Update

app.py

@@ -177,6 +177,7 @@ def generate(
     history: list[dict],
     thinking: bool = False,
     max_new_tokens: int = 1024,
+    max_soft_tokens: int = 280,
     system_prompt: str = "",
 ) -> Iterator[str]:
 
@@ -193,6 +194,7 @@ def generate(
         "return_tensors": "pt",
         "add_generation_prompt": True,
         "load_audio_from_video": _has_media_type(messages, "video"),
+        "processor_kwargs": {"images_kwargs": {"max_soft_tokens": max_soft_tokens}},
     }
     if thinking:
         template_kwargs["enable_thinking"] = True
@@ -313,7 +315,13 @@ demo = gr.ChatInterface(
     multimodal=True,
     additional_inputs=[
         gr.Checkbox(label="Thinking", value=False),
-        gr.Slider(label="Max New Tokens", minimum=100, maximum=4000, step=10, value=1024),
+        gr.Slider(label="Max New Tokens", minimum=100, maximum=4000, step=10, value=2000),
+        gr.Dropdown(
+            label="Image Token Budget",
+            info="Higher values preserve more visual detail (useful for OCR/documents). Lower values are faster.",
+            choices=[70, 140, 280, 560, 1120],
+            value=280,
+        ),
         gr.Textbox(label="System Prompt", value=""),
     ],
     additional_inputs_accordion=gr.Accordion("Settings", open=True),