Collection · CC0

Polyhaven 10

10 multi-view training datasets rendered from CC0 3D models. Ready for NeRF, 3D Gaussian Splatting, and 3D reconstruction research.

Download on HuggingFace Pipeline Details →

196 Views 1024×1024 Full Sphere 7 Modalities CC0 Pre-trained Splats

Trained 3D Gaussian Splats

Each dataset was trained with nerfstudio splatfacto (20k iterations). Drag to orbit, scroll to zoom.

Use ← → to browse all 10 models

Specs

Objects	10
Views	196 per object
Resolution	1024 × 1024
Coverage	Full sphere (±89° elevation)
Distribution	Fibonacci golden-angle spiral
Point cloud	~200k points
Background	Transparent (RGBA)
Modalities	RGB, depth 8-bit, depth 16-bit, normals, masks, point cloud, camera poses
Format	nerfstudio / instant-ngp (transforms.json)
License	CC0 (public domain)
Source	Polyhaven

Camera Distribution

196 camera positions distributed on a full sphere using Fibonacci golden-angle spiral

196 views on a full sphere (±89° elevation), Fibonacci golden-angle spiral distribution

Quick Start

Download

pip install requests
python download_all.py              # all datasets
python download_all.py --splats     # + pre-trained splats
python download_all.py --object apple  # single object

Train with nerfstudio

ns-train splatfacto --data ./dxgl-datasets/apple \
  --max-num-iterations 20000 \
  --pipeline.model.sh-degree 3 \
  --pipeline.model.background-color white \
  --pipeline.model.cull-alpha-thresh 0.2 \
  --pipeline.model.use-scale-regularization True

Or train all at once

python train_all.py                 # ~100 min on RTX 4000 Pro Ada
python train_all.py --object apple  # single object (~10 min)

What's in Each ZIP

dataset/
├── images/           196 RGB frames (PNG, RGBA)
├── depth/            8-bit grayscale depth maps
├── depth_16bit/      16-bit grayscale depth maps
├── normals/          World-space normal maps
├── masks/            Binary alpha masks
├── transforms.json   Camera poses (nerfstudio format)
└── points3D.ply      Sparse point cloud (~200k pts)

Citation

@misc{dxgl_polyhaven10_2026,
  title  = {Polyhaven 10 — Multi-View Datasets for NeRF and 3D Gaussian Splatting},
  author = {DXGL},
  year   = {2026},
  url    = {https://dx.gl/datasets/polyhaven-10},
  note   = {10 CC0 objects, 196 views at 1024x1024, full sphere coverage. Rendered via DX.GL.}
}

Generate your own datasets

Upload any GLB model and render calibrated multi-view datasets with the DX.GL pipeline. 10 model uploads with shareable turntable videos included.

Get Started Free Questions? Contact us →