deploy/gpu_worker/musetalk_server.py

@@ -526,13 +526,12 @@ def _load_musetalk_models():
                         gfpgan_key = "params_ema" if "params_ema" in gfpgan_ckpt else "params"
                         gfpgan_model.load_state_dict(gfpgan_ckpt[gfpgan_key], strict=True)
                         gfpgan_model.eval()
-                        if Config.use_float16:
-                            gfpgan_model = gfpgan_model.half()
+                        # GFPGAN 始终使用 FP32 推理，避免 FP16 色偏导致紫/灰色块
                         gfpgan_model = gfpgan_model.to(device)
                         del gfpgan_ckpt
                         _gfpgan_loaded = True
                         _gfpgan_load_error = None
-                        logger.info("GFPGAN 加载完成 (FP16=%s)", Config.use_float16)
+                        logger.info("GFPGAN 加载完成 (FP32，避免色偏)")
                     else:
                         _gfpgan_loaded = False
                         _gfpgan_load_error = f"模型文件不存在: {gfpgan_path}"
@@ -920,25 +919,35 @@ def _run_inference(
                 _ff_proc.stdin.write(ori_frame.tobytes())
                 continue
 
-            # GFPGAN 人脸超分增强
+            # GFPGAN 人脸超分增强（FP32 + 色彩校正，避免 FP16 色偏）
             if gfpgan_enhancer is not None:
                 try:
                     _fh, _fw = res_frame_resized.shape[:2]
                     _face_up = cv2.resize(res_frame_resized, (512, 512),
                                           interpolation=cv2.INTER_LANCZOS4)
+                    # 保存原始人脸区域用于色彩校正
+                    _face_original_bgr = _face_up.copy()
                     _face_rgb = cv2.cvtColor(_face_up, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
                     _face_t = torch.from_numpy(_face_rgb.transpose(2,0,1)).unsqueeze(0)
+                    # GFPGAN 始终 FP32，避免 FP16 精度导致色偏；归一化到 [-1, 1]
                     _face_t = ((_face_t - 0.5) / 0.5).to(device)
-                    if Config.use_float16:
-                        _face_t = _face_t.half()
                     with torch.no_grad():
-                        _out = gfpgan_enhancer(_face_t, return_rgb=False, weight=0.5)[0]
+                        _out = gfpgan_enhancer(_face_t, return_rgb=False, weight=0.35)[0]
                     _out = _out.squeeze(0).float().cpu().clamp_(-1,1)
                     _out = ((_out + 1)/2*255).numpy().transpose(1,2,0)
                     _out_bgr = cv2.cvtColor(_out.astype(np.uint8), cv2.COLOR_RGB2BGR)
+
+                    # 色彩校正：将增强结果的均值/标准差对齐到原始人脸，消除色调偏移
+                    _orig_mean = _face_original_bgr.mean(axis=(0, 1))
+                    _orig_std = _face_original_bgr.std(axis=(0, 1)) + 1e-6
+                    _enh_mean = _out_bgr.mean(axis=(0, 1))
+                    _enh_std = _out_bgr.std(axis=(0, 1)) + 1e-6
+                    _out_bgr = ((_out_bgr.astype(np.float32) - _enh_mean) * (_orig_std / _enh_std) + _orig_mean)
+                    _out_bgr = np.clip(_out_bgr, 0, 255).astype(np.uint8)
+
                     res_frame_resized = cv2.resize(_out_bgr, (_fw, _fh),
                                                     interpolation=cv2.INTER_LANCZOS4)
-                    del _face_t, _out, _out_bgr
+                    del _face_t, _out, _out_bgr, _face_original_bgr
                 except Exception as _gfpgan_err:
                     logger.warning("GFPGAN 增强失败（帧 %d），使用原图: %s", i, _gfpgan_err)
 
@@ -985,7 +994,7 @@ def _run_inference(
     if video_downsampled:
         logger.info("输入视频 %.1f fps，降帧至 %.1f fps 推理后直接输出（不做插帧还原）", original_fps, inference_fps)
 
-    shutil.copy2(str(final_video), str(output_path))
+    _mux_video_with_audio(final_video, audio_path, output_path)
 
     try:
         torch.cuda.empty_cache()