KAIST AI, Ewha Womans University and KAIST Urban AI Institute (Jinhyeok Choi, JoungBin Lee et al.) · Latent Gaussian world model

L3G-World

L3G-World is a world-generation method from KAIST AI and collaborators that pairs a Latent 3D Gaussian Memory (L3G-Mem) with a video diffusion transformer (L3G-Gen). It aims to keep generated worlds spatially consistent over long horizons in real time. The memory adds Gaussians only for newly observed regions and refines existing ones on revisits, so it grows with scene extent rather than stream length. Each Gaussian stores a VAE latent, so the memory renders straight into the diffusion latent space without RGB rendering and VAE re-encoding. The authors report up to 38.6× fewer 3D primitives, 3.2× faster memory updates and 359× faster memory rendering than the most efficient prior 3D-memory baseline for each metric, with real-time generation at 9 FPS. L3G-Mem reconstructs video streams at 47 FPS on a single A100, and in an ablation direct latent rendering reaches 943.6 FPS against 2.97 FPS without VAE lifting. It works from a single image or a few views, and the page links the paper (arXiv 2610.40614), code and models; the licence is not stated.