Long-horizon world generation faces a counterintuitive bottleneck: a scene can remain visually coherent only if its past is remembered, but accumulating every generated frame makes that memory increasingly expensive. The key idea here is to make memory grow with newly explored space—not with the length of the video—and to store it in the same latent representation used by the generator.
L3G-World: Real-Time Infinite World Generation with Latent 3D Gaussian Memory
Generates explorable video worlds from a single image and camera trajectory while using persistent 3D Gaussian memory to preserve spatial consistency. Storing VAE features in the memory enables direct latent-space conditioning and reported generation at 9 FPS.
Introduction
Information
- Websitearxiv.org
- OrganizationsKAIST AI, Ewha Womans University, KAIST Urban AI Institute
- AuthorsJinhyeok Choi, JoungBin Lee, Jongmin Lee, Minkyeong Jeon, Jiyoung Kim, Kaeun Han, Yoonjin Yoon, Seungryong Kim
- Published date2026/10/29
Categories
More Items
Proves that JEPAs across modalities can learn representations that differ only by a rotation when positive pairs describe the same world states. These “good views” achieve cross-modal alignment with far fewer samples than next-token prediction.
Parses documents, images, audio, and video into structured JSON grounded in space and time. A shared encoder and output schema cover ten parsing tasks; the 5B model scores 96.96 on OmniDocBench v1.6.
Decides when a diffusion vision-language model needs to process an image by checking whether text-only predictions agree with its latest image-conditioned predictions. Retains 99.6% of baseline scores across ten benchmarks at 35.3% of decoding FLOPs.