The Bolt3D method allows for feed-forward 3D scene generation in just 6.25 seconds on a single GPU. By generating multi-view Splatter Images from input images through a diffusion model and Gaussian regression, high-quality 3D scenes are created. The key to success lies in the Geometry VAE, which efficiently compresses pointmaps. When compared to optimization-based methods, Bolt3D offers generative capabilities and reduces inference cost significantly. Unique to this approach is the interactive viewer, allowing for real-time rendering of 3D scenes in a web browser. The method is acknowledged in the research community for its contributions and insights, making it a groundbreaking achievement in 3D scene generation.
https://szymanowiczs.github.io/bolt3d