KAIST researchers propose L3G-World, an interactive world generator that uses a compact 3D Gaussian memory to run at 9 FPS.
Memory: The Latent 3D Gaussian Memory grows with scene extent rather than stream length, and each Gaussian stores a VAE latent feature.
Generation: A memory updater and a video diffusion transformer alternate autoregressively. The memory renders straight into the DiT's latent space, so the RGB renderings don't need VAE re-encoding.
Numbers: Against the most efficient prior 3D-memory baseline for each metric, the paper reports up to 38.6x fewer 3D primitives, 3.2x faster memory updates and 359x faster memory rendering.
Sign in to comment.
TL;DR of the paper, generated automatically.
KAIST researchers propose L3G-World, an interactive world generator that uses a compact 3D Gaussian memory to run at 9 FPS.