L3G-World generates 3D worlds at 9.02 FPS on one H100

L3G-World generates an explorable, spatially consistent world at 9.02 frames per second on a single H100 GPU. The system is described in a paper, "L3G-World: Real-Time Infinite World Generation with Latent 3D Gaussian Memory," submitted to arXiv on October 29 by researchers at KAIST AI, Ewha Womans University and the KAIST Urban AI Institute.

This is a research result, not a product. There is no site to log into, no API and no waitlist. What the paper offers a creator is a method.

L3G-World pairs two modules that take turns. A memory updater, L3G-Mem, keeps a "latent 3D Gaussian memory": a compact set of 3D Gaussians that stands in for the parts of the scene the camera has already seen. A video diffusion transformer, L3G-Gen, generates each new chunk of video conditioned on that memory. Every Gaussian stores a VAE latent feature, so the memory renders straight into the generator's latent space instead of being turned back into pixels and encoded again. The memory also grows with the extent of the scene rather than the length of the video, so a long walk costs about the memory of a short one.

How L3G-World compares on one H100

The paper reports efficiency against other world models, measured over 400 frames of generation on a single H100 (Table 1):

Against NeoVerse, a Gaussian-based world model, L3G-World uses 266 to 417 times fewer 3D primitives and 42 to 67 times less memory, and renders the VAE-space conditioning that feeds its generator 349 to 359 times faster. Against Lyra 2.0, an efficient point-cloud memory, it reconstructs 8.9 to 9.5 times faster while using 7.8 to 10.4 times less memory. The gains come from adding Gaussians only where the camera finds new geometry, rather than adding new primitives every frame.

On the DL3DV benchmark, L3G-World posted the best PSNR (16.00) and LPIPS (0.366) on loop-closure trajectories, and the best FID (40.57), subject consistency (0.820) and background consistency (0.893) among the models tested. A faster variant, L3G-World Fast, gave up some fidelity for the best imaging and aesthetic scores.

Two better-known systems sit further from reach. Google DeepMind's Genie 4 "is available as a research preview for trusted testers," with wider access promised to Google AI Ultra subscribers. Tencent's Hy-Nomas, a technical report posted October 27, targets minute-scale rollouts on a single RTX 4090.

Release is the catch. The team's GitHub repository is up under an MIT license but currently holds only a README, a license file and a teaser image. Its release plan lists the L3G-Mem code and checkpoints, the L3G-Gen code and checkpoints, the inference code and the fast checkpoints as unchecked. When the weights do arrive, the plan splits them. The L3G-Mem weights are set for a non-commercial CC BY-NC 4.0 license, and the generator weights fall under NVIDIA's Open Model License, because the video model is built on NVIDIA Cosmos-Predict2.5.

The metric to watch is whether those checkpoints land.

Read the paper and the sample videos on the project page; no model weights are downloadable as of October 30.