jianhuizhang/4LC-Video

Video generation conditioned on a 4D latent cache.

Models

Profile	Cached representations	Weights
flux-all (default)	RGB / depth / DINOv2 / V-JEPA 2.1 / WorldMirror	Download
flux-vjepa	RGB / depth / V-JEPA 2.1	Download

Each checkpoint contains the condition adapter and VACE control branch for Wan2.1-T2V-14B. Model configurations are provided in the code repository.

Download and Use

Complete the code installation and base-model and encoder setup, then run the commands below from the 4LC code repository root.

Download a profile

hf download jianhuizhang/4LC-Video 4lc-video-flux-all.safetensors \
  --local-dir local-assets/models

Load the weights for video-conditioned generation

source scripts/env.sh
python scripts/infer.py --example examples/spatial-control \
  --output outputs/control.mp4

The bundled example selects flux-all through example.yaml. To use another profile, pass its matching --model and --weights.

Image generation, camera control and rollout · Training

Citation

If you find our models, code, or paper useful in your research, please cite our work:

@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}
}