Update app.py

app.py

@@ -11,7 +11,7 @@ import torch
 
 # 초기 실행 시 필요한 모델들을 다운로드
 cmd = (
-    'python3 -m pip install "huggingface_hub[cli]"; '
+    'python3 -m pip install "huggingface_hub[cli]" accelerate; '  # accelerate도 같이 설치 권장
     'huggingface-cli download LeonJoe13/Sonic --local-dir checkpoints; '
     'huggingface-cli download stabilityai/stable-video-diffusion-img2vid-xt --local-dir  checkpoints/stable-video-diffusion-img2vid-xt; '
     'huggingface-cli download openai/whisper-tiny --local-dir checkpoints/whisper-tiny;'
@@ -20,7 +20,7 @@ os.system(cmd)
 
 pipe = Sonic()
 
-def get_md5(content_bytes):
+def get_md5(content_bytes: bytes):
     """MD5 해시를 계산하여 32자리 문자열을 반환"""
     return hashlib.md5(content_bytes).hexdigest()
 
@@ -29,7 +29,6 @@ res_path = './res_path/'
 os.makedirs(tmp_path, exist_ok=True)
 os.makedirs(res_path, exist_ok=True)
 
-
 @spaces.GPU(duration=600)  # 긴 비디오 처리를 위해 duration 600초로 설정 (10분)
 def get_video_res(img_path, audio_path, res_video_path, dynamic_scale=1.0):
     """
@@ -48,7 +47,7 @@ def get_video_res(img_path, audio_path, res_video_path, dynamic_scale=1.0):
     # 최소 25 프레임, 최대 750 프레임 (60초 => 60*12.5=750)
     inference_steps = min(max(int(duration * 12.5), 25), 750)
     
-    print(f"[INFO] Audio duration: {duration:.2f} seconds, using inference_steps: {inference_steps}")
+    print(f"[INFO] Audio duration: {duration:.2f} seconds, using inference_steps={inference_steps}")
 
     # 얼굴 인식 (face_info는 참고용)
     face_info = pipe.preprocess(img_path, expand_ratio=expand_ratio)
@@ -71,7 +70,6 @@ def get_video_res(img_path, audio_path, res_video_path, dynamic_scale=1.0):
     else:
         return -1
 
-
 def process_sonic(image, audio, dynamic_scale):
     """
     Gradio 인터페이스 상에서 호출되는 함수.
@@ -138,252 +136,10 @@ def process_sonic(image, audio, dynamic_scale):
         video_result = get_video_res(image_path, audio_path, res_video_path, dynamic_scale)
         return video_result
 
-
 def get_example():
     """예시 데이터를 로딩하는 더미 함수 (현재는 빈 리스트)."""
     return []
 
-
-css = """
-.gradio-container {
-    font-family: 'Arial', sans-serif;
-}
-.main-header {
-    text-align: center;
-    color: #2a2a2a;
-    margin-bottom: 2em;
-}
-.parameter-section {
-    background-color: #f5f5f5;
-    padding: 1em;
-    border-radius: 8px;
-    margin: 1em 0;
-}
-.example-section {
-    margin-top: 2em;
-}
-"""
-
-with gr.Blocks(css=css) as demo:
-    gr.HTML("""
-        <div class="main-header">
-            <h1>🎭 Sonic: Advanced Portrait Animation</h1>
-            <p>Transform still images into dynamic videos synchronized with audio (up to 1 minute)</p>
-        </div>
-    """)
-    
-    with gr.Row():
-        with gr.Column():
-            image_input = gr.Image(
-                type='pil',
-                label="Portrait Image",
-                elem_id="image_input"
-            )
-            
-            audio_input = gr.Audio(
-                label="Voice/Audio Input (up to 1 minute)",
-                elem_id="audio_input",
-                type="numpy"
-            )
-            
-            with gr.Column():
-                dynamic_scale = gr.Slider(
-                    minimum=0.5,
-                    maximum=2.0,
-                    value=1.0,
-                    step=0.1,
-                    label="Animation Intensity",
-                    info="Adjust to control movement intensity (0.5: subtle, 2.0: dramatic)"
-                )
-            
-            process_btn = gr.Button(
-                "Generate Animation",
-                variant="primary",
-                elem_id="process_btn"
-            )
-        
-        with gr.Column():
-            video_output = gr.Video(
-                label="Generated Animation",
-                elem_id="video_output"
-            )
-    
-    process_btn.click(
-        fn=process_sonic,
-        inputs=[image_input, audio_input, dynamic_scale],
-        outputs=video_output,
-    )
-    
-    gr.Examples(
-        examples=get_example(),
-        fn=process_sonic,
-        inputs=[image_input, audio_input, dynamic_scale],
-        outputs=video_output,
-        cache_examples=False
-    )
-    
-    gr.HTML("""
-        <div style="text-align: center; margin-top: 2em;">
-            <div style="margin-bottom: 1em;">
-                <a href="https://github.com/jixiaozhong/Sonic" target="_blank" style="text-decoration: none;">
-                    <img src="https://img.shields.io/badge/GitHub-Repo-blue?style=for-the-badge&logo=github" alt="GitHub Repo">
-                </a>
-                <a href="https://arxiv.org/pdf/2411.16331" target="_blank" style="text-decoration: none;">
-                    <img src="https://img.shields.io/badge/Paper-arXiv-red?style=for-the-badge&logo=arxiv" alt="arXiv Paper">
-                </a>
-            </div>
-            <p>🔔 Note: For optimal results, use clear portrait images and high-quality audio (now supports up to 1 minute!)</p>
-        </div>
-    """)
-
-# 공개 링크 생성
-demo.launch(share=True)
-import spaces
-import gradio as gr
-import os
-import numpy as np
-from pydub import AudioSegment
-import hashlib
-import io
-from sonic import Sonic
-from PIL import Image
-import torch
-
-# 초기 실행 시 필요한 모델들을 다운로드
-cmd = (
-    'python3 -m pip install "huggingface_hub[cli]"; '
-    'huggingface-cli download LeonJoe13/Sonic --local-dir checkpoints; '
-    'huggingface-cli download stabilityai/stable-video-diffusion-img2vid-xt --local-dir  checkpoints/stable-video-diffusion-img2vid-xt; '
-    'huggingface-cli download openai/whisper-tiny --local-dir checkpoints/whisper-tiny;'
-)
-os.system(cmd)
-
-pipe = Sonic()
-
-def get_md5(content_bytes):
-    """MD5 해시를 계산하여 32자리 문자열을 반환"""
-    return hashlib.md5(content_bytes).hexdigest()
-
-tmp_path = './tmp_path/'
-res_path = './res_path/'
-os.makedirs(tmp_path, exist_ok=True)
-os.makedirs(res_path, exist_ok=True)
-
-
-@spaces.GPU(duration=600)  # 긴 비디오 처리를 위해 duration 600초로 설정 (10분)
-def get_video_res(img_path, audio_path, res_video_path, dynamic_scale=1.0):
-    """
-    Sonic pipeline으로부터 실제 비디오를 생성하는 함수.
-    최대 60초 길이의 오디오에 대해 inference_steps를 결정하여,
-    얼굴 탐지 후 영상 생성 작업을 수행함.
-    """
-    expand_ratio = 0.0
-    min_resolution = 512
-    
-    # 오디오 길이
-    audio = AudioSegment.from_file(audio_path)
-    duration = len(audio) / 1000.0  # 초 단위
-    
-    # 오디오 길이에 따라 inference_steps 계산 (초당 약 12.5 프레임)
-    # 최소 25 프레임, 최대 750 프레임 (60초 => 60*12.5=750)
-    inference_steps = min(max(int(duration * 12.5), 25), 750)
-    
-    print(f"[INFO] Audio duration: {duration:.2f} seconds, using inference_steps: {inference_steps}")
-
-    # 얼굴 인식 (face_info는 참고용)
-    face_info = pipe.preprocess(img_path, expand_ratio=expand_ratio)
-    print(f"[INFO] Face detection info: {face_info}")
-
-    # 얼굴이 하나라도 검출되면(>0), 원본 이미지 비율 유지
-    if face_info['face_num'] > 0:
-        os.makedirs(os.path.dirname(res_video_path), exist_ok=True)
-        
-        # Sonic pipeline으로 비디오 생성
-        pipe.process(
-            img_path,
-            audio_path,
-            res_video_path,
-            min_resolution=min_resolution,
-            inference_steps=inference_steps,
-            dynamic_scale=dynamic_scale
-        )
-        return res_video_path
-    else:
-        return -1
-
-
-def process_sonic(image, audio, dynamic_scale):
-    """
-    Gradio 인터페이스 상에서 호출되는 함수.
-    1. 이미지/오디오 입력 검증
-    2. MD5 해시 통해 파일명 생성 후 캐싱
-    3. 이미 결과 파일이 있으면 재활용, 없으면 새로 비디오 생성
-    """
-    if image is None:
-        raise gr.Error("Please upload an image")
-    if audio is None:
-        raise gr.Error("Please upload an audio file")
-
-    # 이미지 MD5 해시 계산
-    buf_img = io.BytesIO()
-    image.save(buf_img, format="PNG")
-    img_bytes = buf_img.getvalue()
-    img_md5 = get_md5(img_bytes)
-
-    # 오디오 MD5 해시 계산
-    sampling_rate, arr = audio[:2]
-    if len(arr.shape) == 1:
-        arr = arr[:, None]
-
-    audio_segment = AudioSegment(
-        arr.tobytes(),
-        frame_rate=sampling_rate,
-        sample_width=arr.dtype.itemsize,
-        channels=arr.shape[1]
-    )
-
-    # (중요) Whisper 호환을 위해 mono/16kHz 변환
-    audio_segment = audio_segment.set_channels(1)
-    audio_segment = audio_segment.set_frame_rate(16000)
-
-    # 최대 60초 제한
-    MAX_DURATION_MS = 60000
-    if len(audio_segment) > MAX_DURATION_MS:
-        audio_segment = audio_segment[:MAX_DURATION_MS]
-
-    buf_audio = io.BytesIO()
-    audio_segment.export(buf_audio, format="wav")
-    audio_bytes = buf_audio.getvalue()
-    audio_md5 = get_md5(audio_bytes)
-
-    # 파일 경로 생성
-    image_path = os.path.abspath(os.path.join(tmp_path, f'{img_md5}.png'))
-    audio_path = os.path.abspath(os.path.join(tmp_path, f'{audio_md5}.wav'))
-    res_video_path = os.path.abspath(os.path.join(res_path, f'{img_md5}_{audio_md5}_{dynamic_scale}.mp4'))
-    
-    # 이미지/오디오 파일 캐싱
-    if not os.path.exists(image_path):
-        with open(image_path, "wb") as f:
-            f.write(img_bytes)
-    if not os.path.exists(audio_path):
-        with open(audio_path, "wb") as f:
-            f.write(audio_bytes)
-    
-    # 이미 결과가 존재하면 캐시된 결과 사용
-    if os.path.exists(res_video_path):
-        print(f"[INFO] Using cached result: {res_video_path}")
-        return res_video_path
-    else:
-        print(f"[INFO] Generating new video with dynamic_scale={dynamic_scale}")
-        video_result = get_video_res(image_path, audio_path, res_video_path, dynamic_scale)
-        return video_result
-
-
-def get_example():
-    """예시 데이터를 로딩하는 더미 함수 (현재는 빈 리스트)."""
-    return []
-
-
 css = """
 .gradio-container {
     font-family: 'Arial', sans-serif;