Reshoot-Anything: Self-supervised model for in-the-wild video reshooting

rsasaki0109 · x · 2026-08-19

Reshoot-Anything is a self-supervised model that reshoots dynamic monocular videos under novel camera trajectories. It learns implicit 4D spatiotemporal priors by training on pseudo multi-view triplets extracted from monocular footage. During inference, a 4D point-cloud anchor from the source video provides geometric conditioning for the new trajectory. The repo includes code, visualizers, and pipelines.

Original post →

More from Multimodal

Multimodal channel →