Advances in Human Motion and Interaction Synthesis

The field of human motion and interaction synthesis is moving towards more realistic and controllable generation of human movements and interactions with objects. Researchers are exploring new methods to improve the naturalness and fidelity of generated motions, including the use of diffusion models, reinforcement learning, and novel training strategies. Noteworthy papers in this area include AlignHuman, which proposes a framework for improving motion and fidelity via timestep-segment preference optimization, and GenHOI, which introduces a two-stage framework for generalizing text-driven 4D human-object interaction synthesis to unseen objects. Other notable papers, such as SyncTalk++ and HOIDiNi, demonstrate significant advancements in synchronized talking heads synthesis and human-object interaction generation, respectively.

Sources

AlignHuman: Improving Motion and Fidelity via Timestep-Segment Preference Optimization for Audio-Driven Human Animation

Egocentric Human-Object Interaction Detection: A New Benchmark and Method

Toward Rich Video Human-Motion2D Generation

SyncTalk++: High-Fidelity and Efficient Synchronized Talking Heads Synthesis Using Gaussian Splatting

GenHOI: Generalizing Text-driven 4D Human-Object Interaction Synthesis for Unseen Objects

HOIDiNi: Human-Object Interaction through Diffusion Noise Optimization

Built with on top of