LLaVA-OneVision-2.5: one model for streaming and offline video, codec beats frame sampling

liuziwei7 · x · 2026-11-19

LMMS Lab released LLaVA-OneVision-2.5, a unified VLM where the same weights handle both 24fps streaming and offline video analysis—no separate models needed. Key finding: after exploring both frame sampling and codec-based video input in OV-2 (where codec mainly helped temporal grounding), OV-2.5 shows codec input outperforms uniform frame sampling on general video tasks, treating video as a compressed representation rather than a pile of independent frames.

Related event: LLaVA-OneVision-2.5 Shows Codec Input Beats Uniform Frame Sampling(4 posts)→

Original post →

More from Multimodal

Multimodal channel →