Upload folder using huggingface_hub

app.py

@@ -57,7 +57,19 @@ EOS_TOKEN_ID = tokenizer.eos_token_id
 print("Model ready.", flush=True)
 
 
-@spaces.GPU(duration=90)
+def _estimate_duration(
+    image=None, question=None, max_new_tokens=64, max_visual_tokens=4096, *args, **kwargs
+) -> int:
+    """ZeroGPU reservation: cold-start weight streaming plus input-dependent decode."""
+    try:
+        tokens = int(max_new_tokens)
+        visual = int(max_visual_tokens)
+    except (TypeError, ValueError):
+        tokens, visual = 64, 4096
+    return int(min(75, 22 + tokens * 0.045 + visual * 0.0015))
+
+
+@spaces.GPU(duration=_estimate_duration)
 def answer_question(
     image,
     question: str,
@@ -116,8 +128,10 @@ the native visual rows as fixed evidence. Only the final recurrent state reaches
 answer decoder — the original visual rows and multimodal prefix cache are cut off, which
 makes the latent visual reasoning *causally necessary* rather than merely present.
 
-Ask an open-ended question about an image, or use the benchmark-style multiple-choice
-format the model was tuned on. Decoding is deterministic (greedy) by design.
+Ask a question about an image — open-ended, or in the benchmark-style multiple-choice
+format the model was tuned on. Decoding is deterministic (greedy) by design. The recurrent
+transition is supervised on short VQA answers, so crisp, targeted questions work best;
+long free-form descriptions drift.
 
 [Paper](https://huggingface.co/papers/2609.06746) · [Code](https://github.com/dmis-lab/CVRR) · [Model](https://huggingface.co/dmis-lab/Qwen2.5-VL-7B-CVRR)
 """
@@ -149,7 +163,7 @@ with gr.Blocks() as demo:
                 question_in = gr.Textbox(
                     label="Question",
                     lines=3,
-                    placeholder="What is the person on the left holding?",
+                    placeholder="Is the peacock's tail open or closed?",
                 )
                 run_btn = gr.Button("Answer", variant="primary")
                 answer_out = gr.Textbox(label="Answer", lines=4)