Tighten ZeroGPU duration to 60s based on measured inference time

app.py

@@ -49,7 +49,10 @@ pipeline = RePlanPipeline(
 )
 
 
-@spaces.GPU(duration=120)
+# Measured GPU time for a single-region 4-step edit is ~5-6s. Multi-region
+# instructions produce longer VLM reasoning (up to 2048 tokens) + more diffusion
+# work, so we keep a comfortable-but-lean margin.
+@spaces.GPU(duration=60)
 def edit_image(
     image: Image.Image,
     instruction: str,