Part-Tokenised Graph–Field Flow Matching for Coherent Physics-Grounded 3D Asset Generation

Jongmin Yu, Hyeontaek Oh, Zhongtian Sun, Anoushka Harit, Jinhong Yang

Reviewer: ~Yechan_Kim2

Keywords: physics object generation, physics-grounded asset generation

TL;DR: submission for non-archival

Abstract: Physically grounded 3D asset generation aims to recover simulation-ready objects from a single image by combining geometric fidelity with physically meaningful attributes, including absolute scale, density, affordance, and articulation. We propose Part-Tokenised Graph–Field Flow Matching (PTGFFM), a two-stage framework that augments a frozen geometry-centric structural prior with a trainable part-aware physics branch. PTGFFM represents physical attributes as an unordered set of latent part tokens with existence prediction, and decodes them into dense physical fields and a kinematic graph conditioned on structured 3D features. To improve robustness to long-tailed object sizes, we regress absolute scale in a stabilised asinh space. A conditional flow-matching model is then trained to generate the physical token set from noise. On PhysXNet, PTGFFM improves several physical-property metrics over prior baselines while maintaining geometric fidelity comparable to the underlying structural prior. These results suggest that part-tokenised physical generation is a promising direction for simulation-ready 3D asset synthesis.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 2