Proper handling of large images (4k) with llama-server; --image-max-tokens 10580 doesn't fix the issue.

[wbste]

Hello!
For models like qwen3-vl running via llama-server, the image_min_pixels and image_max_pixels parameter defaults are set as follows I believe:

• Minimum pixels: 8 tokens × (14 × 2)² = 8 × 784 = 6,272 pixels
• Maximum pixels: 4096 tokens × (14 × 2)² = 4096 × 784 = 3,211,264 pixels

Any image sent with a native resolution within this range does great with bbox and point tasks.

If you upload a 4k image (3840 x 2160) the coordinates are off. To fix:

• patch_area = 14 * 14 * 2 * 2 = 784
• Required tokens = 8,294,400 / 784 ≈ 10,580 tokens

If I change add --image-max-tokens 10580 to the server the coordinates returned are off as well.

Am I doing something wrong? It works 100% of the time when the image is within the default token limit, but larger it fails.

Thanks for your time!

[wbste]

Got it figured out. Basically need to use this: https://github.com/QwenLM/Qwen3-VL/blob/main/qwen-vl-utils/src/qwen_vl_utils/vision_process.py

The smart resizing fixed it all!