Update app.py

app.py

@@ -6,9 +6,11 @@ from pydub import AudioSegment
 import hashlib
 from sonic import Sonic
 from PIL import Image
-import torch
+import torch  # 필요 시 사용
 
+# ------------------------------------------------------------------
 # 모델 초기화
+# ------------------------------------------------------------------
 cmd = (
     'python3 -m pip install "huggingface_hub[cli]"; '
     'huggingface-cli download LeonJoe13/Sonic --local-dir  checkpoints; '
@@ -19,175 +21,160 @@ os.system(cmd)
 
 pipe = Sonic()
 
+# ------------------------------------------------------------------
+# 유틸
+# ------------------------------------------------------------------
 def get_md5(content):
+    """바이트/배열에서 md5 해시 문자열 반환"""
     md5hash = hashlib.md5(content)
     return md5hash.hexdigest()
 
-@spaces.GPU(duration=300)  # 긴 비디오 처리를 위해 duration 300초로 설정
+# ------------------------------------------------------------------
+# 비디오 생성
+# ------------------------------------------------------------------
+@spaces.GPU(duration=300)  # 최대 5분까지 GPU 세션 유지
 def get_video_res(img_path, audio_path, res_video_path, dynamic_scale=1.0):
-    expand_ratio = 0.5
+    expand_ratio = 0.0          # ★ 얼굴 크롭 방지
     min_resolution = 512
-    inference_steps = 25  # 2초 분량의 비디오(25 프레임)로 고정
 
-    # 오디오 길이(참고용) 출력
+    # 오디오 길이 → 프레임 수 결정 (fps=25, 최대 60초=1500프레임)
     audio = AudioSegment.from_file(audio_path)
-    duration = len(audio) / 1000.0  # 초 단위
-    print(f"Audio duration: {duration} seconds, using inference_steps: {inference_steps}")
+    duration = len(audio) / 1000.0  # 초
+    fps = 25
+    max_steps = fps * 60            # 1500
+    inference_steps = max(1, min(int(duration * fps), max_steps))
+    print(f"Audio duration: {duration:.2f}s → inference_steps: {inference_steps}")
 
+    # 얼굴 정보는 참고용으로만 출력
     face_info = pipe.preprocess(img_path, expand_ratio=expand_ratio)
     print(f"Face detection info: {face_info}")
+    if face_info["face_num"] == 0:
+        print("Warning: face not detected – proceeding with full image.")
 
-    if face_info['face_num'] > 0:
-        crop_image_path = img_path + '.crop.png'
-        pipe.crop_image(img_path, crop_image_path, face_info['crop_bbox'])
-        img_path = crop_image_path
-        os.makedirs(os.path.dirname(res_video_path), exist_ok=True)
-
-        # 고정된 inference_steps(25)로 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
+    # 출력 폴더 보장
+    os.makedirs(os.path.dirname(res_video_path), exist_ok=True)
+
+    # 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
 
-tmp_path = './tmp_path/'
-res_path = './res_path/'
+# ------------------------------------------------------------------
+# 캐시·경로 설정
+# ------------------------------------------------------------------
+tmp_path = "./tmp_path/"
+res_path = "./res_path/"
 os.makedirs(tmp_path, exist_ok=True)
 os.makedirs(res_path, exist_ok=True)
 
+# ------------------------------------------------------------------
+# Gradio 콜백
+# ------------------------------------------------------------------
 def process_sonic(image, audio, dynamic_scale):
     # 입력 검증
     if image is None:
         raise gr.Error("Please upload an image")
     if audio is None:
         raise gr.Error("Please upload an audio file")
-        
+
     img_md5 = get_md5(np.array(image))
     audio_md5 = get_md5(audio[1])
-    print(f"Processing with image hash: {img_md5}, audio hash: {audio_md5}")
-    
+    print(f"Processing (img={img_md5}, audio={audio_md5})")
+
+    # numpy 오디오 → AudioSegment
     sampling_rate, arr = audio[:2]
-    if len(arr.shape) == 1:
+    if arr.ndim == 1:
         arr = arr[:, None]
-    
-    # numpy array로부터 AudioSegment 생성
     audio_segment = AudioSegment(
         arr.tobytes(),
         frame_rate=sampling_rate,
         sample_width=arr.dtype.itemsize,
-        channels=arr.shape[1]
+        channels=arr.shape[1],
     )
-    audio_segment = audio_segment.set_frame_rate(sampling_rate)
-    
-    # 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'))
-    
-    # 입력 파일이 없으면 저장
+
+    # 경로
+    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")
+    )
+
+    # 저장 / 캐시
     if not os.path.exists(image_path):
         image.save(image_path)
     if not os.path.exists(audio_path):
         audio_segment.export(audio_path, format="wav")
-    
-    # 캐시된 결과가 있으면 반환, 없으면 새로 생성
+
     if os.path.exists(res_video_path):
         print(f"Using cached result: {res_video_path}")
         return res_video_path
-    else:
-        print(f"Generating new video with dynamic scale: {dynamic_scale}")
-        return get_video_res(image_path, audio_path, res_video_path, dynamic_scale)
 
-# 예시 데이터를 위한 dummy 함수 (필요시 실제 예시 데이터를 추가하세요)
+    print(f"Generating new video (dynamic_scale={dynamic_scale})")
+    return get_video_res(image_path, audio_path, res_video_path, dynamic_scale)
+
+# ------------------------------------------------------------------
+# Gradio UI
+# ------------------------------------------------------------------
 def get_example():
+    """예시 데이터 (필요 시 추가)"""
     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;
-}
+.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,theme="apriel") as demo:
-    gr.HTML("""
+with gr.Blocks(css=css, theme="apriel") as demo:
+    gr.HTML(
+        """
         <div class="main-header">
-            <h1>🎭 Sonic: Advanced Portrait Animation</h1>
-            <p>Transform still images into dynamic videos synchronized with audio</p>
+            <h1>🎭 Longer Sonic: Advanced Portrait Animation</h1>
+            <p>Transform still images into dynamic videos synchronized with audio(Demo max 60sec)</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",
-                elem_id="audio_input",
-                type="numpy"
+            image_input = gr.Image(type="pil", label="Portrait Image", elem_id="image_input")
+            audio_input = gr.Audio(label="Voice/Audio Input", elem_id="audio_input", type="numpy")
+            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)",
             )
-            
-            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"
-            )
-        
+            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"
-            )
-    
+            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,
-        api_name="animate"
+        api_name="animate",
     )
-    
+
     gr.Examples(
         examples=get_example(),
         fn=process_sonic,
         inputs=[image_input, audio_input, dynamic_scale],
         outputs=video_output,
-        cache_examples=False
+        cache_examples=False,
     )
-    
-
 
-# 공개 링크 생성: share=True
-demo.launch(share=True)
+# ------------------------------------------------------------------
+# Launch
+# ------------------------------------------------------------------
+demo.launch(share=True)