Skip to content

Download Datasets

This page lists datasets compatible with the application, organized by rendering pipeline. Each section includes download links, licensing information, and recommended visualization settings.

Important

Visualization of 3D Gaussian models is most effective when using the same rendering algorithm and settings used during reconstruction. For example, models generated using Postshot 3DGS with anti-aliasing enabled will be best visualized using one of the available 3DGS pipelines (Mesh or Vertex) and by enabling "Mip Splatting Anti-Aliasing" in the rendering > rasterization parameters. Since no generic format with proper metadata exists, users must manually configure the rendering settings. Note that original 3DGS models from INRIA do not include anti-aliasing, which was introduced later [Yu2023].

3DGS Datasets

Models from INRIA original 3DGS paper

The INRIA dataset of pre-trained models is available for download from the INRIA server.

  • To visualize 3DGS reconstruction output, open PLY files located in the point_cloud subfolders, corresponding to 7,000 or 30,000 iterations.
  • Attention: The input.ply files cannot be loaded, as they represent raw point clouds (not 3DGS) generated by Structure From Motion (SfM) during model reconstruction.

Models from 3ds-scan.de

Thanks to Christian Rochner, you can download:

Both models by 3ds-scan (Christian Rochner) and licensed under the CC BY-SA 4.0 license.

Models from teleportour.com

Thanks to Andrii Schramko, you can download:

Both models by Andrii Schramko.

106M Particles City:

  • Rasterization – Recommended pipeline: Mesh shader 3DGS. Requires ~10 GB of VRAM. Activate screen size culling for best render performance.
  • Ray tracing / Hybrid – Requires ~42 GB of VRAM (using TLAS instances, AABB+parametric for best quality, and stochastic modes for best performance).

15.3M Particles Green House:

  • Rasterization – Recommended pipeline: Mesh shader 3DGS.
  • Ray tracing / Hybrid – Requires ~8 GB of VRAM (using TLAS instances, AABB+parametric for best quality, stochastic modes for best performance).

Models from NVIDIA

The Bouquet of Flowers scene, which is automatically downloaded by CMake, can also be downloaded here. The model is released under the CC BY-SA 4.0 license, courtesy Jean-Eudes Marvie.

Note

Even if not generated by 3DGRT or 3DGUT, the 3DGS bouquet of flowers and the two 3DGS fountain scenes from 3ds-scan will provide acceptable visual results when rendered using 3DGRT or 3DGUT with Kernel degree set to "2 (Quadratic)."

3DGRT Datasets

Attention

The original version of the archive, named 3dgrt-mipnerf360-parametric-1M-v1.zip, was erroneously containing 3DGUT models. The archive has been renamed and updated with proper 3DGRT content.

We provide the reconstruction of the MipNerf 360 image set v1, using 3DGRT with parametric particles intersections and MCCM/Adam, providing 1M splats for each model.

Download: 3dgrt-mipnerf360-v1-parametric-degree-4-mcmc-selective-adam-1M (1.5GB).

Recommended visualization settings for the 3DGRT pipeline:

  • Kernel degree: 4 (Tesseractic)
  • Particle format: AABB + parametric

3DGUT Datasets

We provide the reconstruction of the MipNerf 360 image set v1, using 3DGUT with MCCM/Adam, providing 1M splats for each model.

Download: 3dgut-mipnerf360-v1-mcmc-selective-adam-1M (1.5GB).

Recommended visualization settings for the 3DGUT or hybrid 3DGUT+3DGRT pipelines:

  • Mip splatting anti-aliasing: Enabled
  • Kernel degree: 2 (Quadratic)