The Pulse
4LC Cuts World-Model Conditioning Time by 68.5×
A research team introduced 4LC, a 4D latent cache that reuses visual features to condition video world models without rendering and encoding each new request. The paper reports up to 68.5× faster condition acquisition than a rendering-and-e
AI.info Team · Written with AI ·
A 4D latent cache cut the time spent preparing visual conditions for video world models by as much as 68.5 times in tests reported in a paper posted to arXiv on October 13. The system, called 4LC, also used up to 12.6 times less peak GPU memory than a comparison pipeline that renders a scene and then encodes it. The gains apply to condition acquisition—the preparation of scene information for a generation request—not to the full video-generation process.
4LC reuses features instead of rebuilding each view
World models often need to generate a scene from different camera positions or at different points in time. Existing pipelines can render a fresh RGB view from a reconstructed scene and encode that image for every request. 4LC instead stores visual features alongside shared scene geometry and motion, then reads the cached information at the requested time and viewpoint.
The cache combines representations from visual encoders that capture different information, including appearance, semantics, geometry and motion. Its readout can also combine observations to fill regions not covered by one cache entry. “Once observations are cached, subsequent requests read the stored representations directly instead of rendering and re-encoding them,” write Jianhui Zhang, Zuyao Lin, Pengxin Xu, Jingdong Wang and Xingyu Chen in the paper.
The 68.5× result measures scene preparation
The researchers compare 4LC’s readout with rendering and VAE encoding in NeoVerse, a system that builds conditions from a 4D Gaussian-splatting scene. In that comparison, 4LC’s condition-acquisition step is 68.5 times faster and uses 12.6 times less peak GPU memory. The paper’s shared test setup uses the Wan2.1 model and VAE encoder for requests of 81 frames at 560 × 336 pixels.
The authors also compare end-to-end generation time under the same settings. Condition acquisition accounts for 14.9% of NeoVerse’s 18.47-second generation time, versus 0.25% of 4LC-Video’s 16.19 seconds. Those figures show that 4LC sharply reduces the conditioning share of the workload; they do not mean the complete video is generated 68.5 times faster.
One model handles five generation tasks
Built on the cache, 4LC-Video uses one model for camera control, temporal control, novel-view synthesis, scene revisits and long-term generation, without task-specific fine-tuning. The paper evaluates a single checkpoint across those tasks and reports that it maintains scene consistency while following requested camera and time changes. For long-term generation, each new segment adds observations to the cache for later use.
The results remain the authors’ reported benchmark findings, not evidence that the system will deliver the same speedups in every deployment. 4LC depends on a feed-forward 4D reconstruction model to estimate depth, camera parameters and motion, and errors in those estimates can affect its readouts. The long-term setup also reuses only static content from earlier segments, leaving reuse of dynamic content across segments for future work.
Dynamic history remains an open test
That limitation matters for scenes whose important changes include moving people, objects or other activity: preserving the layout is not the same as carrying forward a scene’s evolving motion. The paper establishes a method for reading cached features at new times and viewpoints, but its own long-term experiments restrict historical reuse to static units. Whether 4LC can extend its memory to dynamic history without losing consistency remains unanswered.