GADA: Geometry-Aware Deformable Aggregation for Image-Based Gaussian Splatting

Siwoo Lim¹ · Sunjae Yoon² · Gwanhyeong Koo¹ · Chang D. Yoo¹

¹ Korea Advanced Institute of Science and Technology (KAIST)  ·  ² Chung-Ang University

We plan to release the full training and evaluation code for GADA in the near future. This repository currently serves as a placeholder; the official implementation, configuration files, and reproduction scripts will be uploaded here once the cleanup is complete.

In the meantime, you can already access:

• 📄 The paper on OpenReview
• 🌐 The project page with qualitative comparisons
• 🖼️ The pretrained rendering results used in our paper, available via the Google Drive link below

Please ⭐ star or watch this repository to get notified when the code is released.

We provide the rendering results used in our paper (Mip-NeRF 360, Tanks & Temples, Deep Blending, and Shiny scenes) for direct qualitative comparison.

🔗 Download: Google Drive — GADA Pretrained Images

The drive contains, for each scene, the rendered outputs of:

• Ground Truth
• GADA (Ours)

These are the exact images used to produce the qualitative comparisons in Figure 6 and the project page slider.

Coming soon — installation instructions will be added together with the code release.

We use the standard benchmark datasets following 3DGS:

• Mip-NeRF 360
• Tanks and Temples
• Deep Blending
• Shiny — preprocessed via COLMAP, following the protocol used by IBGS.

Coming soon — training and evaluation scripts will be released alongside the code.

Please consider citing our work if you find it useful for your research:

@misc{lim2026gadageometryawaredeformableaggregation,
      title={GADA: Geometry-Aware Deformable Aggregation for Image-Based Gaussian Splatting}, 
      author={Siwoo Lim and Sunjae Yoon and Gwanhyeong Koo and Chang D. Yoo},
      year={2026},
      eprint={2607.00595},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.00595}, 
}

We build our method upon the codebase of IBGS, 3DGS, and PGSR. We sincerely thank the authors for releasing their excellent code.