Interleaved Latent Visual Reasoning proposed for video event prediction

Arxiv CS.CV (computer vision)

The paper introduces Interleaved Latent Visual Reasoning (ILVR), which performs future state prediction in latent visual space rather than verbalizing intermediate steps. ILVR uses frame-level temporal abstraction and latent state propagation to capture fine-grained motion and uncertainty.

latent-reasoning

Arxiv CS.CV (computer vision)·Tianxiang JiangLinquan WuSheng XiaSongze LiZiang YanHaoyu YangYu QiaoYi Wang

6 days ago