perf: replace PIL+PNG frame files with direct ffmpeg stdin pipe

Stream raw RGB bytes from tensor directly to ffmpeg stdin.
Eliminates all intermediate PNG file I/O — much faster for large frame counts.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

nodes/feature_extractor.py

@@ -79,37 +79,28 @@ def _hash_inputs(video_tensor, cot_text):
 
 
 def _save_video_tensor_to_mp4(video_tensor, output_path, fps=30):
-    """Save ComfyUI IMAGE tensor [T,H,W,C] to MP4 via PIL + ffmpeg.
+    """Save ComfyUI IMAGE tensor [T,H,W,C] to MP4 by piping raw RGB to ffmpeg.
 
-    torchvision.io.write_video requires the optional 'av' (PyAV) package
+    Avoids intermediate PNG files — frames are streamed directly to ffmpeg stdin.
-    which is not installed in most ComfyUI environments. ffmpeg is always
-    available in ComfyUI Docker images.
     """
-    from PIL import Image
-    import shutil
-
     frames_np = (video_tensor.cpu().numpy() * 255).astype("uint8")
-
+    T, H, W, C = frames_np.shape
-    frame_dir = output_path + "_frames"
-    os.makedirs(frame_dir, exist_ok=True)
-    try:
-        for i, frame in enumerate(frames_np):
-            Image.fromarray(frame).save(os.path.join(frame_dir, f"{i:06d}.png"))
 
     result = subprocess.run(
         [
             "ffmpeg", "-y",
-                "-framerate", str(fps),
+            "-f", "rawvideo", "-vcodec", "rawvideo",
-                "-i", os.path.join(frame_dir, "%06d.png"),
+            "-s", f"{W}x{H}", "-pix_fmt", "rgb24",
+            "-r", str(fps),
+            "-i", "pipe:0",
             "-c:v", "libx264", "-pix_fmt", "yuv420p",
             output_path,
         ],
-            capture_output=True, text=True,
+        input=frames_np.tobytes(),
+        capture_output=True,
     )
     if result.returncode != 0:
-            raise RuntimeError(f"[PrismAudio] ffmpeg failed:\n{result.stderr}")
+        raise RuntimeError(f"[PrismAudio] ffmpeg failed:\n{result.stderr.decode()}")
-    finally:
-        shutil.rmtree(frame_dir, ignore_errors=True)
 
 
 class PrismAudioFeatureExtractor: