leodriesch/open-hdri-1k

Open HDRI 1K

A consolidated, public-domain (CC0-1.0) collection of 3,491 equirectangular HDR environment maps at 1K resolution, gathered from five free HDRI libraries: Poly Haven, BlenderKit, ambientCG, CGEES and Open HDRI.

Every map is stored as a linear, high-dynamic-range .exr file alongside a tonemapped .jpg preview, with a per-asset metadata row (dimensions, source, author, license, SHA-256 checksum and tags).

Contents

Source	Assets	Author(s)	License
BlenderKit	1,990	various (credited per asset)	CC0-1.0
Poly Haven	970	various (credited per asset)	CC0-1.0
ambientCG	407	Lennart Demes (ambientCG)	CC0-1.0
CGEES	65	CGEES	CC0-1.0
Open HDRI	59	Open HDRI	CC0-1.0
Total	3,491		CC0-1.0

All maps are 2:1 equirectangular and stored at exactly 1024×512. BlenderKit assets are distributed at mixed resolutions (912×456 up to 3750×1875 despite the 1K label on the source files), so they were resampled to 1024×512 with area-averaging, which is energy-preserving and therefore the correct way to downsample radiance. Five BlenderKit files whose stored aspect ratio was not 2:1 (and so could not be valid equirectangular maps) were dropped.

Deduplication

The collection is deduplicated across and within sources. Duplicates are detected on the previews by correlating the per-row-mean-subtracted residual (which removes the sky-gradient/horizon structure every panorama shares) at the best yaw rotation, so rotated re-uploads are caught; the search is exhaustive over all pairs via a Cauchy–Schwarz bound on the max-over-rotation correlation. Removed: one BlenderKit re-upload of a Poly Haven asset, 26 duplicate BlenderKit files (byte-identical, re-encoded, or re-graded at a different exposure), and 2 exact duplicates within ambientCG (NightSkyHDRI011, NightSkyHDRI013).

Note that near-identical cloudless sky probes (*_puresky and similar) are not deduplicated: a clear sky carries almost no distinguishing structure, so image similarity cannot separate "duplicate" from "different clear sky" — the distinguishing information (absolute radiance) is not present in the tonemapped preview.

Curation: synthetic assets removed

This dataset targets real-world captured lighting, so obviously non-realistic environments have been removed by manual review of every asset's preview (all 3,845 were inspected, not sampled). 354 assets were removed:

Removed	Count	What it was
Artificial studio light-rigs	331	Abstract emitter setups — glowing shapes floating in a black void. Not scenes at all; used for product renders. Almost all from a single BlenderKit uploader.
Space / planetary environments	15	CG planets seen from orbit, planetary surfaces, star fields.
Other synthetic	6	Miscellaneous non-realistic CG environments.
Real captures, unusable quality	2	satara_night, satara_night_no_lamps (Poly Haven) — genuine night photographs, but so dominated by high-ISO sensor noise as to be unusable. Removed for quality, not because they are synthetic.

352 of the 354 were BlenderKit assets; 2 were Poly Haven.

A note on why this needed human review: neither filenames nor the model-assigned category label identify these reliably. The light-rigs alone appear under at least six unrelated name patterns (hdr-light-*, studio-light-black-white-light-*, black-white-light-studio-*, hdrlight-*, …), and while category == studio had high recall for them, it was not precise — several genuine indoor scenes carry the same label and were deliberately kept.

The remaining BlenderKit assets are still predominantly CG renders — photorealistic ones (rendered interiors, rendered sea-sunsets) rather than abstract. Only obviously non-realistic environments were removed. If your task requires genuine photographic captures, filter to Poly Haven / ambientCG / CGEES / Open HDRI, or treat source as a domain label.

Layout

open-hdri-1k/
├── README.md
├── train.csv
├── validation.csv
├── test.csv
└── data/
    ├── polyhaven/
    │   └── <id>/
    │       ├── <id>_1k.exr
    │       └── <id>_preview.jpg
    ├── blenderkit/
    │   └── <id>/
    │       ├── <id>_1k.exr
    │       └── <id>_preview.jpg
    ├── ambientcg/
    │   └── <id>/
    │       ├── <id>_1k.exr
    │       └── <id>_preview.jpg
    ├── openhdri/
    │   └── <id>/
    │       ├── <id>_1k.exr
    │       └── <id>_preview.jpg
    └── cgees/
        └── <id>/
            ├── <id>_1k.exr
            └── <id>_preview.jpg

Splits

Split	Poly Haven	BlenderKit	ambientCG	CGEES	Open HDRI	Total
train	789	1,587	325	53	47	2,801
validation	90	195	41	6	6	338
test	91	208	41	6	6	352
total	970	1,990	407	65	59	3,491

• ambientCG, Open HDRI and CGEES are split randomly 80 / 10 / 10 within each source (seed = 42).
• BlenderKit is split randomly 80 / 10 / 10 (seed = 20260711).
• Poly Haven uses a fixed held-out evaluation list of 181 assets, split 50 / 50 (90 / 91) into validation / test; all remaining Poly Haven assets go to train.

Split proportions are preserved after curation (80.2 / 9.7 / 10.1); because assets were removed individually rather than re-split, the per-source split counts are no longer exactly 80 / 10 / 10.

Columns

Each split CSV has one row per HDRI with the following columns:

Column	Description
id	Asset slug (unique within its source).
file_path	Path to the HDR .exr, relative to the dataset root.
preview_path	Path to the tonemapped .jpg preview, relative to the dataset root.
width, height	Pixel dimensions of the EXR.
format	Always exr.
source	Poly Haven, BlenderKit, ambientCG, CGEES or Open HDRI.
source_url	Canonical URL of the asset (or its download URL for Open HDRI).
license	Always CC0-1.0.
author	Credited author(s).
sha256	SHA-256 checksum of the EXR file.
tags	Comma-separated tags/categories (may be empty for Open HDRI, CGEES and BlenderKit).
category	Scene-type label (9 classes), assigned by gemini-3.1-flash-lite.
lighting	Lighting / time-of-day label (5 classes), assigned by gemini-3.1-flash-lite.

Classification (category & lighting)

Every HDRI was auto-labeled along two independent axes by the gemini-3.1-flash-lite vision model (low reasoning effort, strict structured output), from its equirectangular preview image. These are model-generated heuristic labels (not human-verified); expect occasional errors.

category (scene type): sky_only (sky dominates), natural_landscape (forest / field / mountain / desert / snow), water_coastal (beach / lake / river / harbor), urban_exterior (streets / buildings / plazas), industrial_outdoor (factory yards / construction / ruins), indoor_residential (homes), indoor_public (offices / classrooms / halls / shops), indoor_industrial (workshops / warehouses / garages), studio (photographic studio / cyclorama).

lighting (time-of-day / light): daylight, golden_hour (sunrise / sunset), overcast, night, artificial (interior electric light).

A peek at some sample HDRIs:

Previews

Previews are tonemapped sRGB JPEGs:

• ambientCG previews are the tonemapped JPEGs bundled with the original assets.
• Poly Haven, Open HDRI and CGEES previews are generated from the EXRs using a Reinhard global tonemap with log-average auto-exposure, followed by sRGB encoding.
• BlenderKit previews are generated from the EXRs by normalising exposure to the 99th percentile of radiance (so a bright sun does not crush the rest of the scene), then gamma-2.2 encoding.

Previews are for inspection only — always use the .exr files for image-based lighting and any HDR-dependent task.

Usage

The CSVs store paths to the image files (not embedded image bytes), so the dataset is convenient to stream while keeping the heavy EXRs on disk.

from datasets import load_dataset

ds = load_dataset("<your-namespace>/open-hdri-1k")
row = ds["train"][0]
print(row["id"], row["width"], row["height"], row["source"])

# Load the actual HDR map (paths are relative to the dataset root)
import cv2, os
os.environ["OPENCV_IO_ENABLE_OPENEXR"] = "1"
hdr = cv2.imread(row["file_path"], cv2.IMREAD_ANYCOLOR | cv2.IMREAD_ANYDEPTH)  # BGR float32

License & Attribution

All assets are released under the Creative Commons CC0 1.0 Universal public-domain dedication. No attribution is legally required, but please credit the original libraries and authors where practical:

• Poly Haven — https://polyhaven.com (individual authors credited in the author column)
• BlenderKit — https://www.blenderkit.com (individual uploaders credited in the author column)
• ambientCG / Lennart Demes — https://ambientcg.com
• Open HDRI — https://openhdri.org
• CGEES — https://cgees.com/hdris

This collection simply re-packages those assets; no rights are claimed over them.