Update app.py

app.py

@@ -19,30 +19,37 @@ os.makedirs(UPLOAD_DIR, exist_ok=True)
 os.makedirs(OUTPUT_DIR, exist_ok=True)
 
 # ===============================
-# Device / dtype (CRITICAL)
+# Device / dtype
 # ===============================
 
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 DTYPE = torch.float16 if DEVICE == "cuda" else torch.float32
 
 # ===============================
-# Job store
+# In-memory job store
 # ===============================
 
 jobs = {}
 
 # ===============================
-# Load models (safe on startup)
+# Load models (startup-safe)
 # ===============================
 
 whisper_model = whisper.load_model("base")
 yolo = YOLO("yolov8n.pt")
 
 svd = StableVideoDiffusionPipeline.from_pretrained(
-    "stabilityai/stable-video-diffusion-img2vid",
-    dtype=DTYPE
+    "stabilityai/stable-video-diffusion-img2vid"
 )
-svd.to(DEVICE)
+svd.to(device=DEVICE, dtype=DTYPE)
+
+# ===============================
+# Health check
+# ===============================
+
+@app.get("/")
+def root():
+    return {"status": "ok"}
 
 # ===============================
 # Endpoints
@@ -99,8 +106,16 @@ async def smart_crop(
     ret, frame = cap.read()
     cap.release()
 
+    if not ret:
+        return {"error": "Failed to read video frame"}
+
     results = yolo(frame)
-    box = results[0].boxes.xyxy[0].cpu().numpy()
+    boxes = results[0].boxes
+
+    if boxes is None or len(boxes) == 0:
+        return {"error": "No subject detected"}
+
+    box = boxes.xyxy[0].cpu().numpy()
 
     return {
         "crop_box": box.tolist(),
@@ -113,12 +128,14 @@ async def smart_crop(
 
 def run_edit_job(job_id: str, video_path: str, frame_path: str):
     try:
+        # Extract single frame safely
         subprocess.run(
             [
                 "ffmpeg", "-y",
                 "-i", video_path,
                 "-vf", "scale=512:512:force_original_aspect_ratio=decrease",
                 "-frames:v", "1",
+                "-update", "1",
                 frame_path
             ],
             check=True
@@ -129,8 +146,8 @@ def run_edit_job(job_id: str, video_path: str, frame_path: str):
         with torch.no_grad():
             output = svd(
                 image=img,
-                num_frames=16,
-                decode_chunk_size=8
+                num_frames=8,          # CPU-safe default
+                decode_chunk_size=4
             )
 
         jobs[job_id]["status"] = "done"