ContactGen: Contact-Constrained Diffusion for 3D Human Generation with a Partner

Abstract

Modeling physical interactions between two humans is essential for realistic 3D human generation, yet it remains challenging due to the high variability of contact configurations and the need to satisfy non- trivial geometric constraints. Existing approaches often (i) rely on simplified interaction templates, (ii) enforce physical plausibility only through post-optimization, or (iii) utilize coarse joint-based represen- tations, limiting their ability to generate diverse and physically consistent interactions. To address this, I introduce ContactGen, a constrained conditional diffusion framework for generat- ing a 3D interacting human given a fixed partner and an interaction label. The proposed framework in- corporates a dynamic, pose-adaptive contact constraint informed by a learned contact prediction network into the diffusion sampling process, enabling the model to enforce realistic contact while preserving the natural variability of human–human interactions. Experiments demonstrate that ContactGen improves label-consistent physical contact while main- taining high distributional fidelity, highlighting its effectiveness for generating diverse and physically plausible human–human interactions.