LMMS-Lab: codec sampling beats uniform frame sampling on general video tasks

liuziwei7 · x · 2026-11-18

LMMS-Lab reports the first clear evidence that codec-based sampling outperforms uniform frame sampling on general video tasks. Most video-LLMs today rely on uniformly sampled frames as visual tokens; this finding offers cross-task empirical support for codec-style video tokenization. The original thread was truncated and a fuller release was teased—treat details as pending.

Original post →

More from Multimodal

Multimodal channel →