A new arXiv preprint introduces Gen4P, a framework that recovers 4D geometry and motion from ordinary monocular video. The key idea is to take a pretrained video generator and repurpose it as a feed-forward 4D perceiver, avoiding the usual iterative denoising steps.
The authors describe Gen4P as a unified approach, suggesting it handles both geometry and motion in a single model. Because the abstract is brief, details on architecture and evaluation are not yet available from the source.
This direction could make 4D scene understanding more practical by leveraging existing video generation models rather than training specialized systems from scratch.