max_pixels is a token budget in disguise — and the right cap depends on the size of what you're looking for
The `max_pixels` configuration in Qwen2.5-VL models is a token budget in disguise, with default settings often leading to a significantly higher budget than recommended. This can result in suboptimal performance, especially for large targets within an image. The optimal token budget is dependent on the size of the specific object being sought, with smaller targets benefiting from larger budgets while larger targets perform best at lower token counts. AI
IMPACT Optimizing `max_pixels` can improve accuracy and efficiency for multimodal models, especially in applications involving object detection or grounding.