Alex Yu*, Sara Fridovich-Keil*, Matthew Tancik, Qinhong Chen, Benjamin Recht, Angjoo Kanazawa

UC Berkeley

Website and video: https://alexyu.net/plenoxels

arXiv: https://arxiv.org/abs/2112.05131

Note: This JAX implementation is intended to be high-level and user-serviceable, but is much slower (roughly 1 hour per epoch) than the CUDA implementation https://github.com/sxyu/svox2 (roughly 1 minute per epoch), and there is not perfect feature alignment between the two versions. This JAX version can likely be sped up significantly, and we may push performance improvements and extra features in the future. Currently, this version only supports bounded scenes and trains using SGD without regularization.

@inproceedings{yu_and_fridovichkeil2021plenoxels,
      title={Plenoxels: Radiance Fields without Neural Networks},
      author={{Sara Fridovich-Keil and Alex Yu} and Matthew Tancik and Qinhong Chen and Benjamin Recht and Angjoo Kanazawa},
      year={2022},
      booktitle={CVPR},
}

Note: joint first-authorship is not really supported in BibTex; you may need to modify the above if not using CVPR's format.

We recommend setup with a conda environment, using the packages provided in requirements.txt.

Currently, this implementation only supports NeRF-Blender, which is available at:

https://drive.google.com/drive/folders/128yBriW1IG_3NJ5Rp7APSTZsJqdJdfc1

The training file is plenoptimize.py; its flags specify many options to control the optimization (scene, resolution, training duration, when to prune and subdivide voxels, where the training data is, where to save rendered images and model checkpoints, etc.). You can also set the frequency of evaluation, which will compute the validation PSNR and render validation images (comparing the reconstruction to the ground truth).