Normalize Wan frames and fix video export

app.py

@@ -125,6 +125,24 @@ def _cover_resize(image: Image.Image, size: tuple[int, int]) -> Image.Image:
     return resized.crop((left, top, left + target_width, top + target_height))
 
 
+def _to_pil_frame(frame) -> Image.Image:
+    """Normalize Diffusers frame outputs to an RGB PIL image."""
+    if isinstance(frame, Image.Image):
+        return frame.convert("RGB")
+    if isinstance(frame, torch.Tensor):
+        frame = frame.detach().cpu().numpy()
+
+    array = np.asarray(frame)
+    if array.ndim == 3 and array.shape[0] in (1, 3, 4) and array.shape[-1] not in (1, 3, 4):
+        array = np.moveaxis(array, 0, -1)
+    if np.issubdtype(array.dtype, np.floating):
+        if array.size and float(array.min()) < 0:
+            array = (array + 1.0) / 2.0
+        array = np.clip(array, 0.0, 1.0) * 255.0
+    array = np.clip(array, 0, 255).astype(np.uint8)
+    return Image.fromarray(array).convert("RGB")
+
+
 def _frame_count(duration_seconds: float) -> int:
     desired = int(round(float(duration_seconds) * FPS))
     return int(np.clip(round((desired - 1) / 4) * 4 + 1, 17, 81))
@@ -230,6 +248,7 @@ def generate_extended_video(
                 (current_seed + index) % MAX_SEED
             ),
         ).frames[0]
+        frames = [_to_pil_frame(frame) for frame in frames]
 
         clip = tempfile.NamedTemporaryFile(suffix=f"-clip-{index + 1}.mp4", delete=False)
         clip.close()

requirements.txt

@@ -5,7 +5,7 @@ safetensors
 sentencepiece
 peft
 ftfy
+imageio
 imageio-ffmpeg
 opencv-python-headless
 torchao==0.11.0
-