input: bool,
+        has_image_pos_emb: bool = False,
+        has_ref_conv: bool = False,
+        add_control_adapter: bool = False,
+        in_dim_control_adapter: int = 24,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.freq_dim = freq_dim
+        self.has_image_input = has_image_input
+        self.patch_size = patch_size
+
+        self.patch_embedding = nn.Conv3d(
+            in_dim, dim, kernel_size=patch_size, stride=patch_size
+        )
+        self.text_embedding = nn.Sequential(
+            nn.Linear(text_dim, dim), nn.GELU(
+                approximate="tanh"), nn.Linear(dim, dim)
+        )
+        self.time_embedding = nn.Sequential(
+            nn.Linear(freq_dim, dim), nn.SiLU(), nn.Linear(dim, dim)
+        )
+        self.time_projection = nn.Sequential(
+            nn.SiLU(), nn.Linear(dim, dim * 6))
+        self.blocks = nn.ModuleList(
+            [
+                DiTBlock(has_image_input, dim, num_heads, ffn_dim, eps)
+                for _ in range(num_layers)
+            ]
+        )
+        self.head = Head(dim, out_dim, patch_size, eps)
+        head_dim = dim // num_heads
+        self.freqs = precompute_freqs_cis_3d(head_dim)
+
+        if has_image_input:
+            self.img_emb = MLP(
+                1280, dim, has_pos_emb=has_image_pos_emb
+            )  # clip_feature_dim = 1280
+        if has_ref_conv:
+            self.ref_conv = nn.Conv2d(
+                16, dim, kernel_size=(2, 2), stride=(2, 2))
+        self.has_image_pos_emb = has_image_pos_emb
+        self.has_ref_conv = has_ref_conv
+
+        self.control_adapter = None
+        self.add_control_adapter = add_control_adapter
+        if add_control_adapter:
+            self.control_adapter = SimpleAdapter(
+                in_dim_control_adapter,
+                dim,
+                kernel_size=patch_size[1:],
+                stride=patch_size[1:],
+            )
+        else:
+            self.control_adapter = None
+
+    def patchify(
+        self, x: torch.Tensor, control_camera_latents_input: torch.Tensor = None
+    ):
+        x = self.patch_embedding(x)
+        if (
+            self.control_adapter is not None
+            and control_camera_latents_input is not None
+        ):
+            y_camera = self.control_adapter(control_camera_latents_input)
+            x = [u + v for u, v in zip(x, y_camera)]
+            # x = x[0].unsqueeze(0)
+        grid_size = x.shape[2:]
+        x = rearrange(x, "b c f h w -> b (f h w) c").contiguous()
+        return x, grid_size  # x, grid_size: (f, h, w)
+
+    def unpatchify(self, x: torch.Tensor, grid_size: torch.Tensor):
+        return rearrange(
+            x,
+            "b (f h w) (x y z c) -> b c (f x) (h y) (w z)",
+            f=grid_size[0],
+            h=grid_size[1],
+            w=grid_size[2],
+            x=self.patch_size[0],
+            y=self.patch_size[1],
+            z=self.patch_size[2],
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        timestep: torch.Tensor,
+        context: torch.Tensor,
+        clip_feature: Optional[torch.Tensor] = None,
+        y: Optional[torch.Tensor] = None,
+        use_gradient_checkpointing: bool = False,
+        use_gradient_checkpointing_offload: bool = False,
+        **kwargs,
+    ):
+        t = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, timestep))
+        t_mod = self.time_projection(t).unflatten(1, (6, self.dim))
+        context = self.text_embedding(context)
+
+        if self.has_image_input:
+            # print(f"x,y shape: {x.shape}, {y.shape if y is not None else 'None'}")
+            x = torch.cat([x, y], dim=1)  # (b, c_x + c_y, f, h, w)
+            clip_embdding = self.img_emb(clip_feature)
+            context = torch.cat([clip_embdding, context], dim=1)
+
+        x, (f, h, w) = self.patchify(x)
+
+        freqs = (
+            torch.cat(
+                [
+                    self.freqs[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1),
+                    self.freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1),
+                    self.freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1),
+                ],
+                dim=-1,
+            )
+            .reshape(f * h * w, 1, -1)
+            .to(x.device)
+        )
+
+        def create_custom_forward(module):
+            def custom_forward(*inputs):
+                return module(*inputs)
+
+            return custom_forward
+
+        for block in self.blocks:
+            if self.training and use_gradient_checkpointing:
+                if use_gradient_checkpointing_offload:
+                    with torch.autograd.graph.save_on_cpu():
+                        x = torch.utils.checkpoint.checkpoint(
+                            create_custom_forward(block),
+                            x,
+                            context,
+                            t_mod,
+                            freqs,
+                            use_reentrant=False,
+                        )
+                else:
+                    x = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(block),
+                        x,
+                        context,
+                        t_mod,
+                        freqs,
+                        use_reentrant=False,
+                    )
+            else:
+                x = block(x, context, t_mod, freqs)
+
+        x = self.head(x, t)
+        x = self.unpatchify(x, (f, h, w))
+
+        if hasattr(self, 'rgb_head'):
+            rgb = self.rgb_head(x, t)
+            rgb = self.unpatchify(rgb, (f, h, w))
+
+            return x, rgb
+
+        return x
+
+    @staticmethod
+    def state_dict_converter():
+        return WanModelStateDictConverter()
+
+
+class WanModelStateDictConverter:
+    def __init__(self):
+        pass
+
+    def from_diffusers(self, state_dict):
+        rename_dict = {
+            "blocks.0.attn1.norm_k.weight": "blocks.0.self_attn.norm_k.weight",
+            "blocks.0.attn1.norm_q.weight": "blocks.0.self_attn.norm_q.weight",
+            "blocks.0.attn1.to_k.bias": "blocks.0.self_attn.k.bias",
+            "blocks.0.attn1.to_k.weight": "blocks.0.self_attn.k.weight",
+            "blocks.0.attn1.to_out.0.bias": "blocks.0.self_attn.o.bias",
+            "blocks.0.attn1.to_out.0.weight": "blocks.0.self_attn.o.weight",
+            "blocks.0.attn1.to_q.bias": "blocks.0.self_attn.q.bias",
+            "blocks.0.attn1.to_q.weight": "blocks.0.self_attn.q.weight",
+            "blocks.0.attn1.to_v.bias": "blocks.0.self_attn.v.bias",
+            "blocks.0.attn1.to_v.weight": "blocks.0.self_attn.v.weight",
+            "blocks.0.attn2.norm_k.weight": "blocks.0.cross_attn.norm_k.weight",
+            "blocks.0.attn2.norm_q.weight": "blocks.0.cross_attn.norm_q.weight",
+            "blocks.0.attn2.to_k.bias": "blocks.0.cross_attn.k.bias",
+            "blocks.0.attn2.to_k.weight": "blocks.0.cross_attn.k.weight",
+            "blocks.0.attn2.to_out.0.bias": "blocks.0.cross_attn.o.bias",
+            "blocks.0.attn2.to_out.0.weight": "blocks.0.cross_attn.o.weight",
+            "blocks.0.attn2.to_q.bias": "blocks.0.cross_attn.q.bias",
+            "blocks.0.attn2.to_q.weight": "blocks.0.cross_attn.q.weight",
+            "blocks.0.attn2.to_v.bias": "blocks.0.cross_attn.v.bias",
+            "blocks.0.attn2.to_v.weight": "blocks.0.cross_attn.v.weight",
+            "blocks.0.ffn.net.0.proj.bias": "blocks.0.ffn.0.bias",
+            "blocks.0.ffn.net.0.proj.weight": "blocks.0.ffn.0.weight",
+            "blocks.0.ffn.net.2.bias": "blocks.0.ffn.2.bias",
+            "blocks.0.ffn.net.2.weight": "blocks.0.ffn.2.weight",
+            "blocks.0.norm2.bias": "blocks.0.norm3.bias",
+            "blocks.0.norm2.weight": "blocks.0.norm3.weight",
+            "blocks.0.scale_shift_table": "blocks.0.modulation",
+            "condition_embedder.text_embedder.linear_1.bias": "text_embedding.0.bias",
+            "condition_embedder.text_embedder.linear_1.weight": "text_embedding.0.weight",
+            "condition_embedder.text_embedder.linear_2.bias": "text_embedding.2.bias",
+            "condition_embedder.text_embedder.linear_2.weight": "text_embedding.2.weight",
+            "condition_embedder.time_embedder.linear_1.bias": "time_embedding.0.bias",
+            "condition_embedder.time_embedder.linear_1.weight": "time_embedding.0.weight",
+            "condition_embedder.time_embedder.linear_2.bias": "time_embedding.2.bias",
+            "condition_embedder.time_embedder.linear_2.weight": "time_embedding.2.weight",
+            "condition_embedder.time_proj.bias": "time_projection.1.bias",
+            "condition_embedder.time_proj.weight": "time_projection.1.weight",
+            "patch_embedding.bias": "patch_embedding.bias",
+            "patch_embedding.weight": "patch_embedding.weight",
+            "scale_shift_table": "head.modulation",
+            "proj_out.bias": "head.head.bias",
+            "proj_out.weight": "head.head.weight",
+        }
+        state_dict_ = {}
+        print(
+            f"hash_state_dict_keys(state_dict): {hash_state_dict_keys(state_dict)}")
+        for name, param in state_dict.items():
+            if name in rename_dict:
+                state_dict_[rename_dict[name]] = param
+            else:
+                name_ = ".".join(name.split(
+                    ".")[:1] + ["0"] + name.split(".")[2:])
+                if name_ in rename_dict:
+                    name_ = rename_dict[name_]
+                    name_ = ".".join(
+                        name_.split(".")[:1]
+                        + [name.split(".")[1]]
+                        + name_.split(".")[2:]
+                    )
+                    state_dict_[name_] = param
+        if hash_state_dict_keys(state_dict) == "cb104773c6c2cb6df4f9529ad5c60d0b":
+            config = {
+                "model_type": "t2v",
+                "patch_size": (1, 2, 2),
+                "text_len": 512,
+                "in_dim": 16,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "window_size": (-1, -1),
+                "qk_norm": True,
+                "cross_attn_norm": True,
+                "eps": 1e-6,
+            }
+        else:
+            config = {}
+        return state_dict_, config
+
+    def from_civitai(self, state_dict):
+        state_dict = {
+            name: param
+            for name, param in state_dict.items()
+            if not name.startswith("vace")
+        }
+        print(
+            f"hash_state_dict_keys(state_dict): {hash_state_dict_keys(state_dict)} from civitai"
+        )
+
+        if hash_state_dict_keys(state_dict) == "9269f8db9040a9d860eaca435be61814":
+            config = {
+                "has_image_input": False,
+                "patch_size": [1, 2, 2],
+                "in_dim": 16,
+                "dim": 1536,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 12,
+                "num_layers": 30,
+                "eps": 1e-6,
+            }
+        elif hash_state_dict_keys(state_dict) == "aafcfd9672c3a2456dc46e1cb6e52c70":
+            config = {
+                "has_image_input": False,
+                "patch_size": [1, 2, 2],
+                "in_dim": 16,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6,
+            }
+        elif hash_state_dict_keys(state_dict) == "6bfcfb3b342cb286ce886889d519a77e":
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 36,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6,
+            }
+        elif hash_state_dict_keys(state_dict) == "6d6ccde6845b95ad9114ab993d917893":
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 36,
+                "dim": 1536,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 12,
+                "num_layers": 30,
+                "eps": 1e-6,
+            }
+        elif hash_state_dict_keys(state_dict) == "6bfcfb3b342cb286ce886889d519a77e":
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 36,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6,
+            }
+        elif hash_state_dict_keys(state_dict) == "349723183fc063b2bfc10bb2835cf677":
+            # 1.3B PAI control
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 48,
+                "dim": 1536,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 12,
+                "num_layers": 30,
+                "eps": 1e-6,
+            }
+        elif hash_state_dict_keys(state_dict) == "efa44cddf936c70abd0ea28b6cbe946c":
+            # 14B PAI control
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 48,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6,
+            }
+        elif hash_state_dict_keys(state_dict) == "3ef3b1f8e1dab83d5b71fd7b617f859f":
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 36,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6,
+                "has_image_pos_emb": True,
+            }
+        elif hash_state_dict_keys(state_dict) == "70ddad9d3a133785da5ea371aae09504":
+            # 1.3B PAI control v1.1
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 48,
+                "dim": 1536,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 12,
+                "num_layers": 30,
+                "eps": 1e-6,
+                "has_ref_conv": True,
+            }
+        elif hash_state_dict_keys(state_dict) == "26bde73488a92e64cc20b0a7485b9e5b":
+            # 14B PAI control v1.1
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 48,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6,
+                "has_ref_conv": True,
+            }
+        elif hash_state_dict_keys(state_dict) == "ac6a5aa74f4a0aab6f64eb9a72f19901":
+            # 1.3B PAI control-camera v1.1
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 32,
+                "dim": 1536,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 12,
+                "num_layers": 30,
+                "eps": 1e-6,
+                "has_ref_conv": False,
+                "add_control_adapter": True,
+                "in_dim_control_adapter": 24,
+            }
+        elif hash_state_dict_keys(state_dict) == "b61c605c2adbd23124d152ed28e049ae":
+            # 14B PAI control-camera v1.1
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 32,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6,
+                "has_ref_conv": False,
+                "add_control_adapter": True,
+                "in_dim_control_adapter": 24,
+            }
+        else:
+            config = {}
+        return state_dict, config

diffsynth/models/wan_video_image_encoder.py

@@ -0,0 +1,902 @@
+"""
+Concise re-implementation of
+``https://github.com/openai/CLIP'' and
+``https://github.com/mlfoundations/open_clip''.
+"""
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision.transforms as T
+from .wan_video_dit import flash_attention
+
+
+class SelfAttention(nn.Module):
+
+    def __init__(self, dim, num_heads, dropout=0.1, eps=1e-5):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.eps = eps
+
+        # layers
+        self.q = nn.Linear(dim, dim)
+        self.k = nn.Linear(dim, dim)
+        self.v = nn.Linear(dim, dim)
+        self.o = nn.Linear(dim, dim)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, x, mask):
+        """
+        x:   [B, L, C].
+        """
+        b, s, c, n, d = *x.size(), self.num_heads, self.head_dim
+
+        # compute query, key, value
+        q = self.q(x).reshape(b, s, n, d).permute(0, 2, 1, 3)
+        k = self.k(x).reshape(b, s, n, d).permute(0, 2, 1, 3)
+        v = self.v(x).reshape(b, s, n, d).permute(0, 2, 1, 3)
+
+        # compute attention
+        p = self.dropout.p if self.training else 0.0
+        x = F.scaled_dot_product_attention(q, k, v, mask, p)
+        x = x.permute(0, 2, 1, 3).reshape(b, s, c)
+
+        # output
+        x = self.o(x)
+        x = self.dropout(x)
+        return x
+
+
+class AttentionBlock(nn.Module):
+
+    def __init__(self, dim, num_heads, post_norm, dropout=0.1, eps=1e-5):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.post_norm = post_norm
+        self.eps = eps
+
+        # layers
+        self.attn = SelfAttention(dim, num_heads, dropout, eps)
+        self.norm1 = nn.LayerNorm(dim, eps=eps)
+        self.ffn = nn.Sequential(
+            nn.Linear(dim, dim * 4), nn.GELU(), nn.Linear(dim * 4, dim),
+            nn.Dropout(dropout))
+        self.norm2 = nn.LayerNorm(dim, eps=eps)
+
+    def forward(self, x, mask):
+        if self.post_norm:
+            x = self.norm1(x + self.attn(x, mask))
+            x = self.norm2(x + self.ffn(x))
+        else:
+            x = x + self.attn(self.norm1(x), mask)
+            x = x + self.ffn(self.norm2(x))
+        return x
+
+
+class XLMRoberta(nn.Module):
+    """
+    XLMRobertaModel with no pooler and no LM head.
+    """
+
+    def __init__(self,
+                 vocab_size=250002,
+                 max_seq_len=514,
+                 type_size=1,
+                 pad_id=1,
+                 dim=1024,
+                 num_heads=16,
+                 num_layers=24,
+                 post_norm=True,
+                 dropout=0.1,
+                 eps=1e-5):
+        super().__init__()
+        self.vocab_size = vocab_size
+        self.max_seq_len = max_seq_len
+        self.type_size = type_size
+        self.pad_id = pad_id
+        self.dim = dim
+        self.num_heads = num_heads
+        self.num_layers = num_layers
+        self.post_norm = post_norm
+        self.eps = eps
+
+        # embeddings
+        self.token_embedding = nn.Embedding(vocab_size, dim, padding_idx=pad_id)
+        self.type_embedding = nn.Embedding(type_size, dim)
+        self.pos_embedding = nn.Embedding(max_seq_len, dim, padding_idx=pad_id)
+        self.dropout = nn.Dropout(dropout)
+
+        # blocks
+        self.blocks = nn.ModuleList([
+            AttentionBlock(dim, num_heads, post_norm, dropout, eps)
+            for _ in range(num_layers)
+        ])
+
+        # norm layer
+        self.norm = nn.LayerNorm(dim, eps=eps)
+
+    def forward(self, ids):
+        """
+        ids: [B, L] of torch.LongTensor.
+        """
+        b, s = ids.shape
+        mask = ids.ne(self.pad_id).long()
+
+        # embeddings
+        x = self.token_embedding(ids) + \
+            self.type_embedding(torch.zeros_like(ids)) + \
+            self.pos_embedding(self.pad_id + torch.cumsum(mask, dim=1) * mask)
+        if self.post_norm:
+            x = self.norm(x)
+        x = self.dropout(x)
+
+        # blocks
+        mask = torch.where(
+            mask.view(b, 1, 1, s).gt(0), 0.0,
+            torch.finfo(x.dtype).min)
+        for block in self.blocks:
+            x = block(x, mask)
+
+        # output
+        if not self.post_norm:
+            x = self.norm(x)
+        return x
+
+
+def xlm_roberta_large(pretrained=False,
+                      return_tokenizer=False,
+                      device='cpu',
+                      **kwargs):
+    """
+    XLMRobertaLarge adapted from Huggingface.
+    """
+    # params
+    cfg = dict(
+        vocab_size=250002,
+        max_seq_len=514,
+        type_size=1,
+        pad_id=1,
+        dim=1024,
+        num_heads=16,
+        num_layers=24,
+        post_norm=True,
+        dropout=0.1,
+        eps=1e-5)
+    cfg.update(**kwargs)
+
+    # init model
+    if pretrained:
+        from sora import DOWNLOAD_TO_CACHE
+
+        # init a meta model
+        with torch.device('meta'):
+            model = XLMRoberta(**cfg)
+
+        # load checkpoint
+        model.load_state_dict(
+            torch.load(
+                DOWNLOAD_TO_CACHE('models/xlm_roberta/xlm_roberta_large.pth'),
+                map_location=device),
+            assign=True)
+    else:
+        # init a model on device
+        with torch.device(device):
+            model = XLMRoberta(**cfg)
+
+    # init tokenizer
+    if return_tokenizer:
+        from sora.data import HuggingfaceTokenizer
+        tokenizer = HuggingfaceTokenizer(
+            name='xlm-roberta-large',
+            seq_len=model.text_len,
+            clean='whitespace')
+        return model, tokenizer
+    else:
+        return model
+
+
+
+def pos_interpolate(pos, seq_len):
+    if pos.size(1) == seq_len:
+        return pos
+    else:
+        src_grid = int(math.sqrt(pos.size(1)))
+        tar_grid = int(math.sqrt(seq_len))
+        n = pos.size(1) - src_grid * src_grid
+        return torch.cat([
+            pos[:, :n],
+            F.interpolate(
+                pos[:, n:].float().reshape(1, src_grid, src_grid, -1).permute(
+                    0, 3, 1, 2),
+                size=(tar_grid, tar_grid),
+                mode='bicubic',
+                align_corners=False).flatten(2).transpose(1, 2)
+        ],
+                         dim=1)
+
+
+class QuickGELU(nn.Module):
+
+    def forward(self, x):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class LayerNorm(nn.LayerNorm):
+
+    def forward(self, x):
+        return super().forward(x).type_as(x)
+
+
+class SelfAttention(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 num_heads,
+                 causal=False,
+                 attn_dropout=0.0,
+                 proj_dropout=0.0):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.causal = causal
+        self.attn_dropout = attn_dropout
+        self.proj_dropout = proj_dropout
+
+        # layers
+        self.to_qkv = nn.Linear(dim, dim * 3)
+        self.proj = nn.Linear(dim, dim)
+
+    def forward(self, x):
+        """
+        x:   [B, L, C].
+        """
+        # compute query, key, value
+        q, k, v = self.to_qkv(x).chunk(3, dim=-1)
+
+        # compute attention
+        x = flash_attention(q, k, v, num_heads=self.num_heads, compatibility_mode=True)
+
+        # output
+        x = self.proj(x)
+        x = F.dropout(x, self.proj_dropout, self.training)
+        return x
+
+
+class SwiGLU(nn.Module):
+
+    def __init__(self, dim, mid_dim):
+        super().__init__()
+        self.dim = dim
+        self.mid_dim = mid_dim
+
+        # layers
+        self.fc1 = nn.Linear(dim, mid_dim)
+        self.fc2 = nn.Linear(dim, mid_dim)
+        self.fc3 = nn.Linear(mid_dim, dim)
+
+    def forward(self, x):
+        x = F.silu(self.fc1(x)) * self.fc2(x)
+        x = self.fc3(x)
+        return x
+
+
+class AttentionBlock(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 mlp_ratio,
+                 num_heads,
+                 post_norm=False,
+                 causal=False,
+                 activation='quick_gelu',
+                 attn_dropout=0.0,
+                 proj_dropout=0.0,
+                 norm_eps=1e-5):
+        assert activation in ['quick_gelu', 'gelu', 'swi_glu']
+        super().__init__()
+        self.dim = dim
+        self.mlp_ratio = mlp_ratio
+        self.num_heads = num_heads
+        self.post_norm = post_norm
+        self.causal = causal
+        self.norm_eps = norm_eps
+
+        # layers
+        self.norm1 = LayerNorm(dim, eps=norm_eps)
+        self.attn = SelfAttention(dim, num_heads, causal, attn_dropout,
+                                  proj_dropout)
+        self.norm2 = LayerNorm(dim, eps=norm_eps)
+        if activation == 'swi_glu':
+            self.mlp = SwiGLU(dim, int(dim * mlp_ratio))
+        else:
+            self.mlp = nn.Sequential(
+                nn.Linear(dim, int(dim * mlp_ratio)),
+                QuickGELU() if activation == 'quick_gelu' else nn.GELU(),
+                nn.Linear(int(dim * mlp_ratio), dim), nn.Dropout(proj_dropout))
+
+    def forward(self, x):
+        if self.post_norm:
+            x = x + self.norm1(self.attn(x))
+            x = x + self.norm2(self.mlp(x))
+        else:
+            x = x + self.attn(self.norm1(x))
+            x = x + self.mlp(self.norm2(x))
+        return x
+
+
+class AttentionPool(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 mlp_ratio,
+                 num_heads,
+                 activation='gelu',
+                 proj_dropout=0.0,
+                 norm_eps=1e-5):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.mlp_ratio = mlp_ratio
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.proj_dropout = proj_dropout
+        self.norm_eps = norm_eps
+
+        # layers
+        gain = 1.0 / math.sqrt(dim)
+        self.cls_embedding = nn.Parameter(gain * torch.randn(1, 1, dim))
+        self.to_q = nn.Linear(dim, dim)
+        self.to_kv = nn.Linear(dim, dim * 2)
+        self.proj = nn.Linear(dim, dim)
+        self.norm = LayerNorm(dim, eps=norm_eps)
+        self.mlp = nn.Sequential(
+            nn.Linear(dim, int(dim * mlp_ratio)),
+            QuickGELU() if activation == 'quick_gelu' else nn.GELU(),
+            nn.Linear(int(dim * mlp_ratio), dim), nn.Dropout(proj_dropout))
+
+    def forward(self, x):
+        """
+        x:  [B, L, C].
+        """
+        b, s, c, n, d = *x.size(), self.num_heads, self.head_dim
+
+        # compute query, key, value
+        q = self.to_q(self.cls_embedding).view(1, 1, n*d).expand(b, -1, -1)
+        k, v = self.to_kv(x).chunk(2, dim=-1)
+
+        # compute attention
+        x = flash_attention(q, k, v, num_heads=self.num_heads, compatibility_mode=True)
+        x = x.reshape(b, 1, c)
+
+        # output
+        x = self.proj(x)
+        x = F.dropout(x, self.proj_dropout, self.training)
+
+        # mlp
+        x = x + self.mlp(self.norm(x))
+        return x[:, 0]
+
+
+class VisionTransformer(nn.Module):
+
+    def __init__(self,
+                 image_size=224,
+                 patch_size=16,
+                 dim=768,
+                 mlp_ratio=4,
+                 out_dim=512,
+                 num_heads=12,
+                 num_layers=12,
+                 pool_type='token',
+                 pre_norm=True,
+                 post_norm=False,
+                 activation='quick_gelu',
+                 attn_dropout=0.0,
+                 proj_dropout=0.0,
+                 embedding_dropout=0.0,
+                 norm_eps=1e-5):
+        if image_size % patch_size != 0:
+            print(
+                '[WARNING] image_size is not divisible by patch_size',
+                flush=True)
+        assert pool_type in ('token', 'token_fc', 'attn_pool')
+        out_dim = out_dim or dim
+        super().__init__()
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_patches = (image_size // patch_size)**2
+        self.dim = dim
+        self.mlp_ratio = mlp_ratio
+        self.out_dim = out_dim
+        self.num_heads = num_heads
+        self.num_layers = num_layers
+        self.pool_type = pool_type
+        self.post_norm = post_norm
+        self.norm_eps = norm_eps
+
+        # embeddings
+        gain = 1.0 / math.sqrt(dim)
+        self.patch_embedding = nn.Conv2d(
+            3,
+            dim,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=not pre_norm)
+        if pool_type in ('token', 'token_fc'):
+            self.cls_embedding = nn.Parameter(gain * torch.randn(1, 1, dim))
+        self.pos_embedding = nn.Parameter(gain * torch.randn(
+            1, self.num_patches +
+            (1 if pool_type in ('token', 'token_fc') else 0), dim))
+        self.dropout = nn.Dropout(embedding_dropout)
+
+        # transformer
+        self.pre_norm = LayerNorm(dim, eps=norm_eps) if pre_norm else None
+        self.transformer = nn.Sequential(*[
+            AttentionBlock(dim, mlp_ratio, num_heads, post_norm, False,
+                           activation, attn_dropout, proj_dropout, norm_eps)
+            for _ in range(num_layers)
+        ])
+        self.post_norm = LayerNorm(dim, eps=norm_eps)
+
+        # head
+        if pool_type == 'token':
+            self.head = nn.Parameter(gain * torch.randn(dim, out_dim))
+        elif pool_type == 'token_fc':
+            self.head = nn.Linear(dim, out_dim)
+        elif pool_type == 'attn_pool':
+            self.head = AttentionPool(dim, mlp_ratio, num_heads, activation,
+                                      proj_dropout, norm_eps)
+
+    def forward(self, x, interpolation=False, use_31_block=False):
+        b = x.size(0)
+
+        # embeddings
+        x = self.patch_embedding(x).flatten(2).permute(0, 2, 1)
+        if self.pool_type in ('token', 'token_fc'):
+            x = torch.cat([self.cls_embedding.expand(b, -1, -1).to(dtype=x.dtype, device=x.device), x], dim=1)
+        if interpolation:
+            e = pos_interpolate(self.pos_embedding, x.size(1))
+        else:
+            e = self.pos_embedding
+        e = e.to(dtype=x.dtype, device=x.device)
+        x = self.dropout(x + e)
+        if self.pre_norm is not None:
+            x = self.pre_norm(x)
+
+        # transformer
+        if use_31_block:
+            x = self.transformer[:-1](x)
+            return x
+        else:
+            x = self.transformer(x)
+            return x
+
+
+class CLIP(nn.Module):
+
+    def __init__(self,
+                 embed_dim=512,
+                 image_size=224,
+                 patch_size=16,
+                 vision_dim=768,
+                 vision_mlp_ratio=4,
+                 vision_heads=12,
+                 vision_layers=12,
+                 vision_pool='token',
+                 vision_pre_norm=True,
+                 vision_post_norm=False,
+                 vocab_size=49408,
+                 text_len=77,
+                 text_dim=512,
+                 text_mlp_ratio=4,
+                 text_heads=8,
+                 text_layers=12,
+                 text_causal=True,
+                 text_pool='argmax',
+                 text_head_bias=False,
+                 logit_bias=None,
+                 activation='quick_gelu',
+                 attn_dropout=0.0,
+                 proj_dropout=0.0,
+                 embedding_dropout=0.0,
+                 norm_eps=1e-5):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.vision_dim = vision_dim
+        self.vision_mlp_ratio = vision_mlp_ratio
+        self.vision_heads = vision_heads
+        self.vision_layers = vision_layers
+        self.vision_pool = vision_pool
+        self.vision_pre_norm = vision_pre_norm
+        self.vision_post_norm = vision_post_norm
+        self.vocab_size = vocab_size
+        self.text_len = text_len
+        self.text_dim = text_dim
+        self.text_mlp_ratio = text_mlp_ratio
+        self.text_heads = text_heads
+        self.text_layers = text_layers
+        self.text_causal = text_causal
+        self.text_pool = text_pool
+        self.text_head_bias = text_head_bias
+        self.norm_eps = norm_eps
+
+        # models
+        self.visual = VisionTransformer(
+            image_size=image_size,
+            patch_size=patch_size,
+            dim=vision_dim,
+            mlp_ratio=vision_mlp_ratio,
+            out_dim=embed_dim,
+            num_heads=vision_heads,
+            num_layers=vision_layers,
+            pool_type=vision_pool,
+            pre_norm=vision_pre_norm,
+            post_norm=vision_post_norm,
+            activation=activation,
+            attn_dropout=attn_dropout,
+            proj_dropout=proj_dropout,
+            embedding_dropout=embedding_dropout,
+            norm_eps=norm_eps)
+        self.textual = TextTransformer(
+            vocab_size=vocab_size,
+            text_len=text_len,
+            dim=text_dim,
+            mlp_ratio=text_mlp_ratio,
+            out_dim=embed_dim,
+            num_heads=text_heads,
+            num_layers=text_layers,
+            causal=text_causal,
+            pool_type=text_pool,
+            head_bias=text_head_bias,
+            activation=activation,
+            attn_dropout=attn_dropout,
+            proj_dropout=proj_dropout,
+            embedding_dropout=embedding_dropout,
+            norm_eps=norm_eps)
+        self.log_scale = nn.Parameter(math.log(1 / 0.07) * torch.ones([]))
+        if logit_bias is not None:
+            self.logit_bias = nn.Parameter(logit_bias * torch.ones([]))
+
+        # initialize weights
+        self.init_weights()
+
+    def forward(self, imgs, txt_ids):
+        """
+        imgs:       [B, 3, H, W] of torch.float32.
+        - mean:     [0.48145466, 0.4578275, 0.40821073]
+        - std:      [0.26862954, 0.26130258, 0.27577711]
+        txt_ids:    [B, L] of torch.long. Encoded by data.CLIPTokenizer.
+        """
+        xi = self.visual(imgs)
+        xt = self.textual(txt_ids)
+        return xi, xt
+
+    def init_weights(self):
+        # embeddings
+        nn.init.normal_(self.textual.token_embedding.weight, std=0.02)
+        nn.init.normal_(self.visual.patch_embedding.weight, std=0.1)
+
+        # attentions
+        for modality in ['visual', 'textual']:
+            dim = self.vision_dim if modality == 'visual' else self.text_dim
+            transformer = getattr(self, modality).transformer
+            proj_gain = (1.0 / math.sqrt(dim)) * (
+                1.0 / math.sqrt(2 * len(transformer)))
+            attn_gain = 1.0 / math.sqrt(dim)
+            mlp_gain = 1.0 / math.sqrt(2.0 * dim)
+            for block in transformer:
+                nn.init.normal_(block.attn.to_qkv.weight, std=attn_gain)
+                nn.init.normal_(block.attn.proj.weight, std=proj_gain)
+                nn.init.normal_(block.mlp[0].weight, std=mlp_gain)
+                nn.init.normal_(block.mlp[2].weight, std=proj_gain)
+
+    def param_groups(self):
+        groups = [{
+            'params': [
+                p for n, p in self.named_parameters()
+                if 'norm' in n or n.endswith('bias')
+            ],
+            'weight_decay': 0.0
+        }, {
+            'params': [
+                p for n, p in self.named_parameters()
+                if not ('norm' in n or n.endswith('bias'))
+            ]
+        }]
+        return groups
+
+
+class XLMRobertaWithHead(XLMRoberta):
+
+    def __init__(self, **kwargs):
+        self.out_dim = kwargs.pop('out_dim')
+        super().__init__(**kwargs)
+
+        # head
+        mid_dim = (self.dim + self.out_dim) // 2
+        self.head = nn.Sequential(
+            nn.Linear(self.dim, mid_dim, bias=False), nn.GELU(),
+            nn.Linear(mid_dim, self.out_dim, bias=False))
+
+    def forward(self, ids):
+        # xlm-roberta
+        x = super().forward(ids)
+
+        # average pooling
+        mask = ids.ne(self.pad_id).unsqueeze(-1).to(x)
+        x = (x * mask).sum(dim=1) / mask.sum(dim=1)
+
+        # head
+        x = self.head(x)
+        return x
+
+
+class XLMRobertaCLIP(nn.Module):
+
+    def __init__(self,
+                 embed_dim=1024,
+                 image_size=224,
+                 patch_size=14,
+                 vision_dim=1280,
+                 vision_mlp_ratio=4,
+                 vision_heads=16,
+                 vision_layers=32,
+                 vision_pool='token',
+                 vision_pre_norm=True,
+                 vision_post_norm=False,
+                 activation='gelu',
+                 vocab_size=250002,
+                 max_text_len=514,
+                 type_size=1,
+                 pad_id=1,
+                 text_dim=1024,
+                 text_heads=16,
+                 text_layers=24,
+                 text_post_norm=True,
+                 text_dropout=0.1,
+                 attn_dropout=0.0,
+                 proj_dropout=0.0,
+                 embedding_dropout=0.0,
+                 norm_eps=1e-5):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.vision_dim = vision_dim
+        self.vision_mlp_ratio = vision_mlp_ratio
+        self.vision_heads = vision_heads
+        self.vision_layers = vision_layers
+        self.vision_pre_norm = vision_pre_norm
+        self.vision_post_norm = vision_post_norm
+        self.activation = activation
+        self.vocab_size = vocab_size
+        self.max_text_len = max_text_len
+        self.type_size = type_size
+        self.pad_id = pad_id
+        self.text_dim = text_dim
+        self.text_heads = text_heads
+        self.text_layers = text_layers
+        self.text_post_norm = text_post_norm
+        self.norm_eps = norm_eps
+
+        # models
+        self.visual = VisionTransformer(
+            image_size=image_size,
+            patch_size=patch_size,
+            dim=vision_dim,
+            mlp_ratio=vision_mlp_ratio,
+            out_dim=embed_dim,
+            num_heads=vision_heads,
+            num_layers=vision_layers,
+            pool_type=vision_pool,
+            pre_norm=vision_pre_norm,
+            post_norm=vision_post_norm,
+            activation=activation,
+            attn_dropout=attn_dropout,
+            proj_dropout=proj_dropout,
+            embedding_dropout=embedding_dropout,
+            norm_eps=norm_eps)
+        self.textual = None
+        self.log_scale = nn.Parameter(math.log(1 / 0.07) * torch.ones([]))
+
+    def forward(self, imgs, txt_ids):
+        """
+        imgs:       [B, 3, H, W] of torch.float32.
+        - mean:     [0.48145466, 0.4578275, 0.40821073]
+        - std:      [0.26862954, 0.26130258, 0.27577711]
+        txt_ids:    [B, L] of torch.long.
+                    Encoded by data.CLIPTokenizer.
+        """
+        xi = self.visual(imgs)
+        xt = self.textual(txt_ids)
+        return xi, xt
+
+    def param_groups(self):
+        groups = [{
+            'params': [
+                p for n, p in self.named_parameters()
+                if 'norm' in n or n.endswith('bias')
+            ],
+            'weight_decay': 0.0
+        }, {
+            'params': [
+                p for n, p in self.named_parameters()
+                if not ('norm' in n or n.endswith('bias'))
+            ]
+        }]
+        return groups
+
+
+def _clip(pretrained=False,
+          pretrained_name=None,
+          model_cls=CLIP,
+          return_transforms=False,
+          return_tokenizer=False,
+          tokenizer_padding='eos',
+          dtype=torch.float32,
+          device='cpu',
+          **kwargs):
+    # init model
+    if pretrained and pretrained_name:
+        from sora import BUCKET, DOWNLOAD_TO_CACHE
+
+        # init a meta model
+        with torch.device('meta'):
+            model = model_cls(**kwargs)
+
+        # checkpoint path
+        checkpoint = f'models/clip/{pretrained_name}'
+        if dtype in (torch.float16, torch.bfloat16):
+            suffix = '-' + {
+                torch.float16: 'fp16',
+                torch.bfloat16: 'bf16'
+            }[dtype]
+            if object_exists(BUCKET, f'{checkpoint}{suffix}.pth'):
+                checkpoint = f'{checkpoint}{suffix}'
+        checkpoint += '.pth'
+
+        # load
+        model.load_state_dict(
+            torch.load(DOWNLOAD_TO_CACHE(checkpoint), map_location=device),
+            assign=True,
+            strict=False)
+    else:
+        # init a model on device
+        with torch.device(device):
+            model = model_cls(**kwargs)
+
+    # set device
+    output = (model,)
+
+    # init transforms
+    if return_transforms:
+        # mean and std
+        if 'siglip' in pretrained_name.lower():
+            mean, std = [0.5, 0.5, 0.5], [0.5, 0.5, 0.5]
+        else:
+            mean = [0.48145466, 0.4578275, 0.40821073]
+            std = [0.26862954, 0.26130258, 0.27577711]
+
+        # transforms
+        transforms = T.Compose([
+            T.Resize((model.image_size, model.image_size),
+                     interpolation=T.InterpolationMode.BICUBIC),
+            T.ToTensor(),
+            T.Normalize(mean=mean, std=std)
+        ])
+        output += (transforms,)
+
+    # init tokenizer
+    if return_tokenizer:
+        from sora import data
+        if 'siglip' in pretrained_name.lower():
+            tokenizer = data.HuggingfaceTokenizer(
+                name=f'timm/{pretrained_name}',
+                seq_len=model.text_len,
+                clean='canonicalize')
+        elif 'xlm' in pretrained_name.lower():
+            tokenizer = data.HuggingfaceTokenizer(
+                name='xlm-roberta-large',
+                seq_len=model.max_text_len - 2,
+                clean='whitespace')
+        elif 'mba' in pretrained_name.lower():
+            tokenizer = data.HuggingfaceTokenizer(
+                name='facebook/xlm-roberta-xl',
+                seq_len=model.max_text_len - 2,
+                clean='whitespace')
+        else:
+            tokenizer = data.CLIPTokenizer(
+                seq_len=model.text_len, padding=tokenizer_padding)
+        output += (tokenizer,)
+    return output[0] if len(output) == 1 else output
+
+
+def clip_xlm_roberta_vit_h_14(
+        pretrained=False,
+        pretrained_name='open-clip-xlm-roberta-large-vit-huge-14',
+        **kwargs):
+    cfg = dict(
+        embed_dim=1024,
+        image_size=224,
+        patch_size=14,
+        vision_dim=1280,
+        vision_mlp_ratio=4,
+        vision_heads=16,
+        vision_layers=32,
+        vision_pool='token',
+        activation='gelu',
+        vocab_size=250002,
+        max_text_len=514,
+        type_size=1,
+        pad_id=1,
+        text_dim=1024,
+        text_heads=16,
+        text_layers=24,
+        text_post_norm=True,
+        text_dropout=0.1,
+        attn_dropout=0.0,
+        proj_dropout=0.0,
+        embedding_dropout=0.0)
+    cfg.update(**kwargs)
+    return _clip(pretrained, pretrained_name, XLMRobertaCLIP, **cfg)
+
+
+class WanImageEncoder(torch.nn.Module):
+
+    def __init__(self):
+        super().__init__()
+        # init model
+        self.model, self.transforms = clip_xlm_roberta_vit_h_14(
+            pretrained=False,
+            return_transforms=True,
+            return_tokenizer=False,
+            dtype=torch.float32,
+            device="cpu")
+
+    def encode_image(self, videos):
+        # preprocess
+        size = (self.model.image_size,) * 2
+        videos = torch.cat([
+            F.interpolate(
+                u,
+                size=size,
+                mode='bicubic',
+                align_corners=False) for u in videos
+        ])
+        videos = self.transforms.transforms[-1](videos.mul_(0.5).add_(0.5))
+
+        # forward
+        dtype = next(iter(self.model.visual.parameters())).dtype
+        videos = videos.to(dtype)
+        out = self.model.visual(videos, use_31_block=True)
+        return out
+        
+    @staticmethod
+    def state_dict_converter():
+        return WanImageEncoderStateDictConverter()
+    
+    
+class WanImageEncoderStateDictConverter:
+    def __init__(self):
+        pass
+
+    def from_diffusers(self, state_dict):
+        return state_dict
+    
+    def from_civitai(self, state_dict):
+        state_dict_ = {}
+        for name, param in state_dict.items():
+            if name.startswith("textual."):
+                continue
+            name = "model." + name
+            state_dict_[name] = param
+        return state_dict_
+

diffsynth/models/wan_video_motion_controller.py

@@ -0,0 +1,44 @@
+import torch
+import torch.nn as nn
+from .wan_video_dit import sinusoidal_embedding_1d
+
+
+
+class WanMotionControllerModel(torch.nn.Module):
+    def __init__(self, freq_dim=256, dim=1536):
+        super().__init__()
+        self.freq_dim = freq_dim
+        self.linear = nn.Sequential(
+            nn.Linear(freq_dim, dim),
+            nn.SiLU(),
+            nn.Linear(dim, dim),
+            nn.SiLU(),
+            nn.Linear(dim, dim * 6),
+        )
+
+    def forward(self, motion_bucket_id):
+        emb = sinusoidal_embedding_1d(self.freq_dim, motion_bucket_id * 10)
+        emb = self.linear(emb)
+        return emb
+
+    def init(self):
+        state_dict = self.linear[-1].state_dict()
+        state_dict = {i: state_dict[i] * 0 for i in state_dict}
+        self.linear[-1].load_state_dict(state_dict)
+
+    @staticmethod
+    def state_dict_converter():
+        return WanMotionControllerModelDictConverter()
+    
+    
+
+class WanMotionControllerModelDictConverter:
+    def __init__(self):
+        pass
+
+    def from_diffusers(self, state_dict):
+        return state_dict
+    
+    def from_civitai(self, state_dict):
+        return state_dict
+

diffsynth/models/wan_video_text_encoder.py

@@ -0,0 +1,269 @@
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def fp16_clamp(x):
+    if x.dtype == torch.float16 and torch.isinf(x).any():
+        clamp = torch.finfo(x.dtype).max - 1000
+        x = torch.clamp(x, min=-clamp, max=clamp)
+    return x
+
+
+class GELU(nn.Module):
+
+    def forward(self, x):
+        return 0.5 * x * (1.0 + torch.tanh(
+            math.sqrt(2.0 / math.pi) * (x + 0.044715 * torch.pow(x, 3.0))))
+
+
+class T5LayerNorm(nn.Module):
+
+    def __init__(self, dim, eps=1e-6):
+        super(T5LayerNorm, self).__init__()
+        self.dim = dim
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x):
+        x = x * torch.rsqrt(x.float().pow(2).mean(dim=-1, keepdim=True) +
+                            self.eps)
+        if self.weight.dtype in [torch.float16, torch.bfloat16]:
+            x = x.type_as(self.weight)
+        return self.weight * x
+
+
+class T5Attention(nn.Module):
+
+    def __init__(self, dim, dim_attn, num_heads, dropout=0.1):
+        assert dim_attn % num_heads == 0
+        super(T5Attention, self).__init__()
+        self.dim = dim
+        self.dim_attn = dim_attn
+        self.num_heads = num_heads
+        self.head_dim = dim_attn // num_heads
+
+        # layers
+        self.q = nn.Linear(dim, dim_attn, bias=False)
+        self.k = nn.Linear(dim, dim_attn, bias=False)
+        self.v = nn.Linear(dim, dim_attn, bias=False)
+        self.o = nn.Linear(dim_attn, dim, bias=False)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, x, context=None, mask=None, pos_bias=None):
+        """
+        x:          [B, L1, C].
+        context:    [B, L2, C] or None.
+        mask:       [B, L2] or [B, L1, L2] or None.
+        """
+        # check inputs
+        context = x if context is None else context
+        b, n, c = x.size(0), self.num_heads, self.head_dim
+
+        # compute query, key, value
+        q = self.q(x).view(b, -1, n, c)
+        k = self.k(context).view(b, -1, n, c)
+        v = self.v(context).view(b, -1, n, c)
+
+        # attention bias
+        attn_bias = x.new_zeros(b, n, q.size(1), k.size(1))
+        if pos_bias is not None:
+            attn_bias += pos_bias
+        if mask is not None:
+            assert mask.ndim in [2, 3]
+            mask = mask.view(b, 1, 1,
+                             -1) if mask.ndim == 2 else mask.unsqueeze(1)
+            attn_bias.masked_fill_(mask == 0, torch.finfo(x.dtype).min)
+
+        # compute attention (T5 does not use scaling)
+        attn = torch.einsum('binc,bjnc->bnij', q, k) + attn_bias
+        attn = F.softmax(attn.float(), dim=-1).type_as(attn)
+        x = torch.einsum('bnij,bjnc->binc', attn, v)
+
+        # output
+        x = x.reshape(b, -1, n * c)
+        x = self.o(x)
+        x = self.dropout(x)
+        return x
+
+
+class T5FeedForward(nn.Module):
+
+    def __init__(self, dim, dim_ffn, dropout=0.1):
+        super(T5FeedForward, self).__init__()
+        self.dim = dim
+        self.dim_ffn = dim_ffn
+
+        # layers
+        self.gate = nn.Sequential(nn.Linear(dim, dim_ffn, bias=False), GELU())
+        self.fc1 = nn.Linear(dim, dim_ffn, bias=False)
+        self.fc2 = nn.Linear(dim_ffn, dim, bias=False)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, x):
+        x = self.fc1(x) * self.gate(x)
+        x = self.dropout(x)
+        x = self.fc2(x)
+        x = self.dropout(x)
+        return x
+
+
+class T5SelfAttention(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 dim_attn,
+                 dim_ffn,
+                 num_heads,
+                 num_buckets,
+                 shared_pos=True,
+                 dropout=0.1):
+        super(T5SelfAttention, self).__init__()
+        self.dim = dim
+        self.dim_attn = dim_attn
+        self.dim_ffn = dim_ffn
+        self.num_heads = num_heads
+        self.num_buckets = num_buckets
+        self.shared_pos = shared_pos
+
+        # layers
+        self.norm1 = T5LayerNorm(dim)
+        self.attn = T5Attention(dim, dim_attn, num_heads, dropout)
+        self.norm2 = T5LayerNorm(dim)
+        self.ffn = T5FeedForward(dim, dim_ffn, dropout)
+        self.pos_embedding = None if shared_pos else T5RelativeEmbedding(
+            num_buckets, num_heads, bidirectional=True)
+
+    def forward(self, x, mask=None, pos_bias=None):
+        e = pos_bias if self.shared_pos else self.pos_embedding(
+            x.size(1), x.size(1))
+        x = fp16_clamp(x + self.attn(self.norm1(x), mask=mask, pos_bias=e))
+        x = fp16_clamp(x + self.ffn(self.norm2(x)))
+        return x
+
+
+class T5RelativeEmbedding(nn.Module):
+
+    def __init__(self, num_buckets, num_heads, bidirectional, max_dist=128):
+        super(T5RelativeEmbedding, self).__init__()
+        self.num_buckets = num_buckets
+        self.num_heads = num_heads
+        self.bidirectional = bidirectional
+        self.max_dist = max_dist
+
+        # layers
+        self.embedding = nn.Embedding(num_buckets, num_heads)
+
+    def forward(self, lq, lk):
+        device = self.embedding.weight.device
+        # rel_pos = torch.arange(lk).unsqueeze(0).to(device) - \
+        #     torch.arange(lq).unsqueeze(1).to(device)
+        rel_pos = torch.arange(lk, device=device).unsqueeze(0) - \
+            torch.arange(lq, device=device).unsqueeze(1)
+        rel_pos = self._relative_position_bucket(rel_pos)
+        rel_pos_embeds = self.embedding(rel_pos)
+        rel_pos_embeds = rel_pos_embeds.permute(2, 0, 1).unsqueeze(
+            0)  # [1, N, Lq, Lk]
+        return rel_pos_embeds.contiguous()
+
+    def _relative_position_bucket(self, rel_pos):
+        # preprocess
+        if self.bidirectional:
+            num_buckets = self.num_buckets // 2
+            rel_buckets = (rel_pos > 0).long() * num_buckets
+            rel_pos = torch.abs(rel_pos)
+        else:
+            num_buckets = self.num_buckets
+            rel_buckets = 0
+            rel_pos = -torch.min(rel_pos, torch.zeros_like(rel_pos))
+
+        # embeddings for small and large positions
+        max_exact = num_buckets // 2
+        rel_pos_large = max_exact + (torch.log(rel_pos.float() / max_exact) /
+                                     math.log(self.max_dist / max_exact) *
+                                     (num_buckets - max_exact)).long()
+        rel_pos_large = torch.min(
+            rel_pos_large, torch.full_like(rel_pos_large, num_buckets - 1))
+        rel_buckets += torch.where(rel_pos < max_exact, rel_pos, rel_pos_large)
+        return rel_buckets
+
+def init_weights(m):
+    if isinstance(m, T5LayerNorm):
+        nn.init.ones_(m.weight)
+    elif isinstance(m, T5FeedForward):
+        nn.init.normal_(m.gate[0].weight, std=m.dim**-0.5)
+        nn.init.normal_(m.fc1.weight, std=m.dim**-0.5)
+        nn.init.normal_(m.fc2.weight, std=m.dim_ffn**-0.5)
+    elif isinstance(m, T5Attention):
+        nn.init.normal_(m.q.weight, std=(m.dim * m.dim_attn)**-0.5)
+        nn.init.normal_(m.k.weight, std=m.dim**-0.5)
+        nn.init.normal_(m.v.weight, std=m.dim**-0.5)
+        nn.init.normal_(m.o.weight, std=(m.num_heads * m.dim_attn)**-0.5)
+    elif isinstance(m, T5RelativeEmbedding):
+        nn.init.normal_(
+            m.embedding.weight, std=(2 * m.num_buckets * m.num_heads)**-0.5)
+
+
+class WanTextEncoder(torch.nn.Module):
+
+    def __init__(self,
+                 vocab=256384,
+                 dim=4096,
+                 dim_attn=4096,
+                 dim_ffn=10240,
+                 num_heads=64,
+                 num_layers=24,
+                 num_buckets=32,
+                 shared_pos=False,
+                 dropout=0.1):
+        super(WanTextEncoder, self).__init__()
+        self.dim = dim
+        self.dim_attn = dim_attn
+        self.dim_ffn = dim_ffn
+        self.num_heads = num_heads
+        self.num_layers = num_layers
+        self.num_buckets = num_buckets
+        self.shared_pos = shared_pos
+
+        # layers
+        self.token_embedding = vocab if isinstance(vocab, nn.Embedding) \
+            else nn.Embedding(vocab, dim)
+        self.pos_embedding = T5RelativeEmbedding(
+            num_buckets, num_heads, bidirectional=True) if shared_pos else None
+        self.dropout = nn.Dropout(dropout)
+        self.blocks = nn.ModuleList([
+            T5SelfAttention(dim, dim_attn, dim_ffn, num_heads, num_buckets,
+                            shared_pos, dropout) for _ in range(num_layers)
+        ])
+        self.norm = T5LayerNorm(dim)
+
+        # initialize weights
+        self.apply(init_weights)
+
+    def forward(self, ids, mask=None):
+        x = self.token_embedding(ids)
+        x = self.dropout(x)
+        e = self.pos_embedding(x.size(1),
+                               x.size(1)) if self.shared_pos else None
+        for block in self.blocks:
+            x = block(x, mask, pos_bias=e)
+        x = self.norm(x)
+        x = self.dropout(x)
+        return x
+    
+    @staticmethod
+    def state_dict_converter():
+        return WanTextEncoderStateDictConverter()
+    
+    
+class WanTextEncoderStateDictConverter:
+    def __init__(self):
+        pass
+
+    def from_diffusers(self, state_dict):
+        return state_dict
+    
+    def from_civitai(self, state_dict):
+        return state_dict

diffsynth/models/wan_video_vace.py

@@ -0,0 +1,113 @@
+import torch
+from .wan_video_dit import DiTBlock
+from .utils import hash_state_dict_keys
+
+class VaceWanAttentionBlock(DiTBlock):
+    def __init__(self, has_image_input, dim, num_heads, ffn_dim, eps=1e-6, block_id=0):
+        super().__init__(has_image_input, dim, num_heads, ffn_dim, eps=eps)
+        self.block_id = block_id
+        if block_id == 0:
+            self.before_proj = torch.nn.Linear(self.dim, self.dim)
+        self.after_proj = torch.nn.Linear(self.dim, self.dim)
+
+    def forward(self, c, x, context, t_mod, freqs):
+        if self.block_id == 0:
+            c = self.before_proj(c) + x
+            all_c = []
+        else:
+            all_c = list(torch.unbind(c))
+            c = all_c.pop(-1)
+        c = super().forward(c, context, t_mod, freqs)
+        c_skip = self.after_proj(c)
+        all_c += [c_skip, c]
+        c = torch.stack(all_c)
+        return c
+
+
+class VaceWanModel(torch.nn.Module):
+    def __init__(
+        self,
+        vace_layers=(0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28),
+        vace_in_dim=96,
+        patch_size=(1, 2, 2),
+        has_image_input=False,
+        dim=1536,
+        num_heads=12,
+        ffn_dim=8960,
+        eps=1e-6,
+    ):
+        super().__init__()
+        self.vace_layers = vace_layers
+        self.vace_in_dim = vace_in_dim
+        self.vace_layers_mapping = {i: n for n, i in enumerate(self.vace_layers)}
+
+        # vace blocks
+        self.vace_blocks = torch.nn.ModuleList([
+            VaceWanAttentionBlock(has_image_input, dim, num_heads, ffn_dim, eps, block_id=i)
+            for i in self.vace_layers
+        ])
+
+        # vace patch embeddings
+        self.vace_patch_embedding = torch.nn.Conv3d(vace_in_dim, dim, kernel_size=patch_size, stride=patch_size)
+
+    def forward(
+        self, x, vace_context, context, t_mod, freqs,
+        use_gradient_checkpointing: bool = False,
+        use_gradient_checkpointing_offload: bool = False,
+    ):
+        c = [self.vace_patch_embedding(u.unsqueeze(0)) for u in vace_context]
+        c = [u.flatten(2).transpose(1, 2) for u in c]
+        c = torch.cat([
+            torch.cat([u, u.new_zeros(1, x.shape[1] - u.size(1), u.size(2))],
+                      dim=1) for u in c
+        ])
+        
+        def create_custom_forward(module):
+            def custom_forward(*inputs):
+                return module(*inputs)
+            return custom_forward
+        
+        for block in self.vace_blocks:
+            if use_gradient_checkpointing_offload:
+                with torch.autograd.graph.save_on_cpu():
+                    c = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(block),
+                        c, x, context, t_mod, freqs,
+                        use_reentrant=False,
+                    )
+            elif use_gradient_checkpointing:
+                c = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    c, x, context, t_mod, freqs,
+                    use_reentrant=False,
+                )
+            else:
+                c = block(c, x, context, t_mod, freqs)
+        hints = torch.unbind(c)[:-1]
+        return hints
+    
+    @staticmethod
+    def state_dict_converter():
+        return VaceWanModelDictConverter()
+    
+    
+class VaceWanModelDictConverter:
+    def __init__(self):
+        pass
+    
+    def from_civitai(self, state_dict):
+        state_dict_ = {name: param for name, param in state_dict.items() if name.startswith("vace")}
+        if hash_state_dict_keys(state_dict_) == '3b2726384e4f64837bdf216eea3f310d': # vace 14B
+            config = {
+                "vace_layers": (0, 5, 10, 15, 20, 25, 30, 35),
+                "vace_in_dim": 96,
+                "patch_size": (1, 2, 2),
+                "has_image_input": False,
+                "dim": 5120,
+                "num_heads": 40,
+                "ffn_dim": 13824,
+                "eps": 1e-06,                
+            }
+        else:
+            config = {}
+        return state_dict_, config

diffsynth/models/wan_video_vae.py

@@ -0,0 +1,828 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange, repeat
+from tqdm import tqdm
+
+CACHE_T = 2
+
+
+def check_is_instance(model, module_class):
+    if isinstance(model, module_class):
+        return True
+    if hasattr(model, "module") and isinstance(model.module, module_class):
+        return True
+    return False
+
+
+def block_causal_mask(x, block_size):
+    # params
+    b, n, s, _, device = *x.size(), x.device
+    assert s % block_size == 0
+    num_blocks = s // block_size
+
+    # build mask
+    mask = torch.zeros(b, n, s, s, dtype=torch.bool, device=device)
+    for i in range(num_blocks):
+        mask[:, :,
+             i * block_size:(i + 1) * block_size, :(i + 1) * block_size] = 1
+    return mask
+
+
+class CausalConv3d(nn.Conv3d):
+    """
+    Causal 3d convolusion.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._padding = (self.padding[2], self.padding[2], self.padding[1],
+                         self.padding[1], 2 * self.padding[0], 0)
+        self.padding = (0, 0, 0)
+
+    def forward(self, x, cache_x=None):
+        padding = list(self._padding)
+        if cache_x is not None and self._padding[4] > 0:
+            cache_x = cache_x.to(x.device)
+            x = torch.cat([cache_x, x], dim=2)
+            padding[4] -= cache_x.shape[2]
+        x = F.pad(x, padding)
+
+        return super().forward(x)
+
+
+class RMS_norm(nn.Module):
+
+    def __init__(self, dim, channel_first=True, images=True, bias=False):
+        super().__init__()
+        broadcastable_dims = (1, 1, 1) if not images else (1, 1)
+        shape = (dim, *broadcastable_dims) if channel_first else (dim,)
+
+        self.channel_first = channel_first
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.ones(shape))
+        self.bias = nn.Parameter(torch.zeros(shape)) if bias else 0.
+
+    def forward(self, x):
+        return F.normalize(
+            x, dim=(1 if self.channel_first else
+                    -1)) * self.scale * self.gamma + self.bias
+
+
+class Upsample(nn.Upsample):
+
+    def forward(self, x):
+        """
+        Fix bfloat16 support for nearest neighbor interpolation.
+        """
+        return super().forward(x.float()).type_as(x)
+
+
+class Resample(nn.Module):
+
+    def __init__(self, dim, mode):
+        assert mode in ('none', 'upsample2d', 'upsample3d', 'downsample2d',
+                        'downsample3d')
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == 'upsample2d':
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Conv2d(dim, dim // 2, 3, padding=1))
+        elif mode == 'upsample3d':
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Conv2d(dim, dim // 2, 3, padding=1))
+            self.time_conv = CausalConv3d(dim,
+                                          dim * 2, (3, 1, 1),
+                                          padding=(1, 0, 0))
+
+        elif mode == 'downsample2d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == 'downsample3d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = CausalConv3d(dim,
+                                          dim, (3, 1, 1),
+                                          stride=(2, 1, 1),
+                                          padding=(0, 0, 0))
+
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        b, c, t, h, w = x.size()
+        if self.mode == 'upsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = 'Rep'
+                    feat_idx[0] += 1
+                else:
+
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] != 'Rep':
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat([
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device), cache_x
+                        ],
+                            dim=2)
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] == 'Rep':
+                        cache_x = torch.cat([
+                            torch.zeros_like(cache_x).to(cache_x.device),
+                            cache_x
+                        ],
+                            dim=2)
+                    if feat_cache[idx] == 'Rep':
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]),
+                                    3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.resample(x)
+        x = rearrange(x, '(b t) c h w -> b c t h w', t=t)
+
+        if self.mode == 'downsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    x = self.time_conv(
+                        torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+        return x
+
+    def init_weight(self, conv):
+        conv_weight = conv.weight
+        nn.init.zeros_(conv_weight)
+        c1, c2, t, h, w = conv_weight.size()
+        one_matrix = torch.eye(c1, c2)
+        init_matrix = one_matrix
+        nn.init.zeros_(conv_weight)
+        conv_weight.data[:, :, 1, 0, 0] = init_matrix
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+    def init_weight2(self, conv):
+        conv_weight = conv.weight.data
+        nn.init.zeros_(conv_weight)
+        c1, c2, t, h, w = conv_weight.size()
+        init_matrix = torch.eye(c1 // 2, c2)
+        conv_weight[:c1 // 2, :, -1, 0, 0] = init_matrix
+        conv_weight[c1 // 2:, :, -1, 0, 0] = init_matrix
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+
+class ResidualBlock(nn.Module):
+
+    def __init__(self, in_dim, out_dim, dropout=0.0):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # layers
+        self.residual = nn.Sequential(
+            RMS_norm(in_dim, images=False), nn.SiLU(),
+            CausalConv3d(in_dim, out_dim, 3, padding=1),
+            RMS_norm(out_dim, images=False), nn.SiLU(), nn.Dropout(dropout),
+            CausalConv3d(out_dim, out_dim, 3, padding=1))
+        self.shortcut = CausalConv3d(in_dim, out_dim, 1) \
+            if in_dim != out_dim else nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        h = self.shortcut(x)
+        for layer in self.residual:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x + h
+
+
+class AttentionBlock(nn.Module):
+    """
+    Causal self-attention with a single head.
+    """
+
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+        # layers
+        self.norm = RMS_norm(dim)
+        self.to_qkv = nn.Conv2d(dim, dim * 3, 1)
+        self.proj = nn.Conv2d(dim, dim, 1)
+
+        # zero out the last layer params
+        nn.init.zeros_(self.proj.weight)
+
+    def forward(self, x):
+        identity = x
+        b, c, t, h, w = x.size()
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.norm(x)
+        # compute query, key, value
+        q, k, v = self.to_qkv(x).reshape(b * t, 1, c * 3, -1).permute(
+            0, 1, 3, 2).contiguous().chunk(3, dim=-1)
+
+        # apply attention
+        x = F.scaled_dot_product_attention(
+            q,
+            k,
+            v,
+            # attn_mask=block_causal_mask(q, block_size=h * w)
+        )
+        x = x.squeeze(1).permute(0, 2, 1).reshape(b * t, c, h, w)
+
+        # output
+        x = self.proj(x)
+        x = rearrange(x, '(b t) c h w-> b c t h w', t=t)
+        return x + identity
+
+
+class Encoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[True, True, False],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(3, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            for _ in range(num_res_blocks):
+                downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    downsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # downsample block
+            if i != len(dim_mult) - 1:
+                mode = 'downsample3d' if temperal_downsample[
+                    i] else 'downsample2d'
+                downsamples.append(Resample(out_dim, mode=mode))
+                scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(ResidualBlock(out_dim, out_dim, dropout),
+                                    AttentionBlock(out_dim),
+                                    ResidualBlock(out_dim, out_dim, dropout))
+
+        # output blocks
+        self.head = nn.Sequential(RMS_norm(out_dim, images=False), nn.SiLU(),
+                                  CausalConv3d(out_dim, z_dim, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        # downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        # middle
+        for layer in self.middle:
+            if check_is_instance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        # head
+        for layer in self.head:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+class Decoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_upsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2**(len(dim_mult) - 2)
+
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(ResidualBlock(dims[0], dims[0], dropout),
+                                    AttentionBlock(dims[0]),
+                                    ResidualBlock(dims[0], dims[0], dropout))
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            if i == 1 or i == 2 or i == 3:
+                in_dim = in_dim // 2
+            for _ in range(num_res_blocks + 1):
+                upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    upsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # upsample block
+            if i != len(dim_mult) - 1:
+                mode = 'upsample3d' if temperal_upsample[i] else 'upsample2d'
+                upsamples.append(Resample(out_dim, mode=mode))
+                scale *= 2.0
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(RMS_norm(out_dim, images=False), nn.SiLU(),
+                                  CausalConv3d(out_dim, 3, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        # conv1
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        # middle
+        for layer in self.middle:
+            if check_is_instance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        # upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        # head
+        for layer in self.head:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+def count_conv3d(model):
+    count = 0
+    for m in model.modules():
+        if check_is_instance(m, CausalConv3d):
+            count += 1
+    return count
+
+
+class VideoVAE_(nn.Module):
+
+    def __init__(self,
+                 dim=96,
+                 z_dim=16,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_downsample, dropout)
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_upsample, dropout)
+
+    def forward(self, x):
+        mu, log_var = self.encode(x)
+        z = self.reparameterize(mu, log_var)
+        x_recon = self.decode(z)
+        return x_recon, mu, log_var
+
+    def encode(self, x, scale):
+        self.clear_cache()
+        # cache
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(x[:, :, :1, :, :],
+                                   feat_cache=self._enc_feat_map,
+                                   feat_idx=self._enc_conv_idx)
+            else:
+                out_ = self.encoder(x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                                    feat_cache=self._enc_feat_map,
+                                    feat_idx=self._enc_conv_idx)
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=mu.dtype, device=mu.device) for s in scale]
+            mu = (mu - scale[0].view(1, self.z_dim, 1, 1, 1)) * scale[1].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=mu.dtype, device=mu.device)
+            mu = (mu - scale[0]) * scale[1]
+        return mu
+
+    def decode(self, z, scale):
+        self.clear_cache()
+        # z: [b,c,t,h,w]
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=z.dtype, device=z.device) for s in scale]
+            z = z / scale[1].view(1, self.z_dim, 1, 1, 1) + scale[0].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=z.dtype, device=z.device)
+            z = z / scale[1] + scale[0]
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(x[:, :, i:i + 1, :, :],
+                                   feat_cache=self._feat_map,
+                                   feat_idx=self._conv_idx)
+            else:
+                out_ = self.decoder(x[:, :, i:i + 1, :, :],
+                                    feat_cache=self._feat_map,
+                                    feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2)  # may add tensor offload
+        return out
+
+    def reparameterize(self, mu, log_var):
+        std = torch.exp(0.5 * log_var)
+        eps = torch.randn_like(std)
+        return eps * std + mu
+
+    def sample(self, imgs, deterministic=False):
+        mu, log_var = self.encode(imgs)
+        if deterministic:
+            return mu
+        std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
+        return mu + std * torch.randn_like(std)
+
+    def clear_cache(self):
+        self._conv_num = count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        # cache encode
+        self._enc_conv_num = count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num
+
+
+class WanVideoVAE(nn.Module):
+
+    def __init__(self, z_dim=16):
+        super().__init__()
+
+        mean = [
+            -0.7571, -0.7089, -0.9113, 0.1075, -0.1745, 0.9653, -0.1517, 1.5508,
+            0.4134, -0.0715, 0.5517, -0.3632, -0.1922, -0.9497, 0.2503, -0.2921
+        ]
+        std = [
+            2.8184, 1.4541, 2.3275, 2.6558, 1.2196, 1.7708, 2.6052, 2.0743,
+            3.2687, 2.1526, 2.8652, 1.5579, 1.6382, 1.1253, 2.8251, 1.9160
+        ]
+        self.mean = torch.tensor(mean)
+        self.std = torch.tensor(std)
+        self.scale = [self.mean, 1.0 / self.std]
+
+        # init model
+        self.model = VideoVAE_(z_dim=z_dim).eval().requires_grad_(False)
+        self.upsampling_factor = 8
+
+    def build_1d_mask(self, length, left_bound, right_bound, border_width):
+        x = torch.ones((length,))
+        if not left_bound:
+            x[:border_width] = (torch.arange(border_width) + 1) / border_width
+        if not right_bound:
+            x[-border_width:] = torch.flip(
+                (torch.arange(border_width) + 1) / border_width, dims=(0,))
+        return x
+
+    def build_mask(self, data, is_bound, border_width):
+        _, _, _, H, W = data.shape
+        h = self.build_1d_mask(H, is_bound[0], is_bound[1], border_width[0])
+        w = self.build_1d_mask(W, is_bound[2], is_bound[3], border_width[1])
+
+        h = repeat(h, "H -> H W", H=H, W=W)
+        w = repeat(w, "W -> H W", H=H, W=W)
+
+        mask = torch.stack([h, w]).min(dim=0).values
+        mask = rearrange(mask, "H W -> 1 1 1 H W")
+        return mask
+
+    def tiled_decode(self, hidden_states, device, tile_size, tile_stride):
+        _, _, T, H, W = hidden_states.shape
+        size_h, size_w = tile_size
+        stride_h, stride_w = tile_stride
+
+        # Split tasks
+        tasks = []
+        for h in range(0, H, stride_h):
+            if (h-stride_h >= 0 and h-stride_h+size_h >= H):
+                continue
+            for w in range(0, W, stride_w):
+                if (w-stride_w >= 0 and w-stride_w+size_w >= W):
+                    continue
+                h_, w_ = h + size_h, w + size_w
+                tasks.append((h, h_, w, w_))
+
+        data_device = "cpu"
+        computation_device = device
+
+        out_T = T * 4 - 3
+        weight = torch.zeros((1, 1, out_T, H * self.upsampling_factor, W *
+                             self.upsampling_factor), dtype=hidden_states.dtype, device=data_device)
+        values = torch.zeros((1, 3, out_T, H * self.upsampling_factor, W *
+                             self.upsampling_factor), dtype=hidden_states.dtype, device=data_device)
+        disable_flag = True
+        for h, h_, w, w_ in tqdm(tasks, desc="VAE decoding", disable=disable_flag):
+            hidden_states_batch = hidden_states[:, :, :, h:h_, w:w_].to(
+                computation_device)
+            hidden_states_batch = self.model.decode(
+                hidden_states_batch, self.scale).to(data_device)
+
+            mask = self.build_mask(
+                hidden_states_batch,
+                is_bound=(h == 0, h_ >= H, w == 0, w_ >= W),
+                border_width=((size_h - stride_h) * self.upsampling_factor,
+                              (size_w - stride_w) * self.upsampling_factor)
+            ).to(dtype=hidden_states.dtype, device=data_device)
+
+            target_h = h * self.upsampling_factor
+            target_w = w * self.upsampling_factor
+            import pdb
+
+            # pdb.set_trace()
+            values[
+                :,
+                :,
+                :,
+                target_h:target_h + hidden_states_batch.shape[3],
+                target_w:target_w + hidden_states_batch.shape[4],
+            ] += hidden_states_batch * mask
+            weight[
+                :,
+                :,
+                :,
+                target_h: target_h + hidden_states_batch.shape[3],
+                target_w: target_w + hidden_states_batch.shape[4],
+            ] += mask
+        values = values / weight
+        values = values.clamp_(-1, 1)
+        return values
+
+    def tiled_encode(self, video, device, tile_size, tile_stride):
+        _, _, T, H, W = video.shape
+        size_h, size_w = tile_size
+        stride_h, stride_w = tile_stride
+
+        # Split tasks
+        tasks = []
+        for h in range(0, H, stride_h):
+            if (h-stride_h >= 0 and h-stride_h+size_h >= H):
+                continue
+            for w in range(0, W, stride_w):
+                if (w-stride_w >= 0 and w-stride_w+size_w >= W):
+                    continue
+                h_, w_ = h + size_h, w + size_w
+                tasks.append((h, h_, w, w_))
+
+        data_device = "cpu"
+        computation_device = device
+
+        out_T = (T + 3) // 4
+        weight = torch.zeros((1, 1, out_T, H // self.upsampling_factor, W //
+                             self.upsampling_factor), dtype=video.dtype, device=data_device)
+        values = torch.zeros((1, 16, out_T, H // self.upsampling_factor, W //
+                             self.upsampling_factor), dtype=video.dtype, device=data_device)
+        disable_flag = True
+        for h, h_, w, w_ in tqdm(tasks, desc="VAE encoding", disable=disable_flag):
+            hidden_states_batch = video[:, :, :,
+                                        h:h_, w:w_].to(computation_device)
+            hidden_states_batch = self.model.encode(
+                hidden_states_batch, self.scale).to(data_device)
+
+            mask = self.build_mask(
+                hidden_states_batch,
+                is_bound=(h == 0, h_ >= H, w == 0, w_ >= W),
+                border_width=((size_h - stride_h) // self.upsampling_factor,
+                              (size_w - stride_w) // self.upsampling_factor)
+            ).to(dtype=video.dtype, device=data_device)
+
+            target_h = h // self.upsampling_factor
+            target_w = w // self.upsampling_factor
+            values[
+                :,
+                :,
+                :,
+                target_h:target_h + hidden_states_batch.shape[3],
+                target_w:target_w + hidden_states_batch.shape[4],
+            ] += hidden_states_batch * mask
+            weight[
+                :,
+                :,
+                :,
+                target_h: target_h + hidden_states_batch.shape[3],
+                target_w: target_w + hidden_states_batch.shape[4],
+            ] += mask
+        values = values / weight
+        return values
+
+    def single_encode(self, video, device):
+        video = video.to(device)
+        x = self.model.encode(video, self.scale)
+        return x
+
+    def single_decode(self, hidden_state, device):
+        hidden_state = hidden_state.to(device)
+        video = self.model.decode(hidden_state, self.scale)
+        return video.clamp_(-1, 1)
+
+    def encode(self, videos, device, tiled=False, tile_size=(34, 34), tile_stride=(18, 16)):
+
+        # videos = [video.to("cpu") for video in videos]
+        hidden_states = []
+        for video in videos:
+            video = video.unsqueeze(0)
+            if tiled:
+                tile_size = (tile_size[0] * 8, tile_size[1] * 8)
+                tile_stride = (tile_stride[0] * 8, tile_stride[1] * 8)
+                hidden_state = self.tiled_encode(
+                    video, device, tile_size, tile_stride)
+            else:
+                hidden_state = self.single_encode(video, device)
+            hidden_state = hidden_state.squeeze(0)
+            hidden_states.append(hidden_state)
+        hidden_states = torch.stack(hidden_states)
+        # TODO
+        # if tiled:
+        #     hidden_states = self.tiled_encode(
+        #         videos, device, tile_size, tile_stride)
+        # else:
+        #     hidden_states = self.single_encode(videos, device)
+        return hidden_states
+
+    def decode(self, hidden_states, device, tiled=False, tile_size=(34, 34), tile_stride=(18, 16)):
+        video = []
+        for _hidden_states in hidden_states:
+            _hidden_states = _hidden_states.unsqueeze(0)
+            if tiled:
+                video.append(
+                    self.tiled_decode(_hidden_states, device, tile_size,
+                                      tile_stride))
+            else:
+                video.append(self.single_decode(_hidden_states, device))
+
+        video = torch.cat(video, dim=0)
+        # TODO
+        # if tiled:
+        #     video = self.tiled_decode(
+        #         hidden_states, device, tile_size, tile_stride)
+        # else:
+        #     video = self.single_decode(hidden_states, device)
+        return video
+
+    @staticmethod
+    def state_dict_converter():
+        return WanVideoVAEStateDictConverter()
+
+
+class WanVideoVAEStateDictConverter:
+
+    def __init__(self):
+        pass
+
+    def from_civitai(self, state_dict):
+        state_dict_ = {}
+        if 'model_state' in state_dict:
+            state_dict = state_dict['model_state']
+        for name in state_dict:
+            state_dict_['model.' + name] = state_dict[name]
+        return state_dict_

diffsynth/pipelines/__init__.py

@@ -0,0 +1 @@
+from .wan_video_new_determine import WanVideoPipeline

diffsynth/pipelines/wan_video_new_determine.py

@@ -0,0 +1,1730 @@
+import glob
+import os
+import time
+import types
+import warnings
+from dataclasses import dataclass
+from typing import List, Optional, Union
+
+import numpy as np
+import torch
+from einops import rearrange, reduce, repeat
+# from modelscope import snapshot_download
+from huggingface_hub import snapshot_download
+from PIL import Image
+from tqdm import tqdm
+from typing_extensions import Literal
+
+from ..models import ModelManager, load_state_dict
+from ..models.wan_video_dit import RMSNorm, WanModel, sinusoidal_embedding_1d
+from ..models.wan_video_image_encoder import WanImageEncoder
+from ..models.wan_video_motion_controller import WanMotionControllerModel
+# from ..model.
+from ..models.wan_video_text_encoder import (T5LayerNorm, T5RelativeEmbedding,
+                                             WanTextEncoder)
+from ..models.wan_video_vace import VaceWanModel
+from ..models.wan_video_vae import (CausalConv3d, RMS_norm, Upsample,
+                                    WanVideoVAE)
+from ..schedulers.flow_match import FlowMatchScheduler
+# from ..prompters import WanPrompter
+from ..vram_management import (AutoWrappedLinear, AutoWrappedModule,
+                               WanAutoCastLayerNorm, enable_vram_management)
+
+
+class BasePipeline(torch.nn.Module):
+
+    def __init__(
+        self,
+        device="cuda",
+        torch_dtype=torch.float16,
+        height_division_factor=64,
+        width_division_factor=64,
+        time_division_factor=None,
+        time_division_remainder=None,
+    ):
+        super().__init__()
+        # The device and torch_dtype is used for the storage of intermediate variables, not models.
+        self.device = device
+        self.torch_dtype = torch_dtype
+        # The following parameters are used for shape check.
+        self.height_division_factor = height_division_factor
+        self.width_division_factor = width_division_factor
+        self.time_division_factor = time_division_factor
+        self.time_division_remainder = time_division_remainder
+        self.vram_management_enabled = False
+
+    def to(self, *args, **kwargs):
+        device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(
+            *args, **kwargs
+        )
+        if device is not None:
+            self.device = device
+        if dtype is not None:
+            self.torch_dtype = dtype
+        super().to(*args, **kwargs)
+        return self
+
+    def check_resize_height_width(self, height, width, num_frames=None):
+        # Shape check
+        # print(
+        #     f"height, width, time division factor: {self.height_division_factor}, {self.width_division_factor}, {self.time_division_factor}, time division remainder: {self.time_division_remainder}"
+        # )
+        assert (
+            height % self.height_division_factor == 0
+        ), f"height {height} is not divisible by {self.height_division_factor}."
+
+        assert (
+            width % self.width_division_factor == 0
+        ), f"width {width} is not divisible by {self.width_division_factor}."
+        assert (num_frames is not None) and (
+            (num_frames + self.time_division_factor) % self.time_division_factor
+            == self.time_division_remainder
+        ), f"num_frames {num_frames} is not divisible by {self.time_division_factor} with remainder {self.time_division_remainder}."
+        return height, width, num_frames
+
+    def preprocess_image(
+        self,
+        image,
+        torch_dtype=None,
+        device=None,
+        pattern="B C H W",
+        min_value=-1,
+        max_value=1,
+    ):
+        # Transform a PIL.Image to torch.Tensor
+        # print(f"Image size: {image.size}, dtype: {image.mode}")
+        # assert isinstance(image, torch.Tensor), "Image must be a torch.Tensor."
+        # C H W
+        if isinstance(image, torch.Tensor):
+            # C H W
+            # print(f"Image shape {image.shape}")
+            assert (len(image.shape) == 3 and image.shape[0] == 3) or (
+                len(image.shape) == 4 and image.shape[1] == 3
+            ), "Image tensor must be in 3 H W or B 3 H W format."
+            image = image.to(
+                dtype=torch_dtype or self.torch_dtype, device=device or self.device
+            )
+            image = image * ((max_value - min_value)) + min_value
+            if len(image.shape) == 3:
+                image = image.unsqueeze(0)  # Add batch dimension
+        else:
+            image = torch.Tensor(np.array(image, dtype=np.float32))
+            image = image.to(
+                dtype=torch_dtype or self.torch_dtype, device=device or self.device
+            )
+            image = image * ((max_value - min_value) / 255) + min_value
+            image = repeat(
+                image, f"H W C -> {pattern}", **({"B": 1} if "B" in pattern else {})
+            )
+        return image
+
+    def preprocess_video(
+        self,
+        video,
+        torch_dtype=None,
+        device=None,
+        pattern="B C T H W",
+        min_value=-1,
+        max_value=1,
+    ):
+        video = [
+            self.preprocess_image(
+                image,
+                torch_dtype=torch_dtype,
+                device=device,
+                min_value=min_value,
+                max_value=max_value,
+            )
+            for image in video
+        ]
+        video = torch.stack(video, dim=pattern.index("T") // 2)
+        return video
+
+    def vae_output_to_image(
+        self, vae_output, pattern="B C H W", min_value=-1, max_value=1
+    ):
+        # Transform a torch.Tensor to PIL.Image
+        if pattern != "H W C":
+            vae_output = reduce(
+                vae_output, f"{pattern} -> H W C", reduction="mean")
+
+        # image = ((vae_output - min_value) * (255 / (max_value - min_value))).clip(
+        #     0, 255
+        # )
+        image = (vae_output - min_value) * (255.0 / (max_value - min_value))
+        image = image.clamp(0.0, 255.0)
+
+        image = image.to(device="cpu", dtype=torch.float32)
+        image = image.numpy()
+        # image = Image.fromarray(image.numpy())
+        return image
+
+    def vae_output_to_video(
+        self, vae_output, pattern="B C T H W", min_value=-1, max_value=1
+    ):
+        # Transform a torch.Tensor to list of PIL.Image
+        # if pattern != "T H W C":
+        #     vae_output = reduce(
+        #         vae_output, f"{pattern} -> T H W C", reduction="mean")
+        if vae_output.ndim == 5:  # B C T H W
+            assert (
+                vae_output.shape[1] == 3
+            ), f"vae_output shape {vae_output.shape} is not valid. Expected 5D tensor with 3 channels on the second dimension."
+            vae_output = vae_output.permute(0, 2, 3, 4, 1)
+            # print(f"vae_output shape after permute: {vae_output.shape}")
+            video = vae_output.to(device="cpu", dtype=torch.float32).numpy()
+            video = (video + 1.0) / 2.0
+            # print(f"Video range before clip: {video.min()} to {video.max()}")
+            video = video.clip(0.0, 1.0)
+
+        #     for _video in vae_output:
+        #         video.append(
+        #             [
+        #                 self.vae_output_to_image(
+        #                     image,
+        #                     pattern="H W C",
+        #                     min_value=min_value,
+        #                     max_value=max_value,
+        #                 )
+        #                 for image in _video
+        #             ]
+        #         )
+        # else:
+        #     raise ValueError(
+        #         f"Invalid vae_output shape {vae_output.shape}. Expected 5D tensor."
+        #     )
+        return video
+
+    def load_models_to_device(self, model_names=[]):
+        if self.vram_management_enabled:
+            # offload models
+            for name, model in self.named_children():
+                if name not in model_names:
+                    if (
+                        hasattr(model, "vram_management_enabled")
+                        and model.vram_management_enabled
+                    ):
+                        for module in model.modules():
+                            if hasattr(module, "offload"):
+                                module.offload()
+                    else:
+                        model.cpu()
+            torch.cuda.empty_cache()
+            # onload models
+            for name, model in self.named_children():
+                if name in model_names:
+                    if (
+                        hasattr(model, "vram_management_enabled")
+                        and model.vram_management_enabled
+                    ):
+                        for module in model.modules():
+                            if hasattr(module, "onload"):
+                                module.onload()
+                    else:
+                        model.to(self.device)
+
+    def generate_noise(
+        self,
+        shape,
+        seed=None,
+        rand_device="cpu",
+        rand_torch_dtype=torch.float32,
+        device=None,
+        torch_dtype=None,
+    ):
+        # Initialize Gaussian noise
+        generator = (
+            None if seed is None else torch.Generator(
+                rand_device).manual_seed(seed)
+        )
+        # TODO multi-res noise
+        noise = torch.randn(
+            shape, generator=generator, device=rand_device, dtype=rand_torch_dtype
+        )
+        noise = noise.to(
+            dtype=torch_dtype or self.torch_dtype, device=device or self.device
+        )
+        return noise
+
+    def enable_cpu_offload(self):
+        warnings.warn(
+            "`enable_cpu_offload` will be deprecated. Please use `enable_vram_management`."
+        )
+        self.vram_management_enabled = True
+
+    def get_vram(self):
+        return torch.cuda.mem_get_info(self.device)[1] / (1024**3)
+
+    def freeze_except(self, model_names):
+        for name, model in self.named_children():
+            if name in model_names:
+                print(f"Unfreezing model {name}.")
+                print(
+                    f"Model parameters: {sum(p.numel() for p in model.parameters())}")
+                model.train()
+                model.requires_grad_(True)
+            else:
+                print(f"Freezing model {name}.")
+                print(
+                    f"Model parameters: {sum(p.numel() for p in model.parameters())}")
+                model.eval()
+                model.requires_grad_(False)
+
+
+@dataclass
+class ModelConfig:
+    path: Union[str, list[str]] = None
+    model_id: str = None
+    origin_file_pattern: Union[str, list[str]] = None
+    download_resource: str = "ModelScope"
+    offload_device: Optional[Union[str, torch.device]] = None
+    offload_dtype: Optional[torch.dtype] = None
+
+    def download_if_necessary(
+        self, local_model_path="./models", skip_download=False, use_usp=False
+    ):
+        if self.path is None:
+            # Check model_id and origin_file_pattern
+            if self.model_id is None:
+                raise ValueError(
+                    f"""No valid model files. Please use `ModelConfig(path="xxx")` or `ModelConfig(model_id="xxx/yyy", origin_file_pattern="zzz")`."""
+                )
+
+            # Skip if not in rank 0
+            if use_usp:
+                import torch.distributed as dist
+
+                skip_download = dist.get_rank() != 0
+
+            # Check whether the origin path is a folder
+            if self.origin_file_pattern is None or self.origin_file_pattern == "":
+                self.origin_file_pattern = ""
+                allow_file_pattern = None
+                is_folder = True
+            elif isinstance(
+                self.origin_file_pattern, str
+            ) and self.origin_file_pattern.endswith("/"):
+                allow_file_pattern = self.origin_file_pattern + "*"
+                is_folder = True
+            else:
+                allow_file_pattern = self.origin_file_pattern
+                is_folder = False
+
+            # Download
+            if not skip_download:
+                downloaded_files = glob.glob(
+                    self.origin_file_pattern,
+                    root_dir=os.path.join(local_model_path, self.model_id),
+                )
+                # snapshot_download(
+                #     self.model_id,
+                #     local_dir=os.path.join(local_model_path, self.model_id),
+                #     allow_file_pattern=allow_file_pattern,
+                #     ignore_file_pattern=downloaded_files,
+                #     local_files_only=False,
+                # )
+                snapshot_download(
+                    self.model_id,
+                    repo_type="model",  # 如果是dataset要改成"dataset"
+                    local_dir=os.path.join(local_model_path, self.model_id),
+                    allow_patterns=allow_file_pattern,
+                    ignore_patterns=downloaded_files,    # 注意这里是 patterns
+                    # ignore_file_pattern=downloaded_files,
+                    # local_files_only=False,
+                    local_files_only=False,
+                    resume_download=True,   # 支持断点续传
+
+                )
+
+            # Let rank 1, 2, ... wait for rank 0
+            if use_usp:
+                import torch.distributed as dist
+
+                dist.barrier(device_ids=[dist.get_rank()])
+
+            # Return downloaded files
+            if is_folder:
+                self.path = os.path.join(
+                    local_model_path, self.model_id, self.origin_file_pattern
+                )
+            else:
+                self.path = glob.glob(
+                    os.path.join(
+                        local_model_path, self.model_id, self.origin_file_pattern
+                    )
+                )
+            if isinstance(self.path, list) and len(self.path) == 1:
+                self.path = self.path[0]
+
+
+class WanVideoPipeline(BasePipeline):
+
+    def __init__(self, device="cuda", torch_dtype=torch.bfloat16, tokenizer_path=None):
+        super().__init__(
+            device=device,
+            torch_dtype=torch_dtype,
+            height_division_factor=16,
+            width_division_factor=16,
+            time_division_factor=4,
+            time_division_remainder=1,
+        )
+        self.scheduler = FlowMatchScheduler(
+            shift=5, sigma_min=0.0, extra_one_step=True)
+        # self.prompter = WanPrompter(tokenizer_path=tokenizer_path)
+        self.text_encoder: WanTextEncoder = None
+        self.image_encoder: WanImageEncoder = None
+        # self.pose_encoder: CameraPoseEncoder = None
+        self.dit: WanModel = None
+        self.vae: WanVideoVAE = None
+        self.motion_controller: WanMotionControllerModel = None
+        self.vace: VaceWanModel = None
+        self.in_iteration_models = ("dit", "motion_controller", "vace")
+        self.unit_runner = PipelineUnitRunner()
+
+        self.units = [
+            WanVideoUnit_ShapeChecker(),  # check if the shape if ok
+            # WanVideoUnit_NoiseInitializer(),
+            WanVideoUnit_InputVideoEmbedder(),
+            WanVideoUnit_PromptEmbedder(),
+            WanVideoUnit_ImageEmbedder(),
+            # WanVideoUnit_FunReference(),
+            # WanVideoUnit_CameraPoseEmbedder(),
+            # WanVideoUnit_SpeedControl(),
+            # WanVideoUnit_VACE(),
+            WanVideoUnit_UnifiedSequenceParallel(),
+            # WanVideoUnit_TeaCache(),
+            # WanVideoUnit_CfgMerger(),
+        ]
+
+        self.model_fn = model_fn_wan_video
+
+    def training_predict(self, **inputs):
+        timestep_id = torch.tensor([0])
+        # print(f"timestep_id: {timestep_id}")
+        timestep = self.scheduler.timesteps[timestep_id].to(
+            dtype=self.torch_dtype, device=self.device
+        )
+        # print(f"Selected timestep {timestep}")
+        inputs["latents"] = inputs['rgb_latents']
+        training_target = self.scheduler.training_target(
+            inputs["depth_latents"], inputs["rgb_latents"], timestep
+        )
+        noise_pred = self.model_fn(**inputs, timestep=timestep)
+
+        return {
+            'rgb_gt': inputs['rgb_latents'],
+            "depth_gt": training_target,
+            "pred": noise_pred,
+            "weight": self.scheduler.training_weight(timestep),
+        }
+
+    def enable_vram_management(
+        self, num_persistent_param_in_dit=None, vram_limit=None, vram_buffer=0.5
+    ):
+        self.vram_management_enabled = True
+        if num_persistent_param_in_dit is not None:
+            vram_limit = None
+        else:
+            if vram_limit is None:
+                vram_limit = self.get_vram()
+            vram_limit = vram_limit - vram_buffer
+        if self.text_encoder is not None:
+            dtype = next(iter(self.text_encoder.parameters())).dtype
+            enable_vram_management(
+                self.text_encoder,
+                module_map={
+                    torch.nn.Linear: AutoWrappedLinear,
+                    torch.nn.Embedding: AutoWrappedModule,
+                    T5RelativeEmbedding: AutoWrappedModule,
+                    T5LayerNorm: AutoWrappedModule,
+                },
+                module_config=dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device="cpu",
+                    computation_dtype=self.torch_dtype,
+                    computation_device=self.device,
+                ),
+                vram_limit=vram_limit,
+            )
+        if self.dit is not None:
+            dtype = next(iter(self.dit.parameters())).dtype
+            device = "cpu" if vram_limit is not None else self.device
+            enable_vram_management(
+                self.dit,
+                module_map={
+                    torch.nn.Linear: AutoWrappedLinear,
+                    torch.nn.Conv3d: AutoWrappedModule,
+                    torch.nn.LayerNorm: WanAutoCastLayerNorm,
+                    RMSNorm: AutoWrappedModule,
+                    torch.nn.Conv2d: AutoWrappedModule,
+                },
+                module_config=dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device=device,
+                    computation_dtype=self.torch_dtype,
+                    computation_device=self.device,
+                ),
+                max_num_param=num_persistent_param_in_dit,
+                overflow_module_config=dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device="cpu",
+                    computation_dtype=self.torch_dtype,
+                    computation_device=self.device,
+                ),
+                vram_limit=vram_limit,
+            )
+        if self.vae is not None:
+            dtype = next(iter(self.vae.parameters())).dtype
+            enable_vram_management(
+                self.vae,
+                module_map={
+                    torch.nn.Linear: AutoWrappedLinear,
+                    torch.nn.Conv2d: AutoWrappedModule,
+                    RMS_norm: AutoWrappedModule,
+                    CausalConv3d: AutoWrappedModule,
+                    Upsample: AutoWrappedModule,
+                    torch.nn.SiLU: AutoWrappedModule,
+                    torch.nn.Dropout: AutoWrappedModule,
+                },
+                module_config=dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device=self.device,
+                    computation_dtype=self.torch_dtype,
+                    computation_device=self.device,
+                ),
+            )
+        if self.image_encoder is not None:
+            dtype = next(iter(self.image_encoder.parameters())).dtype
+            enable_vram_management(
+                self.image_encoder,
+                module_map={
+                    torch.nn.Linear: AutoWrappedLinear,
+                    torch.nn.Conv2d: AutoWrappedModule,
+                    torch.nn.LayerNorm: AutoWrappedModule,
+                },
+                module_config=dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device="cpu",
+                    computation_dtype=dtype,
+                    computation_device=self.device,
+                ),
+            )
+        if self.motion_controller is not None:
+            dtype = next(iter(self.motion_controller.parameters())).dtype
+            enable_vram_management(
+                self.motion_controller,
+                module_map={
+                    torch.nn.Linear: AutoWrappedLinear,
+                },
+                module_config=dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device="cpu",
+                    computation_dtype=dtype,
+                    computation_device=self.device,
+                ),
+            )
+        if self.vace is not None:
+            device = "cpu" if vram_limit is not None else self.device
+            enable_vram_management(
+                self.vace,
+                module_map={
+                    torch.nn.Linear: AutoWrappedLinear,
+                    torch.nn.Conv3d: AutoWrappedModule,
+                    torch.nn.LayerNorm: AutoWrappedModule,
+                    RMSNorm: AutoWrappedModule,
+                },
+                module_config=dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device=device,
+                    computation_dtype=self.torch_dtype,
+                    computation_device=self.device,
+                ),
+                vram_limit=vram_limit,
+            )
+
+    def initialize_usp(self):
+        import torch.distributed as dist
+        from xfuser.core.distributed import (init_distributed_environment,
+                                             initialize_model_parallel)
+
+        dist.init_process_group(backend="nccl", init_method="env://")
+        init_distributed_environment(
+            rank=dist.get_rank(), world_size=dist.get_world_size()
+        )
+        initialize_model_parallel(
+            sequence_parallel_degree=dist.get_world_size(),
+            ring_degree=1,
+            ulysses_degree=dist.get_world_size(),
+        )
+        torch.cuda.set_device(dist.get_rank())
+
++