PhysConvex introduces boundary-driven dynamic convex fields integrated with mesh-free reduced-order convex simulation for dynamic reconstruction, system identification(a), physical generalization(b,c). It recovers appearance, geometry, and physics from videos, improving dynamic and physical reconstruction(d,e), training efficiency(f).
Reconstructing deformable objects from video requires a 4D representation that should preserve geometry, explain motion through physics, and generalize to future or new physical conditions. Existing dynamic NeRF and Gaussian methods achieve strong view synthesis, yet their voxel or ellipsoidal primitives are primarily designed for appearance rendering and typically driven by centers or predefined particle bindings, limiting physically meaningful non-uniform deformation and sharp boundary evolution. We present PhysConvex, a physics-informed dynamic convex field for video-based reconstruction, physical system identification, and simulation. PhysConvex represents a dynamic object as material-space deformable convex primitives whose boundary is advected by physical dynamics. The boundary-driven convex expresses non-uniform deformation and evolving active supports while serving simultaneously as a rendering element, deformation carrier, and physical support where mass, elastic response, forces, and contacts are evaluated. We further introduce a mesh-free reduced-order convex simulator in which neural skinning modes define physics-based deformation bases directly over deformable convex supports. Experiments show improved dynamic and physical reconstruction, efficiency, future prediction, and generalization to changed materials, forces, and boundary conditions.
Video-based deformable reconstruction requires a 4D representation that preserves geometry, captures physics, and generalizes to unseen conditions. Existing dynamic NeRF and Gaussian methods render well, yet their voxel or ellipsoidal primitives are primarily designed for appearance rendering and typically driven by centers or predefined particle bindings, limiting physically meaningful non-uniform deformation and sharp boundary evolution.
Key insight: the rendering primitive should also be the material support for physical deformation and simulation. As such, we propose PhysConvex, a dynamic 4D representation that preserves appearance, geometry, and physical behavior. Specifically, we introduce: 1. Boundary-driven dynamic convex representation that gives each vertex spatially adaptive deformation capacity, rather than forcing primitives to move as a center-driven kernel. 2. Mesh-free reduced-order convex simulation where dynamic convexes are advected under continuum mechanics via a continuous implicit neural skinning field that encodes physics-informed, shape- & material-aware deformation modes.
PhysConvex introduces boundary-driven dynamic convex fields integrated with mesh-free reduced-order convex simulation for dynamic reconstruction, system identification, physical generalization. It recovers appearance, geometry, and physics from videos, improving dynamic and physical reconstruction, and training efficiency. In stage 1, an undeformed convex field is reconstructed over the first multi-view frame. In stage 2, the reduced-order convex simulation advects our boundary-driven dynamic convex field under continuum mechanics, and physical properties can be optimized through joint differential simulation and rendering using video supervision.
@misc{wang2026physconvexphysicsinformed3ddynamic,
title={PhysConvex: Physics-Informed 3D Dynamic Convex Radiance Fields for Reconstruction and Simulation},
author={Dan Wang and Xinrui Cui and Serge Belongie and Ravi Ramamoorthi},
year={2026},
eprint={2602.18886},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2602.18886},
}