app.py

import os, sys
import tempfile
import gradio as gr
from app_tts import infer_tts
from src.gradio_demo import SadTalker  
# from src.utils.text2speech import TTSTalker
from huggingface_hub import snapshot_download

def get_source_image(image):   
        return image

try:
    import webui  # in webui
    in_webui = True
except:
    in_webui = False


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


# New: Gộp 2 nút thành 1, output audio là input cho video
import soundfile as sf

def generate_voice_and_video(ref_audio, ref_text, gen_text, speed, source_image, 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):
    # 1. Sinh audio từ TTS
    (final_sample_rate, final_wave), _ = infer_tts(ref_audio, ref_text, gen_text, speed)
    # Lưu ra file tạm
    tmp_audio = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
    sf.write(tmp_audio.name, final_wave, final_sample_rate)
    # 2. Gọi SadTalker với audio vừa sinh ra
    sad_talker = SadTalker(lazy_load=True)
    video_path = sad_talker.test(
        source_image,
        tmp_audio.name,
        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
    )
    return tmp_audio.name, video_path

def sadtalker_demo():
    download_model()
    with gr.Blocks(analytics_enabled=False) as sadtalker_interface:
        gr.Markdown("""
# 🎤 F5-TTS: Vietnamese Text-to-Speech Synthesis & SadTalker Video
# Nhập text, upload sample voice và ảnh để tạo video nói chuyện.
""")
        with gr.Row():
            ref_audio = gr.Audio(label="🔊 Sample Voice", type="filepath")
            ref_text = gr.Textbox(label="📝 Reference Transcript (optional)", placeholder="Nhập transcript tiếng Việt cho sample voice nếu có...", lines=2)
            gen_text = gr.Textbox(label="📝 Text", placeholder="Enter the text to generate voice...", lines=3)
        speed = gr.Slider(0.3, 2.0, value=1.0, step=0.1, label="⚡ Speed")
        with gr.Row():
            source_image = gr.Image(label="Source image", type="filepath", elem_id="img2img_image")
        with gr.Row():
            # Các setting cho SadTalker
            with gr.Column():
                preprocess_type = gr.Radio(['crop', 'resize','full', 'extcrop', 'extfull'], value='crop', label='preprocess', info="How to handle input image?")
                is_still_mode = gr.Checkbox(label="Still Mode (fewer head motion, works with preprocess `full`)")
                enhancer = gr.Checkbox(label="GFPGAN as Face enhancer")
                batch_size = gr.Slider(label="batch size in generation", step=1, maximum=10, value=1)
                size_of_image = gr.Radio([256, 512], value=256, label='face model resolution', info="use 256/512 model?")
                pose_style = gr.Slider(minimum=0, maximum=45, step=1, label="Pose style", value=0)
                facerender = gr.Radio(['facevid2vid','pirender'], value='facevid2vid', label='facerender', info="which face render?")
                exp_weight = gr.Slider(minimum=0, maximum=3, step=0.1, label="expression scale", value=1)
                use_ref_video = gr.Checkbox(label="Use Reference Video")
                ref_video = gr.Video(label="Reference Video", elem_id="vidref")
                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))")
                use_idle_mode = gr.Checkbox(label="Use Idle Animation")
                length_of_audio = gr.Number(value=5, label="The length(seconds) of the generated video.")
                blink_every = gr.Checkbox(label="use eye blink", value=True)
        btn_generate = gr.Button("🔥 Generate Voice & Video")
        with gr.Row():
            output_audio = gr.Audio(label="🎧 Generated Audio", type="filepath")
            gen_video = gr.Video(label="Generated video", format="mp4", scale=1)
        btn_generate.click(
            generate_voice_and_video,
            inputs=[ref_audio, ref_text, gen_text, speed, source_image, 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=[output_audio, gen_video]
        )
    return sadtalker_interface

if __name__ == "__main__":
    demo = sadtalker_demo()
    demo.queue(max_size=10, api_open=True)
    demo.launch(debug=True, server_name="0.0.0.0")