SJTU's 4LC Cache Cuts Video-World Conditioning 68.5× vs NeoVerse

Caching 4D latents rather than re-rendering them for every request collapses the conditioning cost of dynamic-world video generation, according to a new preprint from Shanghai Jiaotong University, Zhongguancun Academy, the Institute of Automation of the Chinese Academy of Sciences and Harbin Institute of Technology. The authors report that 4LC readout is "68.5× faster, with 12.6× lower peak GPU memory" than the 4DGS rendering plus VAE encoding pipeline in NeoVerse, the baseline they benchmark against.

"Condition acquisition accounts for 14.9% of the 18.47 s end-to-end generation time in NeoVerse, compared with 0.25% of 16.19 s for 4LC-Video," the paper reports. The cache binds representations from multiple encoders, listed in the abstract as "VAE, DINO, V-JEPA, VGGT", to shared geometry and motion, so a single stored object can serve requests at independently specified times and viewpoints.

4LC-Video is pitched as one model handling "camera control, temporal control, novel-view synthesis, scene revisits, and long-term generation" without task-specific fine-tuning. The paper landed the same day our tracker logged Tencent Hunyuan's WorldRecon, part of a steady flow of Chinese-lab work on 4D scene reconstruction and video conditioning. The abstract publishes no fidelity numbers against the NeoVerse baseline, and the headline speed figures are measured against that single comparator.