TL;DR: FlashWorld enables fast (7 seconds on a 1x A100/A800 GPU, 4 seconds on 1x H100/H800 GPU) and high-quality 3D scene generation across diverse scenes, from a single image or text prompt.

A $\text{A}^2\text{TG}$ scene is trained in a two steps manner, first train a 2D Gaussian Splatting model for 30000 iterations, than train $\text{A}^2\text{TG}$ with the same number of Gaussians for ...