Right-size ZeroGPU duration requests (dynamic where workload-dependent)

app.py

@@ -189,7 +189,17 @@ def update_steps_from_mode(mode_choice):
     return DEFAULT_STEPS[mode_choice], DEFAULT_CFG[mode_choice]
 
 
-@spaces.GPU(duration=85)
+def get_duration(prompt, input_images=None, mode_choice="Distilled (4 steps)", seed=42, randomize_seed=False, width=1024, height=1024, num_inference_steps=4, guidance_scale=4.0, prompt_upsampling=False, progress=None):
+    # Observed ~1.2-2.0 it/s at 1MP in production logs, so a distilled 4-step
+    # run finishes in seconds while the old flat 85s request exhausted anonymous
+    # visitors' ZeroGPU quota check up front. Scale with steps and resolution.
+    res_scale = (int(width) * int(height)) / (1024 * 1024)
+    n_images = len(input_images) if input_images else 0
+    estimate = 8 + n_images * 2 + int(num_inference_steps) * 1.1 * res_scale
+    return int(min(max(estimate, 15), 85))
+
+
+@spaces.GPU(duration=get_duration)
 def infer(prompt, input_images=None, mode_choice="Distilled (4 steps)", seed=42, randomize_seed=False, width=1024, height=1024, num_inference_steps=4, guidance_scale=4.0, prompt_upsampling=False, progress=gr.Progress(track_tqdm=True)):
 
     if randomize_seed: