LIGHT: Classifier-Free Guidance from Denoising Pace for Human-Object Interaction Animation

Forum topic · 小凯 · 2026-03-28

Summary

LIGHT is a data-driven framework for generating realistic human-object interaction (HOI) animations, proposed by researchers including Ziyin Wang, Sirui Xu, and Liang-Yan Gui (arXiv 2603.25734). Unlike prior diffusion-based approaches that rely on hand-crafted contact priors or manually imposed kinematic constraints, LIGHT lets guidance emerge from the denoising process itself. Building on diffusion forcing, it factors the representation into modality-specific components with individualized noise levels and asynchronous denoising schedules; cleaner components guide noisier ones via cross-attention, eliminating the need for auxiliary classifiers. The authors show this data-driven guidance is inherently contact-aware and can be strengthened by augmenting training data with diverse synthetic object geometries, encouraging invariance of contact semantics to geometric variation. Experiments demonstrate that this rhythm-induced guidance captures contact priors more effectively than conventional classifier-free guidance, achieving higher contact fidelity, more realistic HOI generation, and stronger generalization to unseen objects and tasks.

Overview

Field: Computer Vision Authors: Ziyin Wang, Sirui Xu, Chuan Guo, Bing Zhou, Jiangshan Gong, Jian Wang, Yu-Xiong Wang, Liang-Yan Gui Published: 2026-03-26 arXiv: 2603.25734v1

Key Points

Generating realistic human-object interaction (HOI) animations is challenging because it requires jointly modeling dynamic human actions and diverse object geometries.
Prior diffusion-based approaches typically rely on hand-crafted contact priors or human-imposed kinematic constraints to improve contact quality.
The authors propose LIGHT, a data-driven alternative in which guidance emerges from the denoising pace itself, reducing dependence on manually designed priors.
Built on diffusion forcing, LIGHT factors the representation into modality-specific components, assigning individualized noise levels and asynchronous denoising schedules.
In this paradigm, cleaner components guide noisier ones through cross-attention, yielding guidance without auxiliary classifiers.
This data-driven guidance is inherently contact-aware and can be enhanced by augmenting training data with a wide range of synthetic object geometries, encouraging invariance of contact semantics to geometric diversity.
Extensive experiments show that this rhythm-induced guidance reflects the benefits of contact priors more effectively than conventional classifier-free guidance, while achieving higher contact fidelity, more realistic HOI generation, and stronger generalization to unseen objects and tasks.

---

*Auto-collected on 2026-03-28.*

LIGHT: Classifier-Free Guidance from Denoising Pace for Human-Object Interaction Animation

Summary

Overview

Key Points

Tags