Zhejiang University & Ant Research · 3D Generation

Diffuman4D

Spatio-temporal diffusion model that generates 4D-consistent multi-view videos from sparse-view input videos, then reconstructs high-fidelity 4D Gaussian Splatting models for free-viewpoint human performance rendering. Uses skeleton-Plücker conditioning and sliding iterative denoising to synthesize novel views of humans in motion with complex clothing.