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

gradio_app.py

@@ -149,7 +149,9 @@ def show_mask_img(input_image: Image) -> Image:
     new_img = img_numpy[..., :3] * alpha[:, :, None] + chkb * (1 - alpha[:, :, None])
     return Image.fromarray(new_img.astype(np.uint8), mode="RGB")
 
-@spaces.GPU
+# Mesh generation alone runs in ~5-10s; the 60s default request was rejecting
+# anonymous visitors whose remaining ZeroGPU quota was below the request.
+@spaces.GPU(duration=20)
 def run_button(
     run_btn,
     input_image,
@@ -243,7 +245,9 @@ def update_foreground_ratio(img_proc, fr):
     )
 
 
-@spaces.GPU(duration=90)
+# Full pipeline (bg removal + mesh + texture bake) measured ~12s end to end;
+# 30s keeps 2.5x headroom without tripping visitors' quota check.
+@spaces.GPU(duration=30)
 def image_to_glb(input_image, foreground_ratio, remesh_option, vertex_count, texture_size):
     # One-call API pipeline for headless callers (Upsampler free tool): background
     # removal + foreground resize + mesh generation, all in a single GPU allocation.