KAIST CV Lab introduces L3G-World — real-time generation of infinite 3D worlds

🧠 KAIST unveils L3G-World: infinite 3D worlds in real time

The framework from KAIST CV Lab generates a continuous 3D world from a single image, text, and camera trajectory. The core is latent 3D Gaussian memory (L3G-Mem): primitives are rendered directly into the latent space of the L3G-Gen DiT, bypassing RGB rendering and VAE re-encoding, while memory grows with scene size rather than stream length.

🌍 Direct latent rendering eliminates the RGB→VAE round-trip that slowed down world models with explicit 3D memory. On an A100, memory updates at 47 FPS, and the world is generated at 9 FPS; an economical 3D-memory baseline has up to 38.6x more primitives.

👤 One of the first real-time generators of "infinite worlds" that maintains consistency when the camera returns to a previous position. Code and weights are not yet available: release is promised after internal review, follow github.com/cvlab-kaist/L3G-World.

Source 1: https://arxiv.org/abs/2610.40614 Source 2: https://cvlab-kaist.github.io/L3G-World/