tighten ZeroGPU duration to measured latency (~1s/step/window)

app.py

@@ -413,10 +413,17 @@ def _write_mp4(frames_u8: np.ndarray, fps: int = 15) -> str:
 
 
 def _estimate_duration(*a, **k):
-    """ZeroGPU reservation: scales with rollout rounds x sampler steps."""
-    rounds = int(a[10]) if len(a) > 10 else 2
-    steps = int(a[11]) if len(a) > 11 else 20
-    return int(min(280, 45 + max(1, rounds) * (8 + max(1, steps) * 2.2)))
+    """ZeroGPU reservation, fitted to latency measured on this Space.
+
+    Client-side end-to-end at 20 sampler steps: 1 window 28.6 s, 2 windows 50.6 s,
+    3 windows 67.2 s -> ~1.0 s per sampler step per window, plus ~12 s of scene
+    prep, VAE encode/decode and mp4 muxing. Keeps ~20% headroom, no more.
+    """
+    rounds = int(a[10]) if len(a) > 10 else int(k.get("rollout_rounds", 2))
+    steps = int(a[11]) if len(a) > 11 else int(k.get("num_steps", 20))
+    rounds = max(1, min(rounds, 3))
+    steps = max(1, min(steps, 40))
+    return int(min(180, 12 + rounds * (4 + steps * 1.0)))
 
 
 @spaces.GPU(duration=_estimate_duration)