HumanNOVA: Photorealistic, Universal and Rapid 3D Human Avatar Modeling from a Single Image

Researchers introduced HumanNOVA, a feed-forward model for 3D human avatar generation from a single RGB image.

It uses a large synthetic and fitted human dataset, token-conditioned fusion, cross-attention, and triplane representations to improve realism and robustness.

HumanNOVA can produce photorealistic avatars in under a second without test-time optimization.

The approach could make 3D avatar creation faster and more scalable for VR, games, and digital media.