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

app.py

@@ -130,7 +130,14 @@ def toggle_name_box(selected_options: list[str]):
 	return gr.update(visible=NAME_OPTION in selected_options)
 
 
-@spaces.GPU()
+def get_duration(input_image, prompt, temperature, top_p, max_new_tokens, log_prompt):
+	# Generation speed is ~40 tok/s on ZeroGPU; a full 512-token caption measured
+	# ~13s wall. Scale the GPU request with the token budget instead of the 60s
+	# default so anonymous visitors with small quotas aren't rejected up front.
+	return min(90, 12 + int(max_new_tokens) // 25)
+
+
+@spaces.GPU(duration=get_duration)
 @torch.no_grad()
 def chat_joycaption(input_image: Image.Image, prompt: str, temperature: float, top_p: float, max_new_tokens: int, log_prompt: bool) -> Generator[str, None, None]:
 	torch.cuda.empty_cache()