A controlled study reveals that query-selected frame allocation is the most impactful lever for long-video models. On LongVideoBench, selecting eight query-specific frames outperformed sixteen uniformly spaced frames by 6.9 points, demonstrating that intelligent selection outweighs spatial compression or reinvestment.
HOW THIS AFFECTS YOU
●
builderYou should prioritize intelligent frame selection logic over simple uniform sampling to save tokens.
●
researcherThis clarifies that frame selection is a primary driver of performance in long-context video models.