Towards Alias-Free 4D Gaussian Representations with Motion-Aware Filtering

Ankit Dhiman^1,2  ·  Kunal A Kathare¹  ·  Pranav Vignesh¹  ·  Lokesh R Boregowda²  ·  R. Venkatesh Babu¹

¹Indian Institute of Science, Bangalore    ²Samsung R&D Institute India – Bangalore

---

Overview

4D Gaussian representations alias badly when novel views are rendered at a scale other than the one they were trained on — most visibly under zoom-in and zoom-out. Static anti-aliasing filters such as the 3D smoothing filter of Mip-Splatting do not fix this, because they assume a constant sampling rate per primitive and therefore ignore local motion.

MAAF replaces that constant sampling rate $\hat{\nu}_k$ with a time-varying one $\hat{\nu}_k(t)$, estimated non-parametrically. For every Gaussian we build a joint density over time and the focal-to-depth ratio using a Parzen-window estimator with a Weibull kernel, then sample from it at render time to pick the filter width. Because the filter acts on the projected 3D Gaussians and touches nothing else, it drops into different 4D backbones unchanged.

Backbones

The filter is representation-agnostic. This repository hosts one self-contained release per backbone; pick the one matching the results you want to reproduce.

Backbone	Directory	Datasets	Status
SpeeDe3DGS	SpeeDe3DGS/	D-NeRF (synthetic)	✅ Released
SaRO-GS	SaRO-GS/	Plenoptic Video, D-NeRF, HyperNeRF	🚧 Coming soon

SaRO-GS is the primary backbone — the main tables of the paper (Plenoptic Video, D-NeRF and HyperNeRF) are produced with it. We further show experiments on SpeeDe3DGS which shows the same filter transfers unchanged to a different 4D representation and mitigates the alisasing artefacts.

Each directory is a standalone codebase with its own environment.

Citation

If you find this work useful, please cite:

@misc{dhiman2026aliasfree4dgaussianrepresentations,
      title={Towards Alias-Free 4D Gaussian Representations with Motion-Aware Filtering}, 
      author={Ankit Dhiman and Kunal A Kathare and Pranav Vignesh and Lokesh R Boregowda and Venkatesh Babu Radhakrishnan},
      year={2026},
      eprint={2608.21828},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.21828}, 
}

Please also cite the backbone you build on — SpeeDe3DGS, SaRO-GS, Deformable-3DGS, Mip-Splatting and 3D Gaussian Splatting.

License

The code original to this project is released under the MIT License.

Acknowledgements

Built on top of SpeeDe3DGS, SaRO-GS and Deformable-3D-Gaussians. The 2D screen-space filter follows Mip-Splatting. We thank the authors for releasing their code.