ContactGen: Contact-Guided Interactive 3D Human Generation for Partners

TL;DR

ContactGen introduces a task of generating interactive 3D humans that maintain natural contact with a given partner under a specified interaction label. It couples a conditional guided diffusion model with a Transformer-based contact prediction module to dynamically identify contact regions and steer sampling via an objective that enforces physical contact. The approach yields physically plausible, diverse interactions and outperforms adapted baselines on the CHI3D dataset, with ablations confirming the critical role of contact-guided guidance. This work advances realistic, contact-aware human–human interaction generation with potential applications in avatar creation and social robotics.

Abstract

Among various interactions between humans, such as eye contact and gestures, physical interactions by contact can act as an essential moment in understanding human behaviors. Inspired by this fact, given a 3D partner human with the desired interaction label, we introduce a new task of 3D human generation in terms of physical contact. Unlike previous works of interacting with static objects or scenes, a given partner human can have diverse poses and different contact regions according to the type of interaction. To handle this challenge, we propose a novel method of generating interactive 3D humans for a given partner human based on a guided diffusion framework. Specifically, we newly present a contact prediction module that adaptively estimates potential contact regions between two input humans according to the interaction label. Using the estimated potential contact regions as complementary guidances, we dynamically enforce ContactGen to generate interactive 3D humans for a given partner human within a guided diffusion model. We demonstrate ContactGen on the CHI3D dataset, where our method generates physically plausible and diverse poses compared to comparison methods.

Figures (5)

• Figure 1: Examples of generated 3D interactive humans using the proposed ContactGen. Given a partner human (red-colored human) and interaction label as conditions, ContactGen generates various interactive 3D humans (blue-colored human) through a guided DDM, where we focus on physical interaction by contact.
• Figure 2: Overview of the ContactGen. Given a partner $\mathbf{x}_p$ and interaction label $\mathbf{l}$, our method generates an interactive 3D human $\mathbf{x}_h^0$ using a guided DDM. In particular, we adaptively estimate potential contact regions between humans by the contact prediction module (orange box). Based on the potential contact regions, we perform interactive optimization by considering physical contact, which provides additional guidance during the sampling.
• Figure 3: Illustration of contact prediction module. We use the contact prediction module to predict the part-wise contact probability map $\mathbf{C}$ between interacting humans. After estimating $\mathbf{C}$, a set of potential contact regions $\mathbb{C}$ (right) can be obtained from a certain threshold $\tau$.
• Figure 4: Qualitative evaluation with comparison methods on the modified CHI3D dataset. We visualize generated interactive humans (blue-colored humans) alongside their given partners (red-colored humans). The first three rows show generated humans by ContactGen according to different partner poses, where ContactGen shows diverse and plausible generations in terms of physical contact. For comparison with SAGA and IMM, we visualize generated humans under the same given partner conditions (last three rows). Compared to our results showing visually satisfactory outcomes, humans generated by SAGA and IMM show improper pose and penetration.
• Figure 5: Comparison of predicted contact regions. In our method, the same colored regions describe the estimated contact regions between two humans. We can observe plausible contact regions according to the interaction label. In SAGA, yellow indicates the estimated contactable vertices between humans, but they do not know the correspondences.