SASNet: Spatially-Adaptive Sinusoidal Networks for INRs

Official implementation of SASNet: Spatially-Adaptive Sinusoidal Networks for INRs, CVPR 2026.

Haoan Feng¹, Diana Aldana², Tiago Novello², Leila De Floriani^{1 1}University of Maryland, College Park    ²Institute for Pure and Applied Mathematics (IMPA)

• arXiv: https://arxiv.org/abs/2503.09750
• Project page: https://fengyee.github.io/SASNet_inr/
• Code: https://github.com/Fengyee/SASNet_inr

TL;DR

SASNet pairs a frozen frequency embedding layer with a lightweight hash-grid MLP that produces spatially-adaptive masks. The masks select, per location and per frequency band, which sinusoidal neurons contribute to the output. This suppresses high-frequency leakage in smooth regions while preserving fine detail around edges and surfaces, yielding higher accuracy and faster convergence on image fitting, volumetric reconstruction, and SDF tasks at only modest overhead from the mask generation network.

Install

conda create -n sasnet python=3.10
conda activate sasnet
pip install -e .

The hash-grid mask generator is required whenever you construct SASNet(use_masks=True) (the full SASNet model used in all paper experiments). It is built on kaolin-wisp; install it from its upstream repository before enabling masks. The unmasked SIREN-style baseline (SASNet(use_masks=False)) runs without kaolin-wisp.

Training scripts and configuration files are not included by now; please refer to the paper and project page for model configurations and hyperparameters.

Citation

@inproceedings{feng2026sasnet,
  title     = {SASNet: Spatially-Adaptive Sinusoidal Networks for INRs},
  author    = {Feng, Haoan and Aldana, Diana and Novello, Tiago and De Floriani, Leila},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}

License

MIT (see LICENSE).