diffsynth 1.0.0__py3-none-any.whl

diffsynth/__init__.py

@@ -0,0 +1,6 @@
+from .data import *
+from .models import *
+from .prompters import *
+from .schedulers import *
+from .pipelines import *
+from .controlnets import *

diffsynth/configs/model_config.py

@@ -0,0 +1,243 @@
+from typing_extensions import Literal, TypeAlias
+
+from ..models.sd_text_encoder import SDTextEncoder
+from ..models.sd_unet import SDUNet
+from ..models.sd_vae_encoder import SDVAEEncoder
+from ..models.sd_vae_decoder import SDVAEDecoder
+
+from ..models.sdxl_text_encoder import SDXLTextEncoder, SDXLTextEncoder2
+from ..models.sdxl_unet import SDXLUNet
+from ..models.sdxl_vae_decoder import SDXLVAEDecoder
+from ..models.sdxl_vae_encoder import SDXLVAEEncoder
+
+from ..models.sd3_text_encoder import SD3TextEncoder1, SD3TextEncoder2, SD3TextEncoder3
+from ..models.sd3_dit import SD3DiT
+from ..models.sd3_vae_decoder import SD3VAEDecoder
+from ..models.sd3_vae_encoder import SD3VAEEncoder
+
+from ..models.sd_controlnet import SDControlNet
+
+from ..models.sd_motion import SDMotionModel
+from ..models.sdxl_motion import SDXLMotionModel
+
+from ..models.svd_image_encoder import SVDImageEncoder
+from ..models.svd_unet import SVDUNet
+from ..models.svd_vae_decoder import SVDVAEDecoder
+from ..models.svd_vae_encoder import SVDVAEEncoder
+
+from ..models.sd_ipadapter import SDIpAdapter, IpAdapterCLIPImageEmbedder
+from ..models.sdxl_ipadapter import SDXLIpAdapter, IpAdapterXLCLIPImageEmbedder
+
+from ..models.hunyuan_dit_text_encoder import HunyuanDiTCLIPTextEncoder, HunyuanDiTT5TextEncoder
+from ..models.hunyuan_dit import HunyuanDiT
+
+
+
+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, "091b0e30e77c76626b3ba62acdf95343", ["sd_controlnet"], [SDControlNet], "civitai"),
+    (None, "4a6c8306a27d916dea81263c8c88f450", ["hunyuan_dit_clip_text_encoder"], [HunyuanDiTCLIPTextEncoder], "civitai"),
+    (None, "f4aec400fe394297961218c768004521", ["hunyuan_dit"], [HunyuanDiT], "civitai"),
+    (None, "9e6e58043a5a2e332803ed42f6ee7181", ["hunyuan_dit_t5_text_encoder"], [HunyuanDiTT5TextEncoder], "civitai"),
+    (None, "13115dd45a6e1c39860f91ab073b8a78", ["sdxl_vae_encoder", "sdxl_vae_decoder"], [SDXLVAEEncoder, SDXLVAEDecoder], "diffusers"),
+    (None, "d78aa6797382a6d455362358a3295ea9", ["sd_ipadapter_clip_image_encoder"], [IpAdapterCLIPImageEmbedder], "diffusers"),
+    (None, "e291636cc15e803186b47404262ef812", ["sd_ipadapter"], [SDIpAdapter], "civitai"),
+    (None, "399c81f2f8de8d1843d0127a00f3c224", ["sdxl_ipadapter_clip_image_encoder"], [IpAdapterXLCLIPImageEmbedder], "diffusers"),
+    (None, "a64eac9aa0db4b9602213bc0131281c7", ["sdxl_ipadapter"], [SDXLIpAdapter], "civitai"),
+    (None, "52817e4fdd89df154f02749ca6f692ac", ["sdxl_unet"], [SDXLUNet], "diffusers"),
+    (None, "03343c606f16d834d6411d0902b53636", ["sd_text_encoder", "sd_unet", "sd_vae_decoder", "sd_vae_encoder"], [SDTextEncoder, SDUNet, SDVAEDecoder, SDVAEEncoder], "civitai"),
+    (None, "d4ba77a7ece070679b4a987f58f201e9", ["sd_text_encoder"], [SDTextEncoder], "civitai"),
+    (None, "d0c89e55c5a57cf3981def0cb1c9e65a", ["sd_vae_decoder", "sd_vae_encoder"], [SDVAEDecoder, SDVAEEncoder], "civitai"),
+    (None, "3926bf373b39a67eeafd7901478a47a7", ["sd_unet"], [SDUNet], "civitai"),
+    (None, "1e0c39ec176b9007c05f76d52b554a4d", ["sd3_text_encoder_1", "sd3_text_encoder_2", "sd3_dit", "sd3_vae_encoder", "sd3_vae_decoder"], [SD3TextEncoder1, SD3TextEncoder2, SD3DiT, SD3VAEEncoder, SD3VAEDecoder], "civitai"),
+    (None, "d9e0290829ba8d98e28e1a2b1407db4a", ["sd3_text_encoder_1", "sd3_text_encoder_2", "sd3_text_encoder_3", "sd3_dit", "sd3_vae_encoder", "sd3_vae_decoder"], [SD3TextEncoder1, SD3TextEncoder2, SD3TextEncoder3, SD3DiT, SD3VAEEncoder, SD3VAEDecoder], "civitai"),
+    (None, "5072d0b24e406b49507abe861cf97691", ["sd3_text_encoder_3"], [SD3TextEncoder3], "civitai"),
+    (None, "4cf64a799d04260df438c6f33c9a047e", ["sdxl_text_encoder", "sdxl_text_encoder_2", "sdxl_unet", "sdxl_vae_decoder", "sdxl_vae_encoder"], [SDXLTextEncoder, SDXLTextEncoder2, SDXLUNet, SDXLVAEDecoder, SDXLVAEEncoder], "civitai"),
+    (None, "d9b008a867c498ab12ad24042eff8e3f", ["sdxl_text_encoder", "sdxl_text_encoder_2", "sdxl_unet", "sdxl_vae_decoder", "sdxl_vae_encoder"], [SDXLTextEncoder, SDXLTextEncoder2, SDXLUNet, SDXLVAEDecoder, SDXLVAEEncoder], "civitai"), # SDXL-Turbo
+    (None, "025bb7452e531a3853d951d77c63f032", ["sdxl_text_encoder", "sdxl_text_encoder_2"], [SDXLTextEncoder, SDXLTextEncoder2], "civitai"),
+    (None, "298997b403a4245c04102c9f36aac348", ["sdxl_unet"], [SDXLUNet], "civitai"),
+    (None, "2a07abce74b4bdc696b76254ab474da6", ["svd_image_encoder", "svd_unet", "svd_vae_decoder", "svd_vae_encoder"], [SVDImageEncoder, SVDUNet, SVDVAEDecoder, SVDVAEEncoder], "civitai"),
+    (None, "c96a285a6888465f87de22a984d049fb", ["sd_motion_modules"], [SDMotionModel], "civitai"),
+    (None, "72907b92caed19bdb2adb89aa4063fe2", ["sdxl_motion_modules"], [SDXLMotionModel], "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)
+    ("ChatGLMModel", "diffsynth.models.kolors_text_encoder", "kolors_text_encoder"),
+    ("MarianMTModel", "transformers.models.marian.modeling_marian", "translator"),
+    ("BloomForCausalLM", "transformers.models.bloom.modeling_bloom", "beautiful_prompt"),
+]
+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)
+    ("9a4ab6869ac9b7d6e31f9854e397c867", ["svd_unet"], [SVDUNet], {"add_positional_conv": 128}),
+]
+
+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"),
+    ],
+}
+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"),
+    ],
+    # 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"),
+    ],
+    # 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")
+    ],
+    # 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"),
+    ],
+    # Beautiful Prompt
+    "BeautifulPrompt": [
+        ("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"),
+    ],
+    # Translator
+    "opus-mt-zh-en": [
+        ("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"),
+    ],
+    # 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": [
+        ("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"),
+    ],
+    "SDXL-vae-fp16-fix": [
+        ("AI-ModelScope/sdxl-vae-fp16-fix", "diffusion_pytorch_model.safetensors", "models/sdxl-vae-fp16-fix")
+    ],
+}
+Preset_model_id: TypeAlias = Literal[
+    "HunyuanDiT",
+    "stable-video-diffusion-img2vid-xt",
+    "ExVideo-SVD-128f-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",
+]

diffsynth/controlnets/__init__.py

@@ -0,0 +1,2 @@
+from .controlnet_unit import ControlNetConfigUnit, ControlNetUnit, MultiControlNetManager
+from .processors import Annotator

diffsynth/controlnets/controlnet_unit.py

@@ -0,0 +1,53 @@
+import torch
+import numpy as np
+from .processors import Processor_id
+
+
+class ControlNetConfigUnit:
+    def __init__(self, processor_id: Processor_id, model_path, scale=1.0):
+        self.processor_id = processor_id
+        self.model_path = model_path
+        self.scale = scale
+
+
+class ControlNetUnit:
+    def __init__(self, processor, model, scale=1.0):
+        self.processor = processor
+        self.model = model
+        self.scale = scale
+
+
+class MultiControlNetManager:
+    def __init__(self, controlnet_units=[]):
+        self.processors = [unit.processor for unit in controlnet_units]
+        self.models = [unit.model for unit in controlnet_units]
+        self.scales = [unit.scale for unit in controlnet_units]
+
+    def process_image(self, image, processor_id=None):
+        if processor_id is None:
+            processed_image = [processor(image) for processor in self.processors]
+        else:
+            processed_image = [self.processors[processor_id](image)]
+        processed_image = torch.concat([
+            torch.Tensor(np.array(image_, dtype=np.float32) / 255).permute(2, 0, 1).unsqueeze(0)
+            for image_ in processed_image
+        ], dim=0)
+        return processed_image
+    
+    def __call__(
+        self,
+        sample, timestep, encoder_hidden_states, conditionings,
+        tiled=False, tile_size=64, tile_stride=32
+    ):
+        res_stack = None
+        for conditioning, model, scale in zip(conditionings, self.models, self.scales):
+            res_stack_ = model(
+                sample, timestep, encoder_hidden_states, conditioning,
+                tiled=tiled, tile_size=tile_size, tile_stride=tile_stride
+            )
+            res_stack_ = [res * scale for res in res_stack_]
+            if res_stack is None:
+                res_stack = res_stack_
+            else:
+                res_stack = [i + j for i, j in zip(res_stack, res_stack_)]
+        return res_stack

diffsynth/controlnets/processors.py

@@ -0,0 +1,51 @@
+from typing_extensions import Literal, TypeAlias
+import warnings
+with warnings.catch_warnings():
+    warnings.simplefilter("ignore")
+    from controlnet_aux.processor import (
+        CannyDetector, MidasDetector, HEDdetector, LineartDetector, LineartAnimeDetector, OpenposeDetector
+    )
+
+
+Processor_id: TypeAlias = Literal[
+    "canny", "depth", "softedge", "lineart", "lineart_anime", "openpose", "tile"
+]
+
+class Annotator:
+    def __init__(self, processor_id: Processor_id, model_path="models/Annotators", detect_resolution=None, device='cuda'):
+        if processor_id == "canny":
+            self.processor = CannyDetector()
+        elif processor_id == "depth":
+            self.processor = MidasDetector.from_pretrained(model_path).to(device)
+        elif processor_id == "softedge":
+            self.processor = HEDdetector.from_pretrained(model_path).to(device)
+        elif processor_id == "lineart":
+            self.processor = LineartDetector.from_pretrained(model_path).to(device)
+        elif processor_id == "lineart_anime":
+            self.processor = LineartAnimeDetector.from_pretrained(model_path).to(device)
+        elif processor_id == "openpose":
+            self.processor = OpenposeDetector.from_pretrained(model_path).to(device)
+        elif processor_id == "tile":
+            self.processor = None
+        else:
+            raise ValueError(f"Unsupported processor_id: {processor_id}")
+
+        self.processor_id = processor_id
+        self.detect_resolution = detect_resolution
+
+    def __call__(self, image):
+        width, height = image.size
+        if self.processor_id == "openpose":
+            kwargs = {
+                "include_body": True,
+                "include_hand": True,
+                "include_face": True
+            }
+        else:
+            kwargs = {}
+        if self.processor is not None:
+            detect_resolution = self.detect_resolution if self.detect_resolution is not None else min(width, height)
+            image = self.processor(image, detect_resolution=detect_resolution, image_resolution=min(width, height), **kwargs)
+        image = image.resize((width, height))
+        return image
+

diffsynth/data/__init__.py

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

diffsynth/data/simple_text_image.py

@@ -0,0 +1,35 @@
+import torch, os
+from torchvision import transforms
+import pandas as pd
+from PIL import Image
+
+
+
+class TextImageDataset(torch.utils.data.Dataset):
+    def __init__(self, dataset_path, steps_per_epoch=10000, height=1024, width=1024, center_crop=True, random_flip=False):
+        self.steps_per_epoch = steps_per_epoch
+        metadata = pd.read_csv(os.path.join(dataset_path, "train/metadata.csv"))
+        self.path = [os.path.join(dataset_path, "train", file_name) for file_name in metadata["file_name"]]
+        self.text = metadata["text"].to_list()
+        self.image_processor = transforms.Compose(
+            [
+                transforms.Resize(max(height, width), interpolation=transforms.InterpolationMode.BILINEAR),
+                transforms.CenterCrop((height, width)) if center_crop else transforms.RandomCrop((height, width)),
+                transforms.RandomHorizontalFlip() if random_flip else transforms.Lambda(lambda x: x),
+                transforms.ToTensor(),
+                transforms.Normalize([0.5], [0.5]),
+            ]
+        )
+
+
+    def __getitem__(self, index):
+        data_id = torch.randint(0, len(self.path), (1,))[0]
+        data_id = (data_id + index) % len(self.path) # For fixed seed.
+        text = self.text[data_id]
+        image = Image.open(self.path[data_id]).convert("RGB")
+        image = self.image_processor(image)
+        return {"text": text, "image": image}
+
+
+    def __len__(self):
+        return self.steps_per_epoch

diffsynth/data/video.py

@@ -0,0 +1,148 @@
+import imageio, os
+import numpy as np
+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(frames, save_path, fps, quality=9):
+    writer = imageio.get_writer(save_path, fps=fps, quality=quality)
+    for frame in tqdm(frames, desc="Saving video"):
+        frame = np.array(frame)
+        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"))

diffsynth/extensions/ESRGAN/__init__.py

@@ -0,0 +1,118 @@
+import torch
+from einops import repeat
+from PIL import Image
+import numpy as np
+
+
+class ResidualDenseBlock(torch.nn.Module):
+
+    def __init__(self, num_feat=64, num_grow_ch=32):
+        super(ResidualDenseBlock, self).__init__()
+        self.conv1 = torch.nn.Conv2d(num_feat, num_grow_ch, 3, 1, 1)
+        self.conv2 = torch.nn.Conv2d(num_feat + num_grow_ch, num_grow_ch, 3, 1, 1)
+        self.conv3 = torch.nn.Conv2d(num_feat + 2 * num_grow_ch, num_grow_ch, 3, 1, 1)
+        self.conv4 = torch.nn.Conv2d(num_feat + 3 * num_grow_ch, num_grow_ch, 3, 1, 1)
+        self.conv5 = torch.nn.Conv2d(num_feat + 4 * num_grow_ch, num_feat, 3, 1, 1)
+        self.lrelu = torch.nn.LeakyReLU(negative_slope=0.2, inplace=True)
+
+    def forward(self, x):
+        x1 = self.lrelu(self.conv1(x))
+        x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1)))
+        x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1)))
+        x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1)))
+        x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
+        return x5 * 0.2 + x
+
+
+class RRDB(torch.nn.Module):
+
+    def __init__(self, num_feat, num_grow_ch=32):
+        super(RRDB, self).__init__()
+        self.rdb1 = ResidualDenseBlock(num_feat, num_grow_ch)
+        self.rdb2 = ResidualDenseBlock(num_feat, num_grow_ch)
+        self.rdb3 = ResidualDenseBlock(num_feat, num_grow_ch)
+
+    def forward(self, x):
+        out = self.rdb1(x)
+        out = self.rdb2(out)
+        out = self.rdb3(out)
+        return out * 0.2 + x
+
+
+class RRDBNet(torch.nn.Module):
+
+    def __init__(self, num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32):
+        super(RRDBNet, self).__init__()
+        self.conv_first = torch.nn.Conv2d(num_in_ch, num_feat, 3, 1, 1)
+        self.body = torch.torch.nn.Sequential(*[RRDB(num_feat=num_feat, num_grow_ch=num_grow_ch) for _ in range(num_block)])
+        self.conv_body = torch.nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        # upsample
+        self.conv_up1 = torch.nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_up2 = torch.nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_hr = torch.nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_last = torch.nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+        self.lrelu = torch.nn.LeakyReLU(negative_slope=0.2, inplace=True)
+
+    def forward(self, x):
+        feat = x
+        feat = self.conv_first(feat)
+        body_feat = self.conv_body(self.body(feat))
+        feat = feat + body_feat
+        # upsample
+        feat = repeat(feat, "B C H W -> B C (H 2) (W 2)")
+        feat = self.lrelu(self.conv_up1(feat))
+        feat = repeat(feat, "B C H W -> B C (H 2) (W 2)")
+        feat = self.lrelu(self.conv_up2(feat))
+        out = self.conv_last(self.lrelu(self.conv_hr(feat)))
+        return out
+
+
+class ESRGAN(torch.nn.Module):
+    def __init__(self, model):
+        super().__init__()
+        self.model = model
+
+    @staticmethod
+    def from_pretrained(model_path):
+        model = RRDBNet()
+        state_dict = torch.load(model_path, map_location="cpu")["params_ema"]
+        model.load_state_dict(state_dict)
+        model.eval()
+        return ESRGAN(model)
+
+    def process_image(self, image):
+        image = torch.Tensor(np.array(image, dtype=np.float32) / 255).permute(2, 0, 1)
+        return image
+    
+    def process_images(self, images):
+        images = [self.process_image(image) for image in images]
+        images = torch.stack(images)
+        return images
+    
+    def decode_images(self, images):
+        images = (images.permute(0, 2, 3, 1) * 255).clip(0, 255).numpy().astype(np.uint8)
+        images = [Image.fromarray(image) for image in images]
+        return images
+    
+    @torch.no_grad()
+    def upscale(self, images, batch_size=4, progress_bar=lambda x:x):
+        # Preprocess
+        input_tensor = self.process_images(images)
+
+        # Interpolate
+        output_tensor = []
+        for batch_id in progress_bar(range(0, input_tensor.shape[0], batch_size)):
+            batch_id_ = min(batch_id + batch_size, input_tensor.shape[0])
+            batch_input_tensor = input_tensor[batch_id: batch_id_]
+            batch_input_tensor = batch_input_tensor.to(
+                device=self.model.conv_first.weight.device,
+                dtype=self.model.conv_first.weight.dtype)
+            batch_output_tensor = self.model(batch_input_tensor)
+            output_tensor.append(batch_output_tensor.cpu())
+        
+        # Output
+        output_tensor = torch.concat(output_tensor, dim=0)
+
+        # To images
+        output_images = self.decode_images(output_tensor)
+        return output_images