Size ZeroGPU duration from measured 1.89 s/step + 3 s fixed

app.py

@@ -120,8 +120,11 @@ MAX_SEED = 2**31 - 1
 DEFAULT_SEED = 20260508
 
 
-STEP_SECONDS = 5.0  # replaced with the measured per-step cost after smoke test
-BASE_SECONDS = 45.0  # text encode + VAE decode + mux
+# Measured on this Space's zero-a10g: 20 steps -> 40.9 s, 50 steps -> 97.6 s
+# wall inside `generate` => ~1.89 s/step plus ~3 s fixed (text encode, VAE
+# decode of 121 frames, WAV write, ffmpeg mux). Sized with a ~15-20% margin.
+STEP_SECONDS = 2.2
+BASE_SECONDS = 5.0
 
 
 def _duration(*args, **kwargs):
@@ -137,7 +140,7 @@ def _duration(*args, **kwargs):
         steps = int(steps)
     except (TypeError, ValueError):
         steps = 50
-    return int(min(900, BASE_SECONDS + STEP_SECONDS * steps))
+    return int(min(300, BASE_SECONDS + STEP_SECONDS * steps))
 
 
 @spaces.GPU(duration=_duration)
@@ -254,7 +257,7 @@ with gr.Blocks() as demo:
             than dubbed afterwards.
 
             Output is the checkpoint's native resolution: **121 frames · 256x256 · 24 fps ·
-            48 kHz audio (~5 s)**. A run takes a few minutes.
+            48 kHz audio (~5 s)**. A default 50-step run takes about 1.5 minutes.
             """
         )