Brief

Researchers have developed a novel generative model for human geometry that significantly improves the quality and efficiency of creating realistic 3D avatars. This new approach encodes geometry distributions as 2D feature maps and utilizes SMPL models, refining the flow velocity field for better accuracy. The framework employs a two-stage training process, first compressing distributions into a latent space with a diffusion flow model and then training another flow model on this latent space. Experiments show a 57% improvement in geometry quality for tasks like pose-conditioned avatar generation and novel pose synthesis. AI