deploy/gpu_worker/musetalk_server.py

@@ -919,35 +919,30 @@ def _run_inference(
                 _ff_proc.stdin.write(ori_frame.tobytes())
                 continue
 
-            # GFPGAN 人脸超分增强（FP32 + 色彩校正，避免 FP16 色偏）
+            # GFPGAN 人脸超分增强（FP32 推理，避免 FP16 色偏）
+            # 色彩通道约定：ori_frame / res_frame / _face_up 均为 BGR（OpenCV 默认）；
+            # GFPGAN 输出用 return_rgb=True 拿到 RGB，再转 BGR，与后续 face_parsing 融合保持一致。
             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)
+                    _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)
                     with torch.no_grad():
-                        _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)
+                        _out = gfpgan_enhancer(_face_t, return_rgb=True, weight=0.35)[0]
+                    # 输出 tensor: RGB, [-1, 1] 范围 → clamp → 映射到 [0, 255] uint8
+                    _out = _out.squeeze(0).float().cpu().clamp_(-1.0, 1.0)
+                    _out = ((_out + 1.0) / 2.0 * 255.0).numpy().transpose(1, 2, 0)
+                    _out_rgb = _out.astype(np.uint8)
+                    # RGB → BGR，与 ori_frame 保持一致，确保 face_parsing 融合时通道正确
+                    _out_bgr = cv2.cvtColor(_out_rgb, cv2.COLOR_RGB2BGR)
 
                     res_frame_resized = cv2.resize(_out_bgr, (_fw, _fh),
                                                     interpolation=cv2.INTER_LANCZOS4)
-                    del _face_t, _out, _out_bgr, _face_original_bgr
+                    del _face_t, _out, _out_rgb, _out_bgr
                 except Exception as _gfpgan_err:
                     logger.warning("GFPGAN 增强失败（帧 %d），使用原图: %s", i, _gfpgan_err)