import os import random import sys from typing import Sequence, Mapping, Any, Union import spaces import torch import gradio as gr from huggingface_hub import hf_hub_download from comfy import model_management from PIL import Image # --- Helper Functions from original script --- def get_value_at_index(obj: Union[Sequence, Mapping], index: int) -> Any: try: return obj[index] except KeyError: return obj["result"][index] def find_path(name: str, path: str = None) -> str: if path is None: path = os.getcwd() if name in os.listdir(path): path_name = os.path.join(path, name) print(f"{name} found: {path_name}") return path_name parent_directory = os.path.dirname(path) if parent_directory == path: return None return find_path(name, parent_directory) def add_comfyui_directory_to_sys_path() -> None: comfyui_path = find_path("ComfyUI") if comfyui_path is not None and os.path.isdir(comfyui_path): sys.path.append(comfyui_path) print(f"'{comfyui_path}' added to sys.path") def add_extra_model_paths() -> None: try: from main import load_extra_path_config except ImportError: from utils.extra_config import load_extra_path_config extra_model_paths = find_path("extra_model_paths.yaml") if extra_model_paths is not None: load_extra_path_config(extra_model_paths) else: print("Could not find the extra_model_paths config file.") def import_custom_nodes() -> None: import asyncio import execution from nodes import init_extra_nodes import server loop = asyncio.new_event_loop() asyncio.set_event_loop(loop) server_instance = server.PromptServer(loop) execution.PromptQueue(server_instance) init_extra_nodes() # --- Setup and Model Downloads --- add_comfyui_directory_to_sys_path() add_extra_model_paths() import_custom_nodes() from nodes import NODE_CLASS_MAPPINGS print("Downlading models from Hugging Face Hub...") # Text Encoder hf_hub_download(repo_id="Comfy-Org/Wan_2.1_ComfyUI_repackaged", filename="split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors", local_dir="models/text_encoders") # UNETs hf_hub_download(repo_id="Comfy-Org/Wan_2.2_ComfyUI_Repackaged", filename="split_files/diffusion_models/wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors", local_dir="models/unet") hf_hub_download(repo_id="Comfy-Org/Wan_2.2_ComfyUI_Repackaged", filename="split_files/diffusion_models/wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors", local_dir="models/unet") # VAE hf_hub_download(repo_id="Comfy-Org/Wan_2.1_ComfyUI_repackaged", filename="split_files/vae/wan_2.1_vae.safetensors", local_dir="models/vae") # CLIP Vision hf_hub_download(repo_id="Comfy-Org/Wan_2.1_ComfyUI_repackaged", filename="split_files/clip_vision/clip_vision_h.safetensors", local_dir="models/clip_vision") # LoRAs hf_hub_download(repo_id="Kijai/WanVideo_comfy", filename="Wan22-Lightning/Wan2.2-Lightning_I2V-A14B-4steps-lora_HIGH_fp16.safetensors", local_dir="models/loras") hf_hub_download(repo_id="Kijai/WanVideo_comfy", filename="Wan22-Lightning/Wan2.2-Lightning_I2V-A14B-4steps-lora_LOW_fp16.safetensors", local_dir="models/loras") print("Downloads complete.") # --- ZeroGPU: Pre-load models and instantiate nodes globally --- # Instantiate Nodes cliploader = NODE_CLASS_MAPPINGS["CLIPLoader"]() cliptextencode = NODE_CLASS_MAPPINGS["CLIPTextEncode"]() unetloader = NODE_CLASS_MAPPINGS["UNETLoader"]() vaeloader = NODE_CLASS_MAPPINGS["VAELoader"]() clipvisionloader = NODE_CLASS_MAPPINGS["CLIPVisionLoader"]() loadimage = NODE_CLASS_MAPPINGS["LoadImage"]() clipvisionencode = NODE_CLASS_MAPPINGS["CLIPVisionEncode"]() loraloadermodelonly = NODE_CLASS_MAPPINGS["LoraLoaderModelOnly"]() modelsamplingsd3 = NODE_CLASS_MAPPINGS["ModelSamplingSD3"]() pathchsageattentionkj = NODE_CLASS_MAPPINGS["PathchSageAttentionKJ"]() wanfirstlastframetovideo = NODE_CLASS_MAPPINGS["WanFirstLastFrameToVideo"]() ksampleradvanced = NODE_CLASS_MAPPINGS["KSamplerAdvanced"]() vaedecode = NODE_CLASS_MAPPINGS["VAEDecode"]() createvideo = NODE_CLASS_MAPPINGS["CreateVideo"]() savevideo = NODE_CLASS_MAPPINGS["SaveVideo"]() imageresize = NODE_CLASS_MAPPINGS["ImageResize+"]() # For dynamic resizing # Load Models cliploader_38 = cliploader.load_clip(clip_name="umt5_xxl_fp8_e4m3fn_scaled.safetensors", type="wan", device="cpu") unetloader_37_low_noise = unetloader.load_unet(unet_name="wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors", weight_dtype="default") unetloader_91_high_noise = unetloader.load_unet(unet_name="wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors", weight_dtype="default") vaeloader_39 = vaeloader.load_vae(vae_name="wan_2.1_vae.safetensors") clipvisionloader_49 = clipvisionloader.load_clip(clip_name="clip_vision_h.safetensors") # Apply LoRAs and Patches loraloadermodelonly_94_high = loraloadermodelonly.load_lora_model_only(lora_name="Wan2.2-Lightning_I2V-A14B-4steps-lora_HIGH_fp16.safetensors", strength_model=0.8, model=get_value_at_index(unetloader_91_high_noise, 0)) loraloadermodelonly_95_low = loraloadermodelonly.load_lora_model_only(lora_name="Wan2.2-Lightning_I2V-A14B-4steps-lora_LOW_fp16.safetensors", strength_model=0.8, model=get_value_at_index(unetloader_37_low_noise, 0)) modelsamplingsd3_93_low = modelsamplingsd3.patch(shift=8, model=get_value_at_index(loraloadermodelonly_95_low, 0)) pathchsageattentionkj_98_low = pathchsageattentionkj.patch(sage_attention="auto", model=get_value_at_index(modelsamplingsd3_93_low, 0)) modelsamplingsd3_79_high = modelsamplingsd3.patch(shift=8, model=get_value_at_index(loraloadermodelonly_94_high, 0)) pathchsageattentionkj_96_high = pathchsageattentionkj.patch(sage_attention="auto", model=get_value_at_index(modelsamplingsd3_79_high, 0)) # Pre-load models to GPU model_loaders = [cliploader_38, unetloader_37_low_noise, unetloader_91_high_noise, vaeloader_39, clipvisionloader_49, loraloadermodelonly_94_high, loraloadermodelonly_95_low] valid_models = [getattr(loader[0], 'patcher', loader[0]) for loader in model_loaders if not isinstance(loader[0], dict) and not isinstance(getattr(loader[0], 'patcher', None), dict)] model_management.load_models_gpu(valid_models) # --- Custom Logic for this App --- def calculate_dimensions(image_path): with Image.open(image_path) as img: width, height = img.size if width == height: return 480, 480 if width > height: new_width = 832 new_height = int(height * (832 / width)) else: new_height = 832 new_width = int(width * (832 / height)) # Ensure dimensions are multiples of 16 new_width = (new_width // 16) * 16 new_height = (new_height // 16) * 16 return new_width, new_height # --- Main Generation Function --- @spaces.GPU(duration=120) def generate_video(prompt, first_image_path, last_image_path): # This function now only handles per-request logic with torch.inference_mode(): # Calculate target dimensions based on the first image target_width, target_height = calculate_dimensions(first_image_path) # 1. Load and resize images # Since LoadImage returns a tensor, we pass it to the resize node loaded_first_image = loadimage.load_image(image=first_image_path) resized_first_image = imageresize.execute( width=target_width, height=target_height, interpolation="bicubic", method="stretch", condition="always", multiple_of=1, image=get_value_at_index(loaded_first_image, 0) ) loaded_last_image = loadimage.load_image(image=last_image_path) resized_last_image = imageresize.execute( width=target_width, height=target_height, interpolation="bicubic", method="stretch", condition="always", multiple_of=1, image=get_value_at_index(loaded_last_image, 0) ) # 2. Encode text and images cliptextencode_6 = cliptextencode.encode(text=prompt, clip=get_value_at_index(cliploader_38, 0)) cliptextencode_7_negative = cliptextencode.encode( text="low quality, worst quality, jpeg artifacts, ugly, deformed, blurry", clip=get_value_at_index(cliploader_38, 0), ) clipvisionencode_51 = clipvisionencode.encode(crop="none", clip_vision=get_value_at_index(clipvisionloader_49, 0), image=get_value_at_index(resized_first_image, 0)) clipvisionencode_87 = clipvisionencode.encode(crop="none", clip_vision=get_value_at_index(clipvisionloader_49, 0), image=get_value_at_index(resized_last_image, 0)) # 3. Prepare latents for video generation wanfirstlastframetovideo_83 = wanfirstlastframetovideo.EXECUTE_NORMALIZED( width=target_width, height=target_height, length=33, batch_size=1, positive=get_value_at_index(cliptextencode_6, 0), negative=get_value_at_index(cliptextencode_7_negative, 0), vae=get_value_at_index(vaeloader_39, 0), clip_vision_start_image=get_value_at_index(clipvisionencode_51, 0), clip_vision_end_image=get_value_at_index(clipvisionencode_87, 0), start_image=get_value_at_index(resized_first_image, 0), end_image=get_value_at_index(resized_last_image, 0), ) # 4. KSampler pipeline ksampleradvanced_101 = ksampleradvanced.sample( add_noise="enable", noise_seed=random.randint(1, 2**64), steps=8, cfg=1, sampler_name="euler", scheduler="simple", start_at_step=0, end_at_step=4, return_with_leftover_noise="enable", model=get_value_at_index(pathchsageattentionkj_96_high, 0), positive=get_value_at_index(wanfirstlastframetovideo_83, 0), negative=get_value_at_index(wanfirstlastframetovideo_83, 1), latent_image=get_value_at_index(wanfirstlastframetovideo_83, 2), ) ksampleradvanced_102 = ksampleradvanced.sample( add_noise="disable", noise_seed=random.randint(1, 2**64), steps=8, cfg=1, sampler_name="euler", scheduler="simple", start_at_step=4, end_at_step=10000, return_with_leftover_noise="disable", model=get_value_at_index(pathchsageattentionkj_98_low, 0), positive=get_value_at_index(wanfirstlastframetovideo_83, 0), negative=get_value_at_index(wanfirstlastframetovideo_83, 1), latent_image=get_value_at_index(ksampleradvanced_101, 0), ) # 5. Decode and save video vaedecode_8 = vaedecode.decode(samples=get_value_at_index(ksampleradvanced_102, 0), vae=get_value_at_index(vaeloader_39, 0)) createvideo_104 = createvideo.create_video(fps=16, images=get_value_at_index(vaedecode_8, 0)) savevideo_103 = savevideo.save_video(filename_prefix="ComfyUI_Video", format="mp4", codec="libx264", video=get_value_at_index(createvideo_104, 0)) # Return the path to the saved video video_filename = savevideo_103['ui']['videos'][0]['filename'] return f"output/{video_filename}" # --- Gradio Interface --- with gr.Blocks() as app: gr.Markdown("# Wan 2.2 First/Last Frame to Video") gr.Markdown("Provide a starting image, an ending image, and a text prompt to generate a video transitioning between them.") with gr.Row(): with gr.Column(scale=1): prompt_input = gr.Textbox(label="Prompt", value="the guy turns") first_image = gr.Image(label="First Frame", type="filepath") last_image = gr.Image(label="Last Frame", type="filepath") generate_btn = gr.Button("Generate Video") with gr.Column(scale=2): output_video = gr.Video(label="Generated Video") generate_btn.click( fn=generate_video, inputs=[prompt_input, first_image, last_image], outputs=[output_video] ) gr.Examples( examples=[ ["a beautiful woman, cinematic", "examples/start.png", "examples/end.png"] ], inputs=[prompt_input, first_image, last_image] ) if __name__ == "__main__": # Create example images if they don't exist if not os.path.exists("examples"): os.makedirs("examples") if not os.path.exists("examples/start.png"): Image.new('RGB', (512, 512), color = 'red').save('examples/start.png') if not os.path.exists("examples/end.png"): Image.new('RGB', (512, 512), color = 'blue').save('examples/end.png') app.launch()