KAIST's L3G-World Generates Infinite Interactive Worlds in Real Time with Latent 3D Gaussian Memory
kaist-ai · hf · 2026-10-30
Researchers at KAIST AI and Ewha Womans University present L3G-World, which pairs a compact latent 3D Gaussian memory with a video diffusion transformer via alternating autoregressive updates to produce spatially consistent, interactive worlds in real time. Compared to prior approaches relying on direct diffusion rendering, it renders novel views faster, keeps memory flat as scenes grow, and supports infinite generation from a single input image plus a text prompt and camera trajectory.
More from Embodied
- Amazon's DreamFAR world model nearly doubles π₀ real-world performance for robot training — CyberRobooo · 2026-10-30
- Tesla Robotaxi hosts Accessibility Day, disabled user feedback to shape Cybercab — XFreeze · 2026-10-30
- Against All Odds: Maker Gets His Robot Live on Perth Radio After Four Disasters — JohnVial · 2026-10-30
- Chinese startup sells $1,000 AI laser device that shoots mosquitoes mid-flight — nordicinst · 2026-10-30
- NVIDIA open-sources Metatron v1, a unified humanoid control policy for Unitree G1 — Balance- · 2026-10-30
- Sim-trained quadruped HOUND lands Science Robotics cover with one policy — Scobleizer · 2026-10-30