webgui.py

#!/usr/bin/env python
# -*- coding: UTF-8 -*-
'''
webui
'''

import os
import random
from datetime import datetime
from pathlib import Path

import cv2
import numpy as np
import torch
from diffusers import AutoencoderKL, DDIMScheduler
from omegaconf import OmegaConf
from PIL import Image
from src.models.unet_2d_condition import UNet2DConditionModel
from src.models.unet_3d_echo import EchoUNet3DConditionModel
from src.models.whisper.audio2feature import load_audio_model
from src.pipelines.pipeline_echo_mimic import Audio2VideoPipeline
from src.utils.util import save_videos_grid, crop_and_pad
from src.models.face_locator import FaceLocator
from moviepy.editor import VideoFileClip, AudioFileClip
from facenet_pytorch import MTCNN
import argparse

import gradio as gr
from gradio_client import Client, handle_file
from pydub import AudioSegment
import huggingface_hub 
import spaces  # Import spaces module for ZeroGPU support

huggingface_hub.snapshot_download(
    repo_id='BadToBest/EchoMimic',
    local_dir='./pretrained_weights'
)

is_shared_ui = True if "fffiloni/EchoMimic" in os.environ.get('SPACE_ID', '') else False
available_property = False if is_shared_ui else True
advanced_settings_label = "Advanced Configuration (only for duplicated spaces)" if is_shared_ui else "Advanced Configuration"

default_values = {
    "width": 512,
    "height": 512,
    "length": 1200,
    "seed": 420,
    "facemask_dilation_ratio": 0.1,
    "facecrop_dilation_ratio": 0.5,
    "context_frames": 12,
    "context_overlap": 3,
    "cfg": 2.5,
    "steps": 30,
    "sample_rate": 16000,
    "fps": 24,
    "device": "cuda"
}

ffmpeg_path = os.getenv('FFMPEG_PATH')
if ffmpeg_path is None:
    print("please download ffmpeg-static and export to FFMPEG_PATH. \nFor example: export FFMPEG_PATH=/musetalk/ffmpeg-4.4-amd64-static")
elif ffmpeg_path not in os.getenv('PATH', ''):
    print("add ffmpeg to path")
    os.environ["PATH"] = f"{ffmpeg_path}:{os.environ['PATH']}"


config_path = "./configs/prompts/animation.yaml"
config = OmegaConf.load(config_path)
if config.weight_dtype == "fp16":
    weight_dtype = torch.float16
else:
    weight_dtype = torch.float32

device = "cuda"
if not torch.cuda.is_available():
    device = "cpu"

inference_config_path = config.inference_config
infer_config = OmegaConf.load(inference_config_path)

# Model initialization is performed on-demand with ZeroGPU

# Function to initialize models when needed
@spaces.GPU
def initialize_models():
    global vae, reference_unet, denoising_unet, face_locator, audio_processor, face_detector, pipe
    
    ## vae init
    vae = AutoencoderKL.from_pretrained(config.pretrained_vae_path).to(device, dtype=weight_dtype)

    ## reference net init
    reference_unet = UNet2DConditionModel.from_pretrained(
        config.pretrained_base_model_path,
        subfolder="unet",
    ).to(dtype=weight_dtype, device=device)
    reference_unet.load_state_dict(torch.load(config.reference_unet_path, map_location="cpu"))

    ## denoising net init
    if os.path.exists(config.motion_module_path):
        ### stage1 + stage2
        denoising_unet = EchoUNet3DConditionModel.from_pretrained_2d(
            config.pretrained_base_model_path,
            config.motion_module_path,
            subfolder="unet",
            unet_additional_kwargs=infer_config.unet_additional_kwargs,
        ).to(dtype=weight_dtype, device=device)
    else:
        ### only stage1
        denoising_unet = EchoUNet3DConditionModel.from_pretrained_2d(
            config.pretrained_base_model_path,
            "",
            subfolder="unet",
            unet_additional_kwargs={
                "use_motion_module": False,
                "unet_use_temporal_attention": False,
                "cross_attention_dim": infer_config.unet_additional_kwargs.cross_attention_dim
            }
        ).to(dtype=weight_dtype, device=device)

    denoising_unet.load_state_dict(torch.load(config.denoising_unet_path, map_location="cpu"), strict=False)

    ## face locator init
    face_locator = FaceLocator(320, conditioning_channels=1, block_out_channels=(16, 32, 96, 256)).to(dtype=weight_dtype, device=device)
    face_locator.load_state_dict(torch.load(config.face_locator_path))

    ## load audio processor params
    audio_processor = load_audio_model(model_path=config.audio_model_path, device=device)

    ## load face detector params
    face_detector = MTCNN(image_size=320, margin=0, min_face_size=20, thresholds=[0.6, 0.7, 0.7], factor=0.709, post_process=True, device=device)

    sched_kwargs = OmegaConf.to_container(infer_config.noise_scheduler_kwargs)
    scheduler = DDIMScheduler(**sched_kwargs)

    pipe = Audio2VideoPipeline(
        vae=vae,
        reference_unet=reference_unet,
        denoising_unet=denoising_unet,
        audio_guider=audio_processor,
        face_locator=face_locator,
        scheduler=scheduler,
    ).to(device, dtype=weight_dtype)

# Global variables for models
vae = None
reference_unet = None
denoising_unet = None
face_locator = None
audio_processor = None
face_detector = None
pipe = None

def ensure_png(image_path):
    # Load the image with Pillow
    with Image.open(image_path) as img:
        # Check if the image is already a PNG
        if img.format != "PNG":
            # Convert and save as PNG
            png_path = os.path.splitext(image_path)[0] + ".png"
            img.save(png_path, format="PNG")
            print(f"Image converted to PNG and saved as {png_path}")
            return png_path
        else:
            print("Image is already a PNG.")
            return image_path

def select_face(det_bboxes, probs):
    ## max face from faces that the prob is above 0.8
    ## box: xyxy
    if det_bboxes is None or probs is None:
        return None
    filtered_bboxes = []
    for bbox_i in range(len(det_bboxes)):
        if probs[bbox_i] > 0.8:
            filtered_bboxes.append(det_bboxes[bbox_i])
    if len(filtered_bboxes) == 0:
        return None
    sorted_bboxes = sorted(filtered_bboxes, key=lambda x:(x[3]-x[1]) * (x[2] - x[0]), reverse=True)
    return sorted_bboxes[0]

@spaces.GPU(duration=120)  # Allow up to 2 minutes for video processing (maximum allowed)
def process_video(uploaded_img, uploaded_audio, width, height, length, seed, facemask_dilation_ratio, facecrop_dilation_ratio, context_frames, context_overlap, cfg, steps, sample_rate, fps, device):
    # Ensure models are initialized
    if vae is None:
        initialize_models()

    if seed is not None and seed > -1:
        generator = torch.manual_seed(seed)
    else:
        generator = torch.manual_seed(random.randint(100, 1000000))

    uploaded_img = ensure_png(uploaded_img)

    #### face mask prepare
    face_img = cv2.imread(uploaded_img)
    
    # Get the original dimensions
    original_height, original_width = face_img.shape[:2]
    
    # Set the new width to 512 pixels
    new_width = 512
    
    # Calculate the new height with the same aspect ratio
    new_height = int(original_height * (new_width / original_width))
    
    # Ensure both width and height are divisible by 8
    new_width = (new_width // 8) * 8  # Force target width to be divisible by 8
    new_height = (new_height // 8) * 8   # Floor the height to the nearest multiple of 8
    

    # Resize the image to the calculated dimensions
    face_img = cv2.resize(face_img, (new_width, new_height))
    
    face_mask = np.zeros((face_img.shape[0], face_img.shape[1])).astype('uint8')
    det_bboxes, probs = face_detector.detect(face_img)
    select_bbox = select_face(det_bboxes, probs)
    if select_bbox is None:
        print("SELECT_BBOX IS NONE")
        face_mask[:, :] = 255
        face_img = cv2.resize(face_img, (width, height))
        face_mask = cv2.resize(face_mask, (width, height))
        raise gr.Error("Face Detector could not detect a face in your image. Try with a 512 squared image where the face is clearly visible.")
    else:
        print("SELECT_BBOX IS NOT NONE")
        xyxy = select_bbox[:4]
        xyxy = np.round(xyxy).astype('int')
        rb, re, cb, ce = xyxy[1], xyxy[3], xyxy[0], xyxy[2]
        r_pad = int((re - rb) * facemask_dilation_ratio)
        c_pad = int((ce - cb) * facemask_dilation_ratio)
        face_mask[rb - r_pad : re + r_pad, cb - c_pad : ce + c_pad] = 255
        
        #### face crop
        r_pad_crop = int((re - rb) * facecrop_dilation_ratio)
        c_pad_crop = int((ce - cb) * facecrop_dilation_ratio)
        crop_rect = [max(0, cb - c_pad_crop), max(0, rb - r_pad_crop), min(ce + c_pad_crop, face_img.shape[1]), min(re + r_pad_crop, face_img.shape[0])]
        face_img = crop_and_pad(face_img, crop_rect)
        face_mask = crop_and_pad(face_mask, crop_rect)
        face_img = cv2.resize(face_img, (width, height))
        face_mask = cv2.resize(face_mask, (width, height))

    ref_image_pil = Image.fromarray(face_img[:, :, [2, 1, 0]])
    face_mask_tensor = torch.Tensor(face_mask).to(dtype=weight_dtype, device=device).unsqueeze(0).unsqueeze(0).unsqueeze(0) / 255.0
    
    video = pipe(
        ref_image_pil,
        uploaded_audio,
        face_mask_tensor,
        width,
        height,
        length,
        steps,
        cfg,
        generator=generator,
        audio_sample_rate=sample_rate,
        context_frames=context_frames,
        fps=fps,
        context_overlap=context_overlap
    ).videos

    save_dir = Path("output/tmp")
    save_dir.mkdir(exist_ok=True, parents=True)
    output_video_path = save_dir / "output_video.mp4"
    save_videos_grid(video, str(output_video_path), n_rows=1, fps=fps)

    video_clip = VideoFileClip(str(output_video_path))
    audio_clip = AudioFileClip(uploaded_audio)
    final_output_path = save_dir / "output_video_with_audio.mp4"
    video_clip = video_clip.set_audio(audio_clip)
    video_clip.write_videofile(str(final_output_path), codec="libx264", audio_codec="aac")

    return final_output_path

@spaces.GPU(duration=60)  # Allow 1 minute for voice cloning
def get_maskGCT_TTS(prompt_audio_maskGCT, audio_to_clone):
    try:
        client = Client("amphion/maskgct")
    except:
        raise gr.Error(f"amphion/maskgct space's api might not be ready, please wait, or upload an audio instead.")
    
    result = client.predict(
        prompt_wav = handle_file(audio_to_clone),
        target_text = prompt_audio_maskGCT,
        target_len=-1,
        n_timesteps=25,
        api_name="/predict"
    )
    print(result) 
    return result, gr.update(value=result, visible=True)
  
with gr.Blocks() as demo:
    gr.Markdown('# EchoMimic')
    gr.Markdown('## Lifelike Audio-Driven Portrait Animations through Editable Landmark Conditioning')
    gr.Markdown('Running on Spaces ZeroGPU: Dynamic GPU allocation for optimal resource usage')
    gr.HTML("""
    <div style="display:flex;column-gap:4px;">
        <a href='https://badtobest.github.io/echomimic.html'><img src='https://img.shields.io/badge/Project-Page-blue'></a>
        <a href='https://huggingface.co/BadToBest/EchoMimic'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20HuggingFace-Model-yellow'></a>
        <a href='https://arxiv.org/abs/2407.08136'><img src='https://img.shields.io/badge/Paper-Arxiv-red'></a>
    </div>
    """)
    with gr.Row():
        with gr.Column():
            uploaded_img = gr.Image(type="filepath", label="Reference Image")
            uploaded_audio = gr.Audio(type="filepath", label="Input Audio", format="wav")
            preprocess_audio_file = gr.File(visible=False)
            with gr.Accordion(label="Voice cloning with MaskGCT", open=False):
                prompt_audio_maskGCT = gr.Textbox(
                    label = "Text to synthetize",
                    lines = 2,
                    max_lines = 2,
                    elem_id = "text-synth-maskGCT"
                )
                audio_to_clone_maskGCT = gr.Audio(
                    label = "Voice to clone",
                    type = "filepath",
                    elem_id = "audio-clone-elm-maskGCT"
                )
                gen_maskGCT_voice_btn = gr.Button("Generate voice clone (optional)")
            with gr.Accordion(label=advanced_settings_label, open=False):
                with gr.Row():
                    width = gr.Slider(label="Width", minimum=128, maximum=1024, value=default_values["width"], interactive=available_property)
                    height = gr.Slider(label="Height", minimum=128, maximum=1024, value=default_values["height"], interactive=available_property)
                with gr.Row():
                    length = gr.Slider(label="Length", minimum=100, maximum=5000, value=default_values["length"], interactive=available_property)
                    seed = gr.Slider(label="Seed", minimum=0, maximum=10000, value=default_values["seed"], interactive=available_property)
                with gr.Row():
                    facemask_dilation_ratio = gr.Slider(label="Facemask Dilation Ratio", minimum=0.0, maximum=1.0, step=0.01, value=default_values["facemask_dilation_ratio"], interactive=available_property)
                    facecrop_dilation_ratio = gr.Slider(label="Facecrop Dilation Ratio", minimum=0.0, maximum=1.0, step=0.01, value=default_values["facecrop_dilation_ratio"], interactive=available_property)
                with gr.Row():
                    context_frames = gr.Slider(label="Context Frames", minimum=0, maximum=50, step=1, value=default_values["context_frames"], interactive=available_property)
                    context_overlap = gr.Slider(label="Context Overlap", minimum=0, maximum=10, step=1, value=default_values["context_overlap"], interactive=available_property)
                with gr.Row():
                    cfg = gr.Slider(label="CFG", minimum=0.0, maximum=10.0, step=0.1, value=default_values["cfg"], interactive=available_property)
                    steps = gr.Slider(label="Steps", minimum=1, maximum=100, step=1, value=default_values["steps"], interactive=available_property)
                with gr.Row():
                    sample_rate = gr.Slider(label="Sample Rate", minimum=8000, maximum=48000, step=1000, value=default_values["sample_rate"], interactive=available_property)
                    fps = gr.Slider(label="FPS", minimum=1, maximum=60, step=1, value=default_values["fps"], interactive=available_property)
                    device = gr.Radio(label="Device", choices=["cuda", "cpu"], value=default_values["device"], interactive=available_property)
            generate_button = gr.Button("Generate Video")
        with gr.Column():
            output_video = gr.Video()
            gr.Examples(
                label = "Portrait examples",
                examples = [
                    ['assets/test_imgs/a.png'],
                    ['assets/test_imgs/b.png'],
                    ['assets/test_imgs/c.png'],
                    ['assets/test_imgs/d.png'],
                    ['assets/test_imgs/e.png']
                ],
                inputs = [uploaded_img]
            )
            gr.Examples(
                label = "Audio examples",
                examples = [
                    ['assets/test_audios/chunnuanhuakai.wav'],
                    ['assets/test_audios/chunwang.wav'],
                    ['assets/test_audios/echomimic_en_girl.wav'],
                    ['assets/test_audios/echomimic_en.wav'],
                    ['assets/test_audios/echomimic_girl.wav'],
                    ['assets/test_audios/echomimic.wav'],
                    ['assets/test_audios/jane.wav'],
                    ['assets/test_audios/mei.wav'],
                    ['assets/test_audios/walden.wav'],
                    ['assets/test_audios/yun.wav'],
                ],
                inputs = [uploaded_audio]
            )
            gr.HTML("""
            <div style="display:flex;column-gap:4px;">
                <a href="https://huggingface.co/spaces/fffiloni/EchoMimic?duplicate=true">
                    <img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/duplicate-this-space-xl.svg" alt="Duplicate this Space">
                </a>
                <a href="https://huggingface.co/fffiloni">
                    <img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/follow-me-on-HF-xl-dark.svg" alt="Follow me on HF">
                </a>
            </div>
            """)

    def trim_audio(file_path, output_path, max_duration=10):
        # Load the audio file
        audio = AudioSegment.from_wav(file_path)
    
        # Convert max duration to milliseconds
        max_duration_ms = max_duration * 1000
    
        # Trim the audio if it's longer than max_duration
        if len(audio) > max_duration_ms:
            audio = audio[:max_duration_ms]
    
        # Export the trimmed audio
        audio.export(output_path, format="wav")
        print(f"Audio trimmed and saved as {output_path}")
        return output_path

    def generate_video(uploaded_img, uploaded_audio, width, height, length, seed, facemask_dilation_ratio, facecrop_dilation_ratio, context_frames, context_overlap, cfg, steps, sample_rate, fps, device, progress=gr.Progress(track_tqdm=True)):
        # First, check and trim audio if needed
        if is_shared_ui:
            gr.Info("Trimming audio to max 10 seconds. Duplicate the space for unlimited audio length.")
            uploaded_audio = trim_audio(uploaded_audio, "trimmed_audio.wav")

        # Process the video with ZeroGPU support
        final_output_path = process_video(
            uploaded_img, uploaded_audio, width, height, length, seed, facemask_dilation_ratio, facecrop_dilation_ratio, context_frames, context_overlap, cfg, steps, sample_rate, fps, device
        )        
        return final_output_path

    gen_maskGCT_voice_btn.click(
        fn = get_maskGCT_TTS,
        inputs = [prompt_audio_maskGCT, audio_to_clone_maskGCT],
        outputs = [uploaded_audio, preprocess_audio_file],
        queue = False,
        show_api = False
    )

    generate_button.click(
        generate_video,
        inputs=[
            uploaded_img,
            uploaded_audio,
            width,
            height,
            length,
            seed,
            facemask_dilation_ratio,
            facecrop_dilation_ratio,
            context_frames,
            context_overlap,
            cfg,
            steps,
            sample_rate,
            fps,
            device
        ],
        outputs=output_video,
        show_api=False
    )
parser = argparse.ArgumentParser(description='EchoMimic with ZeroGPU Support')
parser.add_argument('--server_name', type=str, default='0.0.0.0', help='Server name')
parser.add_argument('--server_port', type=int, default=7680, help='Server port')
args = parser.parse_args()

if __name__ == '__main__':
    demo.queue(max_size=3).launch(show_api=False, show_error=True, ssr_mode=False)
    #demo.launch(server_name=args.server_name, server_port=args.server_port, inbrowser=True)