import subprocess subprocess.run( 'pip install numpy==1.26.4', shell=True ) import os import gradio as gr import torch import spaces import random from PIL import Image import numpy as np from glob import glob from pathlib import Path from typing import Optional #Core functions from https://github.com/modelscope/DiffSynth-Studio from diffsynth import save_video, ModelManager, SVDVideoPipeline from diffsynth import SDVideoPipeline, ControlNetConfigUnit, VideoData, save_frames from diffsynth.extensions.RIFE import RIFESmoother import cv2 # Constants MAX_SEED = np.iinfo(np.int32).max CSS = """ footer { visibility: hidden; } """ JS = """function () { gradioURL = window.location.href if (!gradioURL.endsWith('?__theme=dark')) { window.location.replace(gradioURL + '?__theme=dark'); } }""" # Ensure model and scheduler are initialized in GPU-enabled function if torch.cuda.is_available(): model_manager2 = ModelManager(torch_dtype=torch.float16, device="cuda") model_manager2.load_textual_inversions("models/textual_inversion") model_manager2.load_models([ "models/stable_diffusion/flat2DAnimerge_v45Sharp.safetensors", "models/AnimateDiff/mm_sd_v15_v2.ckpt", "models/ControlNet/control_v11p_sd15_lineart.pth", "models/ControlNet/control_v11f1e_sd15_tile.pth", "models/RIFE/flownet.pkl" ]) pipe2 = SDVideoPipeline.from_model_manager( model_manager2, [ ControlNetConfigUnit( processor_id="lineart", model_path="models/ControlNet/control_v11p_sd15_lineart.pth", scale=0.5 ), ControlNetConfigUnit( processor_id="tile", model_path="models/ControlNet/control_v11f1e_sd15_tile.pth", scale=0.5 ) ] ) smoother = RIFESmoother.from_model_manager(model_manager2) def update_frames(video_in): up_video = VideoData( video_file=video_in) frame_len = len(up_video) video_path = video_in cap = cv2.VideoCapture(video_path) fps_in = cap.get(cv2.CAP_PROP_FPS) width_in = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)) height_in = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) cap.release() return gr.update(maximum=frame_len), gr.update(value=fps_in), gr.update(value=width_in), gr.update(value=height_in) @spaces.GPU(duration=180) def generate( video_in, image_in, prompt: str = "best quality", seed: int = -1, num_inference_steps: int = 10, num_frames: int = 30, height: int = 512, width: int = 512, animatediff_batch_size: int = 32, animatediff_stride: int = 16, fps_id: int = 25, output_folder: str = "outputs", progress=gr.Progress(track_tqdm=True)): video = "" if seed == -1: seed = random.randint(0, MAX_SEED) torch.manual_seed(seed) os.makedirs(output_folder, exist_ok=True) base_count = len(glob(os.path.join(output_folder, "*.mp4"))) video_path = os.path.join(output_folder, f"{base_count:06d}.mp4") up_video = VideoData( video_file=video_in, height=height, width=width) input_video = [up_video[i] for i in range(1, num_frames)] video = pipe2( prompt=prompt, negative_prompt="verybadimagenegative_v1.3", cfg_scale=3, clip_skip=2, controlnet_frames=input_video, num_frames=len(input_video), num_inference_steps=num_inference_steps, height=height, width=width, animatediff_batch_size=animatediff_batch_size, animatediff_stride=animatediff_stride, unet_batch_size=8, controlnet_batch_size=8, vram_limit_level=0, ) video = smoother(video) save_video(video, video_path, fps=fps_id) return video_path, seed examples = [ ['./dancing.mp4', None, "best quality, perfect anime illustration, light, a girl is dancing, smile, solo"], ] # Gradio Interface with gr.Blocks(css=CSS, js=JS, theme="soft") as demo: gr.HTML("