LLaVA-OneVision-2: Towards Next-Generation Perceptual Intelligence

Researchers unveiled LLaVA-OneVision-2, a next-generation vision-language model for multimodal perception.

It uses codec-stream tokenization, windowed attention, and shared 3D positional encoding to better handle long videos and spatial inputs.

Trained on millions of open video and spatial samples, it delivered strong gains in video understanding, grounding, and tracking benchmarks.

The model shows improved fine-grained temporal localization and unified performance across video and spatial tasks, outperforming a leading rival.