jadgardner/neusky-synthetic

NeuSky Synthetic

NeuSky Synthetic is a five-scene, multi-illumination outdoor dataset for evaluating neural inverse rendering, intrinsic decomposition, novel-view synthesis and illumination estimation. It contains 1,500 rendered views at 1920 x 1080 resolution, with 250 training, 25 validation and 25 test views per scene.

This benchmark was developed as a thesis extension to The Sky's the Limit. It was not part of the original ECCV 2024 evaluation.

Property	Value
Scenes	5
Total views	1,500
Training views	1,250
Validation views	125
Test views	125
Resolution	1920 x 1080
Illumination environments	167
Release	1.0
Generator commit	8e074a09385d975766742c2e4b3bf4e6866a7211

Scenes

• abandoned_buildings
• apartment_building
• arlanda_uppsala_cathedral
• glass_building
• interstellar_house

Each scene is rendered under many Poly Haven HDR environments with known horizontal rotation. Training views additionally vary camera focal length and exposure. Camera intrinsics, camera-to-world transforms, exposure, environment identity and environment rotation are recorded per frame.

Splits And Layers

The training split contains only the inputs used for reconstruction:

• tone-mapped RGB;
• foreground/sky semantic masks.

Validation and test additionally provide the ground-truth decomposition:

• diffuse albedo;
• world-space surface normals;
• metric depth;
• roughness;
• metallic;
• transmission;
• index of refraction.

Every scene also contains points3d.ply, used for scene centring and Gaussian-splatting initialisation, and transforms.json, containing all camera and illumination metadata.

Each independently downloadable archive extracts to:

scenes/<scene>/
  transforms.json
  points3d.ply
  train/
    rgb/
    cityscapes_mask/
  validation/
    rgb/
    cityscapes_mask/
    albedo/
    normal/
    depth/
    roughness/
    metallic/
    transmission/
    ior/
  test/
    <same layers as validation>

The mask images use the Cityscapes colours expected by NeuSky: sky is (70, 130, 180) and the reconstructed scene is (70, 70, 70).

Download

Download the complete release with the Hugging Face CLI, verify it, then extract the five scene archives:

hf download jadgardner/neusky-synthetic \
  --repo-type dataset \
  --revision v1.0 \
  --local-dir neusky-synthetic

(cd neusky-synthetic && sha256sum -c SHA256SUMS)
for archive in neusky-synthetic/archives/*.tar.zst; do
  tar --zstd -xf "$archive" -C neusky-synthetic
done

Download one scene:

hf download jadgardner/neusky-synthetic \
  --repo-type dataset \
  --revision v1.0 \
  archives/interstellar_house.tar.zst \
  --local-dir neusky-synthetic

tar --zstd -xf \
  neusky-synthetic/archives/interstellar_house.tar.zst \
  -C neusky-synthetic

An individual scene can also be downloaded without the CLI:

curl -L \
  "https://huggingface.co/datasets/jadgardner/neusky-synthetic/resolve/v1.0/archives/interstellar_house.tar.zst?download=true" \
  -o interstellar_house.tar.zst

zstd is required for extraction. MANIFEST.json records the size and SHA256 of every release file, and SHA256SUMS can be checked before extraction.

NeuSky

The loader, training code and deterministic generation pipeline are in the NeuSky repository.

git clone --recurse-submodules https://github.com/JADGardner/neusky.git
cd neusky

PYTHONPATH=. python scripts/train_synthetic.py \
  --data /path/to/neusky-synthetic/scenes/abandoned_buildings

See provenance/scene_sources.json for the source models and scene hashes, and provenance/scene_render_configs.json for the accepted camera and rendering profiles.

Baseline Results

These are the accepted means over all five scenes. The full per-scene results and additional raw/exposure-aligned columns are in benchmark/synthetic_results.csv.

Method	Known PSNR EA ↑	SSIM ↑	LPIPS ↓	Left-Half PSNR ↑	SSIM ↑	LPIPS ↓	Albedo PSNR ↑	Normal MAE ↓	Depth RMSE ↓
NeuSky	16.54	0.506	0.602	15.94	0.575	0.570	17.84	46.49	24.91
NeRF-OSR	15.48	0.335	0.591	12.79	0.468	0.523	18.43	38.44	27.13
GS-IR	16.11	0.597	0.415	-	-	-	17.30	55.46	25.86

Known-illumination novel-view synthesis uses the ground-truth HDR environment and rotation. Its PSNR is exposure-aligned. In the left-half fitting track, illumination is fitted using only the left half of each held-out image and the full image is scored. GS-IR has no entry because its scene-wide lighting has no per-image estimation mechanism. Decomposition metrics use non-sky, valid-object masks; albedo PSNR is scale-aligned and normal MAE is measured in degrees.

Provenance

The 167 HDR environments are CC0 assets from Poly Haven. Their identifiers, current download metadata and accepted-generation checksums are included under provenance/. Poly Haven has made minor upstream revisions to 36 EXRs since the accepted render. Current upstream files are supported for source re-rendering; this prepared release is canonical for reproducing the reported benchmark.

The editable Blender scenes are not included. They contain extractable third-party models that cannot be redistributed in source form under their licences. provenance/scene_sources.json records their authors, source pages, licences and accepted scene hashes.

Licence

The prepared renders, masks, camera metadata, point clouds and ground-truth layers in this repository are released under the Creative Commons Attribution 4.0 International licence. See LICENSE.md.

This licence does not apply to the external HDRIs or editable third-party models referenced by the provenance files. The HDRIs are distributed separately by Poly Haven under CC0; the editable models are not included.

Limitations

• The dataset contains five synthetic outdoor architectural scenes and does not represent the full diversity of real outdoor captures.
• Camera paths are sampled around the principal scene content and do not reproduce a specific handheld capture trajectory.
• The training data contains deliberately broad illumination and exposure variation; methods assuming a single fixed scene illumination are outside their intended setting.
• Ground-truth decomposition layers are reserved for validation and test and must not be used as training supervision when reproducing the reported protocol.

Citation

The synthetic benchmark should be cited as part of the accompanying thesis. For NeuSky, cite:

@inproceedings{gardner2024sky,
  author    = {James A. D. Gardner and Evgenii Kashin and
               Bernhard Egger and William A. P. Smith},
  title     = {The Sky's the Limit: Relightable Outdoor Scenes via a
               Sky-Pixel Constrained Illumination Prior and Outside-In
               Visibility},
  booktitle = {European Conference on Computer Vision},
  pages     = {126--143},
  year      = {2024}
}

Release prepared on 2026-07-30.