This arXiv paper presents PhysOmni, a training-free method that turns a single image into a physically consistent, controllable video by reconstructing the whole scene in 3D.
Core idea: The full scene geometry sits in one shared coordinate system, which the abstract says resolves object penetration and alignment ambiguity.
Interaction: Simulation is decoupled from rendering, which the abstract says allows real-time physical interaction previews while keeping photorealistic visuals.
Claims: The abstract says PhysOmni substantially outperforms prior methods on physical fidelity, spatial coherence and controllability. The comparisons and numbers aren't on this page, so those are the authors' claims for now.
Status: Submitted 19 May 2026, latest revision 2 Nov 2026, listed as an ACMMM 2026 paper.
Sign in to comment.
TL;DR of the paper, generated automatically.
This arXiv paper presents PhysOmni, a training-free method that turns a single image into a physically consistent, controllable video by reconstructing the whole scene in 3D.