4LC: 4D Latent Cache for Efficient Conditioning in Dynamic World Modeling
4LC caches reusable latent representations so video models can condition on time and viewpoint without rendering and re-encoding each request.
The paper proposes a 4D latent cache that stores static and dynamic visual content tied to shared geometry and motion. It can combine representations from encoders such as VAE, DINO, V-JEPA, and VGGT. Its access interface lets a model read cached content at specified times and viewpoints, then merge observations into a common conditioning format. The authors build 4LC-Video and report more efficient condition acquisition, spatiotemporal control, and scene consistency across tasks without task-specific fine-tuning. HF Daily Papers' note
The paper proposes a 4D latent cache that stores static and dynamic visual content tied to shared geometry and motion. It can combine representations from encoders such as VAE, DINO, V-JEPA, and VGGT. Its access interface lets a model read cached content at specified times and viewpoints, then merge observations into a common conditioning format. The authors build 4LC-Video and report more efficient condition acquisition, spatiotemporal control, and scene consistency across tasks without task-specific fine-tuning. HF Daily Papers' note
score 4