UniVerse checkpoints

Weights of the video diffusion model of UniVerse: Unleashing the Scene Prior of Video Diffusion Models for Robust Radiance Field Reconstruction (ICCV 2025).

Code | Paper | Project Page

File	Model	Resolution	Training	sha256
universe_512.ckpt	UniVerse-512	320 x 512	stage 1, 14,520 iterations	7c86adcd48bb32054d62e4bb47831014b5e9c481f2cea028dfb429e08547bdea

The file is a weights-only PyTorch Lightning checkpoint (fp32, 10.4 GB) that contains the whole model, including the VAE and the OpenCLIP ViT-H/14 encoders.

Usage

In the GitHub repository, inference.py downloads the checkpoint into checkpoints/ on first use:

python inference.py --image_dir data/demo/images --out_dir output/demo

or download it manually:

hf download TmaKiss/UniVerse universe_512.ckpt --local-dir checkpoints

Terms

The models are fine-tuned from ViewCrafter's ViewCrafter_25_sparse (Apache-2.0), contain the OpenCLIP ViT-H/14 weights trained on LAION-2B (MIT), and were trained on DL3DV-10K, which is released for non-commercial use (CC BY-NC 4.0). The UniVerse release is distributed under the ZJU3DV Project Registration License (PRL) v1.0, see the repository's LICENSE.

Citation

@misc{cao2025universeunleashingsceneprior,
  title={UniVerse: Unleashing the Scene Prior of Video Diffusion Models for Robust Radiance Field Reconstruction},
  author={Jin Cao and Hongrui Wu and Ziyong Feng and Hujun Bao and Xiaowei Zhou and Sida Peng},
  year={2025},
  eprint={2510.01669},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2510.01669},
}