Shikun Feng

Keyframe-residual captioning unlocks high-fidelity video-language supervision, surpassing direct VLM captioning in capturing fine-grained visual details.

LLaVA-OV-2's codec-stream tokenization lets it crush existing video-language models, especially in tasks requiring fine-grained temporal understanding of high-frequency motion.

A 1.7B parameter model can now rival much larger audio language models, thanks to a novel architecture and data synthesis pipeline.