""" Hallo2: Long-Duration and High-Resolution Audio-driven Portrait Image Animation Gradio App for HuggingFace Spaces """ import os import sys import argparse import tempfile import traceback from pathlib import Path import gradio as gr import torch # --------------------------------------------------------------------------- # Environment # --------------------------------------------------------------------------- ROOT_DIR = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, ROOT_DIR) DEVICE = "cuda" if torch.cuda.is_available() else "cpu" print(f"Running on device: {DEVICE}") # --------------------------------------------------------------------------- # Model Download # --------------------------------------------------------------------------- def download_pretrained_models(): """ Download Hallo2 pretrained models if not already present. """ pretrained_dir = "./pretrained_models" os.makedirs(pretrained_dir, exist_ok=True) try: marker_file = os.path.join(pretrained_dir, ".download_complete") if os.path.exists(marker_file): print("Pretrained models already present.") return print("Downloading Hallo2 pretrained models...") from huggingface_hub import snapshot_download snapshot_download( repo_id="fudan-generative-ai/hallo2", local_dir=pretrained_dir, local_dir_use_symlinks=False, resume_download=True, ) with open(marker_file, "w") as f: f.write("ok") print("Model download completed.") except Exception as e: print(f"[WARNING] Model download failed:") print(str(e)) # --------------------------------------------------------------------------- # Inference # --------------------------------------------------------------------------- def run_inference( source_image_path, driving_audio_path, pose_weight, face_weight, lip_weight, face_expand_ratio, inference_steps, cfg_scale, ): try: from omegaconf import OmegaConf from scripts.inference_long import inference_process from hallo.utils.util import merge_videos output_dir = tempfile.mkdtemp(prefix="hallo2_out_") config_path = "configs/inference/long.yaml" if not os.path.exists(config_path): raise FileNotFoundError(f"Config not found: {config_path}") cfg = OmegaConf.load(config_path) cfg.inference_steps = int(inference_steps) cfg.cfg_scale = float(cfg_scale) cfg.save_path = output_dir cfg.source_image = source_image_path cfg.driving_audio = driving_audio_path cfg.pose_weight = float(pose_weight) cfg.face_weight = float(face_weight) cfg.lip_weight = float(lip_weight) cfg.face_expand_ratio = float(face_expand_ratio) temp_cfg = os.path.join(output_dir, "temp_config.yaml") OmegaConf.save(cfg, temp_cfg) args = argparse.Namespace( config=temp_cfg, source_image=source_image_path, driving_audio=driving_audio_path, pose_weight=float(pose_weight), face_weight=float(face_weight), lip_weight=float(lip_weight), face_expand_ratio=float(face_expand_ratio), audio_ckpt_dir=None, ) seg_path = inference_process(args) merged_path = os.path.join(Path(seg_path).parent, "merge_video.mp4") merge_videos(seg_path, merged_path) if os.path.exists(merged_path): return merged_path mp4_files = list(Path(seg_path).glob("*.mp4")) if mp4_files: return str(mp4_files[0]) raise RuntimeError("No output video generated.") except Exception: raise RuntimeError(traceback.format_exc()) # --------------------------------------------------------------------------- # Gradio Inference Wrapper # --------------------------------------------------------------------------- def infer( source_image, driving_audio, pose_weight, face_weight, lip_weight, face_expand_ratio, inference_steps, cfg_scale, ): if source_image is None: return None, "Please upload a portrait image." if driving_audio is None: return None, "Please upload a WAV audio file." try: video_path = run_inference( source_image_path=source_image, driving_audio_path=driving_audio, pose_weight=pose_weight, face_weight=face_weight, lip_weight=lip_weight, face_expand_ratio=face_expand_ratio, inference_steps=inference_steps, cfg_scale=cfg_scale, ) return video_path, "Video generated successfully." except Exception as e: return None, f"Error:\n{str(e)}" # --------------------------------------------------------------------------- # UI # --------------------------------------------------------------------------- with gr.Blocks(title="Hallo2 Audio Driven Portrait Animation") as demo: gr.Markdown( """ # Hallo2 Upload: - Portrait image - WAV audio Generate realistic talking-head animation. """ ) with gr.Row(): with gr.Column(): source_image = gr.Image(label="Portrait Image", type="filepath") driving_audio = gr.Audio(label="Driving Audio", type="filepath") with gr.Accordion("Advanced Settings", open=False): pose_weight = gr.Slider( 0.0, 2.0, value=1.0, step=0.1, label="Pose Weight" ) face_weight = gr.Slider( 0.0, 2.0, value=1.0, step=0.1, label="Face Weight" ) lip_weight = gr.Slider( 0.0, 2.0, value=1.0, step=0.1, label="Lip Weight" ) face_expand_ratio = gr.Slider( 1.0, 2.0, value=1.2, step=0.05, label="Face Expand Ratio" ) inference_steps = gr.Slider( 10, 100, value=40, step=5, label="Inference Steps" ) cfg_scale = gr.Slider( 1.0, 10.0, value=3.5, step=0.5, label="CFG Scale" ) run_btn = gr.Button("Generate Video", variant="primary") with gr.Column(): output_video = gr.Video(label="Generated Video") status_box = gr.Textbox(label="Status", lines=8, interactive=False) run_btn.click( fn=infer, inputs=[ source_image, driving_audio, pose_weight, face_weight, lip_weight, face_expand_ratio, inference_steps, cfg_scale, ], outputs=[output_video, status_box], ) # --------------------------------------------------------------------------- # Startup # --------------------------------------------------------------------------- try: download_pretrained_models() except Exception as e: print(f"Startup warning: {e}") if __name__ == "__main__": demo.queue(max_size=5).launch()