REST3D: Reconstructing Physically Stable 3D Scenes from a Single Image

REST3D is a framework that reconstructs physically stable 3D scenes from a single image by combining physical scene understanding with physics-constrained refinement. The method uses an "agentic" approach to construct a scene-tree representation capturing object relationships and physical states, then refines the initial 3D reconstruction to eliminate physical violations like object floating and interpenetration while maintaining visual consistency. Experiments demonstrate that REST3D produces simulation-ready scenes with improved stability compared to existing methods and enables practical applications like VR-based human-object interaction.