Fast 4D Mesh Generation by Spatio-Temporal Attention Chains

Researchers proposed a training-free 4D mesh generation method that uses spatio-temporal attention chains to map anchor-mesh vertices into latent space and propagate correspondences over time.

The framework reconstructs frame-specific vertices with attention, reducing generation time to about 9 seconds while improving mesh quality and temporal consistency.

It scales to longer video sequences and supports downstream tasks such as 2D object tracking, 4D tracking, and camera estimation.

Overall, the method makes dynamic 3D reconstruction faster, more scalable, and more useful for practical applications.