ConeGaussian

TL;DR. A camera is a sampling operator that maps each finite pixel to a bundle of rays. ConeGaussian builds an anisotropic pixel footprint directly from the neighbouring rays of any calibrated central camera's inverse projection, filters the exact ray–Gaussian response with it in closed form, and derives a per-Gaussian training-frequency floor from the same geometry. One implementation gives anti-aliased zoom-out and artifact-free zoom-in for pinhole and strongly distorted fisheye cameras, and ports unchanged across two Gaussian ray-rendering backbones (3DGEER and 3DGUT).

Method

Build the footprint yourself interactive

Move the pixel across the sensor, switch the camera model or reshape the Gaussian. The 3D geometry and both rows of filters update together.

Zoom-out: minification without aliasing

3DGEERConeGaussian

Mip-NeRF 360 · bicyclepinhole · 19.1 → 22.7 dB (+3.7)

3DGEERConeGaussian

Mip-NeRF 360 · gardenpinhole · 21.3 → 27.0 dB (+5.7)

3DGEERConeGaussian

Zip-NeRF londonfisheye · 28.7 → 30.6 dB (+1.9)

Zoom-in: magnification without needle artifacts

3DGEERConeGaussian

bicyclegrass tuft at the front wheel

3DGEERConeGaussian

gardenhedge foliage

3DGEERConeGaussian

stumptall grass blades

One scene, any central camera

• garden
• bicycle

Qualitative comparisons against ground truth

Choose Zoom-out or Zoom-in, then a scene, to compare 3DGEER → ConeGaussian → ground truth. The 6 zoom-out and 5 zoom-in comparisons are kept separate, with their original detail insets and scale labels. Scene/view names are descriptive because the source figures do not list scene IDs.

• Zoom-out
• Zoom-in

Zoom-out · 1/8 sampling rate. Select a scene below.

• Office · windows
• Garden
• Flowers
• Room
• Bicycle
• Backyard

Window sill Zoom-out · 1/8 sampling rate

Supplementary zoomout figure, row 6. Full source figure ↗

Zoom-in · ×8. Select a scene below.

• Office · windows
• Meeting room
• Office · ceiling
• Office · desk

Window frame Zoom-in · ×8

Supplementary zoomin figure, row 5. Full source figure ↗

Comparisons with other methods

Every cell reads PSNR↑ / SSIM↑ / LPIPS↓. Real-scene benchmarks: ScanNet++ and Zip-NeRF (fisheye, metrics inside a 3-pixel-eroded valid-domain mask) and Mip-NeRF 360 (pinhole); every eighth image is held out. Zoom-in / zoom-out change only the image sampling resolution with consistently scaled calibration, never the optical zoom or the pose. Green = best, orange = second best per column and metric among the generalizable ray-based methods.

Multi-scale training and testing, two backbones

Multi-scale training and testing at {1, ½, ¼, ⅛}. PSNR↑ / SSIM↑ / LPIPS↓. Gray rows are author-reported, pinhole-only screen-space methods (3DGS rasterizer) that do not support generic cameras; green / orange mark the best / second-best generalizable ray-based method per column and metric.

Citation

If you use ConeGaussian in your work, please cite our paper. The BibTeX below follows the public arXiv record.

@article{zhang2026conegaussian,
  title   = {ConeGaussian: Anti-Aliased Gaussian Ray-Tracing for Generic Central Cameras},
  author  = {Zhang, Deheng and Shi, Letian and Yang, Runyi and Li, Zhendong and Sun, Lei and
             Wu, Kanzhi and Chhatkuli, Ajad and Paudel, Danda Pani and Van Gool, Luc},
  journal = {arXiv preprint arXiv:2609.13397},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.13397}
}

References

1. Deheng Zhang et al. ConeGaussian: Anti-Aliased Gaussian Ray-Tracing for Generic Central Cameras. arXiv, 2026.
2. Zixun Huang, Cho-Ying Wu, Yuliang Guo, Xinyu Huang, and Liu Ren. 3DGEER: 3D Gaussian Rendering Made Exact and Efficient for Generic Cameras. ICLR, 2026. Code.
3. Qi Wu, Janick Martinez Esturo, Ashkan Mirzaei, Nicolas Moënne-Loccoz, and Zan Gojcic. 3DGUT: Enabling Distorted Cameras and Secondary Rays in Gaussian Splatting. CVPR, 2025.
4. Samuel Rota Bulò, Nemanja Bartolovic, Lorenzo Porzi, and Peter Kontschieder. Hardware-Rasterized Ray-Based Gaussian Splatting (VKRayGS). CVPR, 2025.
5. Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-Splatting: Alias-free 3D Gaussian Splatting. CVPR, 2024.