app.py

import os
import platform
import torch
import gradio as gr
from huggingface_hub import snapshot_download
import uuid
import shutil
from pydub import AudioSegment
import spaces

from src.facerender.pirender_animate import AnimateFromCoeff_PIRender
from src.utils.preprocess import CropAndExtract
from src.test_audio2coeff import Audio2Coeff
from src.facerender.animate import AnimateFromCoeff
from src.generate_batch import get_data
from src.generate_facerender_batch import get_facerender_data
from src.utils.init_path import init_path


def get_source_image(image):
    return image


def toggle_audio_file(choice):
    if choice == False:
        return gr.update(visible=True), gr.update(visible=False)
    else:
        return gr.update(visible=False), gr.update(visible=True)


def ref_video_fn(path_of_ref_video):
    if path_of_ref_video is not None:
        return gr.update(value=True)
    else:
        return gr.update(value=False)


if torch.cuda.is_available():
    device = "cuda"
elif platform.system() == 'Darwin':  # macos
    device = "mps"
else:
    device = "cpu"

os.environ['TORCH_HOME'] = 'checkpoints'

checkpoint_path = 'checkpoints'
config_path = 'src/config'

snapshot_download(repo_id='vinthony/SadTalker-V002rc',
                  local_dir='./checkpoints', local_dir_use_symlinks=True)


def mp3_to_wav(mp3_filename, wav_filename, frame_rate):
    mp3_file = AudioSegment.from_file(file=mp3_filename)
    mp3_file.set_frame_rate(frame_rate).export(wav_filename, format="wav")


@spaces.GPU()
def test(source_image, driven_audio, preprocess='crop',
         still_mode=False,  use_enhancer=False, batch_size=1, size=256,
         pose_style=0,
         facerender='facevid2vid',
         exp_scale=1.0,
         use_ref_video=False,
         ref_video=None,
         ref_info=None,
         use_idle_mode=False,
         length_of_audio=0, use_blink=True,
         result_dir='./results/'):

    sadtalker_paths = init_path(
        checkpoint_path, config_path, size, False, preprocess)

    audio_to_coeff = Audio2Coeff(sadtalker_paths, device)
    preprocess_model = CropAndExtract(sadtalker_paths, device)

    if facerender == 'facevid2vid' and device != 'mps':
        animate_from_coeff = AnimateFromCoeff(
            sadtalker_paths, device)
    elif facerender == 'pirender' or device == 'mps':
        animate_from_coeff = AnimateFromCoeff_PIRender(
            sadtalker_paths, device)
        facerender = 'pirender'
    else:
        raise (RuntimeError('Unknown model: {}'.format(facerender)))

    time_tag = str(uuid.uuid4())
    save_dir = os.path.join(result_dir, time_tag)
    os.makedirs(save_dir, exist_ok=True)

    input_dir = os.path.join(save_dir, 'input')
    os.makedirs(input_dir, exist_ok=True)

    print(source_image)
    pic_path = os.path.join(input_dir, os.path.basename(source_image))
    shutil.move(source_image, input_dir)

    if driven_audio is not None and os.path.isfile(driven_audio):
        audio_path = os.path.join(input_dir, os.path.basename(driven_audio))

        # mp3 to wav
        if '.mp3' in audio_path:
            mp3_to_wav(driven_audio, audio_path.replace('.mp3', '.wav'), 16000)
            audio_path = audio_path.replace('.mp3', '.wav')
        else:
            shutil.move(driven_audio, input_dir)

    elif use_idle_mode:
        # generate audio from this new audio_path
        audio_path = os.path.join(
            input_dir, 'idlemode_'+str(length_of_audio)+'.wav')
        from pydub import AudioSegment
        one_sec_segment = AudioSegment.silent(
            duration=1000*length_of_audio)  # duration in milliseconds
        one_sec_segment.export(audio_path, format="wav")
    else:
        print(use_ref_video, ref_info)
        assert use_ref_video == True and ref_info == 'all'

    if use_ref_video and ref_info == 'all':  # full ref mode
        ref_video_videoname = os.path.basename(ref_video)
        audio_path = os.path.join(save_dir, ref_video_videoname+'.wav')
        print('new audiopath:', audio_path)
        # if ref_video contains audio, set the audio from ref_video.
        cmd = r"ffmpeg -y -hide_banner -loglevel error -i %s %s" % (
            ref_video, audio_path)
        os.system(cmd)

    os.makedirs(save_dir, exist_ok=True)

    # crop image and extract 3dmm from image
    first_frame_dir = os.path.join(save_dir, 'first_frame_dir')
    os.makedirs(first_frame_dir, exist_ok=True)
    first_coeff_path, crop_pic_path, crop_info = preprocess_model.generate(
        pic_path, first_frame_dir, preprocess, True, size)

    if first_coeff_path is None:
        raise AttributeError("No face is detected")

    if use_ref_video:
        print('using ref video for genreation')
        ref_video_videoname = os.path.splitext(os.path.split(ref_video)[-1])[0]
        ref_video_frame_dir = os.path.join(save_dir, ref_video_videoname)
        os.makedirs(ref_video_frame_dir, exist_ok=True)
        print('3DMM Extraction for the reference video providing pose')
        ref_video_coeff_path, _, _ = preprocess_model.generate(
            ref_video, ref_video_frame_dir, preprocess, source_image_flag=False)
    else:
        ref_video_coeff_path = None

    if use_ref_video:
        if ref_info == 'pose':
            ref_pose_coeff_path = ref_video_coeff_path
            ref_eyeblink_coeff_path = None
        elif ref_info == 'blink':
            ref_pose_coeff_path = None
            ref_eyeblink_coeff_path = ref_video_coeff_path
        elif ref_info == 'pose+blink':
            ref_pose_coeff_path = ref_video_coeff_path
            ref_eyeblink_coeff_path = ref_video_coeff_path
        elif ref_info == 'all':
            ref_pose_coeff_path = None
            ref_eyeblink_coeff_path = None
        else:
            raise ('error in refinfo')
    else:
        ref_pose_coeff_path = None
        ref_eyeblink_coeff_path = None

    # audio2ceoff
    if use_ref_video and ref_info == 'all':
        # audio_to_coeff.generate(batch, save_dir, pose_style, ref_pose_coeff_path)
        coeff_path = ref_video_coeff_path
    else:
        batch = get_data(first_coeff_path, audio_path, device, ref_eyeblink_coeff_path=ref_eyeblink_coeff_path, still=still_mode,
                         idlemode=use_idle_mode, length_of_audio=length_of_audio, use_blink=use_blink)  # longer audio?
        coeff_path = audio_to_coeff.generate(
            batch, save_dir, pose_style, ref_pose_coeff_path)

    # coeff2video
    data = get_facerender_data(coeff_path, crop_pic_path, first_coeff_path, audio_path, batch_size, still_mode=still_mode,
                               preprocess=preprocess, size=size, expression_scale=exp_scale, facemodel=facerender)
    return_path = animate_from_coeff.generate(
        data, save_dir,  pic_path, crop_info, enhancer='gfpgan' if use_enhancer else None, preprocess=preprocess, img_size=size)
    video_name = data['video_name']
    print(f'The generated video is named {video_name} in {save_dir}')

    return return_path


with gr.Blocks(analytics_enabled=False) as demo:
    with gr.Row():
        with gr.Column(variant='panel'):
            with gr.Tabs(elem_id="sadtalker_source_image"):
                with gr.TabItem('Source image'):
                    with gr.Row():
                        source_image = gr.Image(
                            label="Source image", sources="upload", type="filepath", elem_id="img2img_image")

            with gr.Tabs(elem_id="sadtalker_driven_audio"):
                with gr.TabItem('Driving Methods'):
                    gr.Markdown(
                        "Possible driving combinations: <br> 1. Audio only 2. Audio/IDLE Mode + Ref Video(pose, blink, pose+blink) 3. IDLE Mode only 4. Ref Video only (all) ")

                    with gr.Row():
                        driven_audio = gr.Audio(
                            label="Input audio", sources="upload", type="filepath")
                        driven_audio_no = gr.Audio(
                            label="Use IDLE mode, no audio is required", sources="upload", type="filepath", visible=False)

                        with gr.Column():
                            use_idle_mode = gr.Checkbox(
                                label="Use Idle Animation")
                            length_of_audio = gr.Number(
                                value=5, label="The length(seconds) of the generated video.")
                            use_idle_mode.change(toggle_audio_file, inputs=use_idle_mode, outputs=[
                                driven_audio, driven_audio_no])  # todo

                    with gr.Row():
                        ref_video = gr.Video(
                            label="Reference Video", sources="upload", elem_id="vidref")

                        with gr.Column():
                            use_ref_video = gr.Checkbox(
                                label="Use Reference Video")
                            ref_info = gr.Radio(['pose', 'blink', 'pose+blink', 'all'], value='pose', label='Reference Video',
                                                info="How to borrow from reference Video?((fully transfer, aka, video driving mode))")

                        ref_video.change(ref_video_fn, inputs=ref_video, outputs=[
                            use_ref_video])  # todo

        with gr.Column(variant='panel'):
            with gr.Tabs(elem_id="sadtalker_checkbox"):
                with gr.TabItem('Settings'):
                    with gr.Column(variant='panel'):
                        # width = gr.Slider(minimum=64, elem_id="img2img_width", maximum=2048, step=8, label="Manually Crop Width", value=512) # img2img_width
                        # height = gr.Slider(minimum=64, elem_id="img2img_height", maximum=2048, step=8, label="Manually Crop Height", value=512) # img2img_width
                        with gr.Row():
                            pose_style = gr.Slider(
                                minimum=0, maximum=45, step=1, label="Pose style", value=0)
                            exp_weight = gr.Slider(
                                minimum=0, maximum=3, step=0.1, label="expression scale", value=1)
                            blink_every = gr.Checkbox(
                                label="use eye blink", value=True)

                        with gr.Row():
                            size_of_image = gr.Radio(
                                [256, 512], value=256, label='face model resolution', info="use 256/512 model?")
                            preprocess_type = gr.Radio(
                                ['crop', 'resize', 'full', 'extcrop', 'extfull'], value='crop', label='preprocess', info="How to handle input image?")

                        with gr.Row():
                            is_still_mode = gr.Checkbox(
                                label="Still Mode (fewer head motion, works with preprocess `full`)")
                            facerender = gr.Radio(
                                ['facevid2vid', 'pirender'], value='facevid2vid', label='facerender', info="which face render?")

                        with gr.Row():
                            batch_size = gr.Slider(
                                label="batch size in generation", step=1, maximum=10, value=1)
                            enhancer = gr.Checkbox(
                                label="GFPGAN as Face enhancer")

                        submit = gr.Button(
                            'Generate', elem_id="sadtalker_generate", variant='primary')

            with gr.Tabs(elem_id="sadtalker_genearted"):
                gen_video = gr.Video(label="Generated video")

    submit.click(
        fn=test,
        inputs=[source_image,
                driven_audio,
                preprocess_type,
                is_still_mode,
                enhancer,
                batch_size,
                size_of_image,
                pose_style,
                facerender,
                exp_weight,
                use_ref_video,
                ref_video,
                ref_info,
                use_idle_mode,
                length_of_audio,
                blink_every
                ],
        outputs=[gen_video],
    )

    with gr.Row():
        gr.Examples(examples=[
            [
                'examples/source_image/full_body_1.png',
                'examples/driven_audio/bus_chinese.wav',
                'crop',
                True,
                False
            ],
            [
                'examples/source_image/full_body_2.png',
                'examples/driven_audio/japanese.wav',
                'crop',
                False,
                False
            ],
            [
                'examples/source_image/full3.png',
                'examples/driven_audio/deyu.wav',
                'crop',
                False,
                True
            ],
            [
                'examples/source_image/full4.jpeg',
                'examples/driven_audio/eluosi.wav',
                'full',
                False,
                True
            ],
            [
                'examples/source_image/full4.jpeg',
                'examples/driven_audio/imagine.wav',
                'full',
                True,
                True
            ],
            [
                'examples/source_image/full_body_1.png',
                'examples/driven_audio/bus_chinese.wav',
                'full',
                True,
                False
            ],
            [
                'examples/source_image/art_13.png',
                'examples/driven_audio/fayu.wav',
                'resize',
                True,
                False
            ],
            [
                'examples/source_image/art_5.png',
                'examples/driven_audio/chinese_news.wav',
                'resize',
                False,
                False
            ],
            [
                'examples/source_image/art_5.png',
                'examples/driven_audio/RD_Radio31_000.wav',
                'resize',
                True,
                True
            ],
        ],
            inputs=[
            source_image,
            driven_audio,
            preprocess_type,
            is_still_mode,
            enhancer],
            outputs=[gen_video],
            fn=test)

demo.launch(debug=True)