4LC: 4D Latent Cache

TL;DR4LC caches visual representations of static and dynamic content for reuse across times and viewpoints.Built on it, a single 4LC-Video model enables efficient conditioning for multi-task video generation.

From Cache Readout to Video

Camera Control

Time Control

Single View to Multiple Views

Image to Video

Image to World

Video Editing

Condition Acquisition

One segment: share of end-to-end generation time spent on condition acquisition

NeoVerse

2.75 s of 18.47 s

4LC

0.04 s of 16.19 s

Condition acquisitionRest of generation

up to68.5×faster condition acquisition

12.6×lower peak GPU memory

Long-horizon: condition acquisition adds up over 8 segments

NeoVerse

4LC

All timings: 81-frame segments at 560×336 on the same Wan2.1 backbone.

Comparison with NeoVerse

Abstract

For video generation in dynamic world modeling, observed static and dynamic content should be efficiently reused to provide conditions for repeated generation at different times and viewpoints. We introduce 4LC, a 4D latent cache with composite visual representations and a flexible spatiotemporal access interface. As a latent cache, 4LC directly reuses visual representations of static and dynamic content to provide generation conditions, avoiding per-request RGB rendering and re-encoding. Building on 4LC, a single 4LC-Video model handles camera control, temporal control, novel-view synthesis, scene revisits and long-term generation without task-specific fine-tuning.

Overview of 4LC: the cache, cache-level generation, condition-acquisition cost and multi-task results — **Overview of 4LC.** **(a)** Static and dynamic scene content is retained in multiple visual representations for reuse across time and viewpoint requests. **(b)** Non-cache pipelines render and re-encode a reconstructed scene for each request; 4LC encodes source observations once and reads out the cache for later requests. **(c)** With the same Wan2.1 VAE, 4LC needs less time and peak GPU memory to acquire conditions than 4DGS and RGB point clouds. **(d)** A single 4LC-Video model performs well across five video generation tasks.

4D Latent Cache

Architecture of 4LC: construction, readout and composition — **Architecture of 4LC.** **(a) Construction** binds visual representations from multiple encoders to shared geometry and motion in one cache entry. **(b) Readout** propagates the entry to a requested time and projects it through a target camera, yielding readout features and coverage masks. **(c) Composition** fills uncovered regions with readouts from further entries.

4LC-Video

Task configurations for 4LC-Video. Requests use one entry $\mathcal C_q$ or selected entries from $\mathcal C$ . $\bar t_j$ preserves the source or default single-image timeline, and $\bar\Gamma_j$ the source camera trajectory. Scene revisits use one readout along a camera path that returns to its start.

Task	Initial frames	Target request $(t_j,\Gamma_j)$	Condition source
Camera control	$\{\mathbf I_k\}_{k=0}^{K-1}$	$(\bar t_j,\Gamma_j)$	$\mathcal C_q$
Temporal control	$\{\mathbf I_k\}_{k=0}^{K-1}$	$(t_j,\bar\Gamma_j)$	$\mathcal C_q$
Novel-view synthesis	$\mathbf I_0$	$(\bar t_j,\Gamma_j)$	$\mathcal C_q$
Scene revisits	$\mathbf I_0$	$(\bar t_j,\Gamma_j)$	$\mathcal C_q$
Long-term generation	$\mathbf I_0$	$(\bar t_j,\Gamma_j)$	$\mathcal C$

**4LC-Video framework.** The condition adapter aligns readouts to the Wan2.1 VAE latent grid and fuses them with masks and camera rays for the control branch. Generated observations extend the cache.

BibTeX

@article{zhang2026fourlc,
  title   = {4LC: 4D Latent Cache for Efficient Conditioning in Dynamic World Modeling},
  author  = {Zhang, Jianhui and Lin, Zuyao and Xu, Pengxin and Wang, Jingdong and Chen, Xingyu},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026}
}