2026-09-23 00:08:09 +08:00
1 changed files with 3 additions and 2 deletions
@@ -805,8 +805,9 @@ def _run_inference(
            crop = frame[y1:y2_eff, x1:x2]
            if crop.size == 0:
                continue
-            crop_rgb = cv2.cvtColor(crop, cv2.COLOR_BGR2RGB)
-            crop_resized = cv2.resize(crop_rgb, (256, 256), interpolation=cv2.INTER_LANCZOS4)
+            # 直接传 BGR 给 VAE：VAE 内部 preprocess_img 已做 BGR→RGB 转换，
+            # 此处再 cvtColor 会造成双重转换、R/B 通道互换（蓝色块根因）。
+            crop_resized = cv2.resize(crop, (256, 256), interpolation=cv2.INTER_LANCZOS4)
            # 使用 VAE 的 get_latents_for_unet 得到 8 通道输入
            # get_latents_for_unet 内部: preprocess(half_mask=True) encode + preprocess(half_mask=False) encode → cat → [1,8,32,32]
            latents = vae.get_latents_for_unet(crop_resized).detach().cpu()