Overview

Code implementation of Spatial Annealing for Efficient Few-shot Neural Rendering (AAAI 2025)
Comparisons:

This code is implemented based on TriMipRF.

Installation

First, create a new sanerf environment:

conda create -n sanerf python==3.8

Next, activate the environment:

conda activate sanerf

Install the following dependency:
PyTorch (1.13.1 + CUDA 11.6)
nvdiffrast
tiny-cuda-nn

pip install -r requirements.txt

DATA

nerf_synthetic dataset

Please download the nerf_synthetic dataset from the NeRF official project.

Reproduce SANeRF's results

bash ./scripts/reproduce_sanerf.sh

The reproduced results have an error margin of 0.05dB.

Reproduce baseline's results

Replace method=sanerf with method=trimiprf in "./scripts/reproduce_sanerf.sh" , then run:

bash ./scripts/reproduce_sanerf.sh

Citation

If you find our work useful, please cite it as

@inproceedings{xiao2025spatial,
  title={Spatial Annealing for Efficient Few-shot Neural Rendering},
  author={Xiao, Yuru and Zhai, Deming and Zhao, Wenbo and Jiang, Kui and Jiang, Junjun and Liu, Xianming},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={8},
  pages={8691--8699},
  year={2025}
}