fix: increase GPU duration estimate for 22B model with CPU offload overhead

app.py

@@ -277,9 +277,11 @@ def _estimate_duration(*args, **kwargs):
         num_frames = int(num_frames)
         num_inference_steps = int(num_inference_steps)
     except (TypeError, ValueError):
-        return 120
-    base = 30 + num_inference_steps * 2
-    frame_factor = max(1.0, num_frames / 121)
+        return 240
+    # 22B model with enable_model_cpu_offload(): first step ~28s (cold),
+    # subsequent ~5s each, plus VAE decode + video encoding ~20s.
+    base = 60 + num_inference_steps * 8
+    frame_factor = max(1.0, num_frames / 41)
     return min(240, int(base * frame_factor))