FL4N: Feed-forward Latent Model for 4D Novel View Synthesis

[arXiv] | [Project Page] | [Code]

FL4N is a feed-forward latent 4D novel view synthesis model that renders dynamic scenes from unseen viewpoints and timestamps without explicit 3D representations. It models temporal dynamics as residuals on top of a pretrained static latent prior, enabling effective 4D scene modeling with limited dynamic training data.

Quick Start

git clone https://github.com/cvlab-kaist/FL4N.git
cd FL4N
conda create -n fl4n python=3.11 -y && conda activate fl4n
pip install torch torchvision && pip install -r requirements.txt

# Download the checkpoint
python -c "from huggingface_hub import hf_hub_download; hf_hub_download('eunbeen-hong/FL4N', 'fl4n.ckpt', local_dir='checkpoints')"

# Render one NVIDIA Dynamic Scenes scene (data under $FL4N_DATA/NVIDIA)
FL4N_DATA=/path/to/data python benchmark/run_eval.py --models fl4n --protocols ctvs --datasets nvidia --scenes Balloon1 --gpus 0

Citation

@misc{hong2026fl4n,
  title         = {{FL4N}: Feed-forward Latent Model for 4D Novel View Synthesis},
  author        = {Hong, Eunbeen and Yu, Hyeonseo and An, Honggyu and Son, Jooho and Ko, Hyunah and Son, Soowon and Kim, Chaehyun and Koo, Jahyeok and Kim, Jisoo and Hur, Junhwa and Kim, Seungryong},
  year          = {2026},
  eprint        = {2610.35854},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
}

Acknowledgements

Built upon LagerNVS, and VGGT.