L3G-World Hits 9 FPS Real-Time World Generation on One H100

A paper posted to Hugging Face on October 30 claims to run a feed-forward world model at 9.02 frames per second on a single H100, well above the sub-two-FPS numbers the paper reports for the temporal-context baselines it tests against.

The system, called L3G-World, splits generation into two alternating modules: a memory updater that maintains a 3D Gaussian representation of the scene, and a video diffusion transformer that renders frames from it. The twist is in what each Gaussian stores. Rather than holding RGB values that must be re-encoded through a VAE for every new frame, each primitive carries a VAE latent feature directly, so the DiT reads from latent space without a re-encoding pass.

"Each Gaussian further stores a VAE latent feature, allowing the memory to be rendered directly into the DiT's latent space without VAE re-encoding," the abstract states. The authors claim "up to 38.6× fewer 3D primitives, 3.2× faster memory reconstruction, and 359× faster VAE feature rendering relative to the most efficient prior 3D-memory baseline for each metric."

Against NeoVerse, the strongest 3D-memory baseline listed, the paper reports 266 to 417 times fewer primitives, 42 to 67 times less memory use, and 349 to 359 times higher VAE-conditioning throughput. On the DL3DV benchmark L3G-World records a PSNR of 16.00, LPIPS of 0.366, and subject-consistency of 0.820. The abstract pitches the architectural reason plainly: "the memory grows with scene extent rather than stream length."

The retrieved page lists "Equal contribution" and "Co-corresponding authors" but surfaces no author names or institutional affiliation. Training ran on four A100 GPUs and inference on one H100. The paper lands amid 324 generative-AI stories we have tracked over the last ninety days.