LLaVA-OneVision-2.5 shows codec input beats uniform frame sampling on video tasks

liuziwei7 · x · 2026-11-19

LMMS Lab released LLaVA-OneVision-2.5, the first clear evidence that codec (compressed bitstream) input beats uniform frame sampling on general video understanding. Pixeltable adds a practical caveat: codec mode selects patches by encoder bit-cost (dense I-frames, motion-rich P-frame patches), a signal lost if you decode to JPEG and discard the source; frame sampling still wins some benchmarks (TOMATO 39.2 vs 36.0, LVBench 62.0 vs 61.6), so keep source video and treat frames as recomputable.

Original post →

More from Models

Models channel →