LuxRemix: Lighting Decomposition and Remixing for Indoor Scenes
Ruofan Liang, Norman Müller, Ethan Weber, Duncan Zauss, Nandita Vijaykumar, Peter Kontschieder, Christian Richardt
CVPR 2026 · Project page · arXiv:2601.15283 · Download
LuxRemix is a Blender-rendered dataset of indoor scenes with per-light decomposition (one-light-at-a-time / OLAT), produced as panoramic equirectangular HDR. It is designed for training models that decompose and re-mix indoor illumination from a single image.
This repository contains the tooling, reference dataloader, and documentation needed to download, validate, and consume the released dataset. The dataset itself is hosted at https://ai.meta.com/datasets/luxremix-dataset/.
| Scenes | 12,439 (12,039 training + 400 test) |
| Total files | ~1.48 M |
| Total size | ~9.7 TB |
| Shards | 1,244 tar archives (1,204 training + 40 test), ~8 GB each |
| Image resolution | 2048 × 1024 (equirectangular panorama) |
| Viewpoints per scene | 4 for ~99% of scenes; 114 have 3 views and 1 has 1 view (see DATACARD.md) |
| Lights per scene | 2–7 (mean ~5) |
| File formats | EXR (HDR), PNG (LDR + masks), JPEG (auxiliary), JSON (metadata) |
See FORMAT.md for the per-file layout and DATACARD.md for composition, intended uses, and known limitations.
# 1. Install dependencies
pip install -r requirements.txt
git submodule update --init # AgX OCIO config
# 2. Get the current shard URL list from the dataset portal:
# https://ai.meta.com/datasets/luxremix-dataset/ (download `dataset-shards.txt`,
# a TSV with header `file_name<TAB>cdn_link` and one row per shard)
# 3. Download (and unpack) one test shard as a smoke test
(head -1 dataset-shards.txt && sed -n '2p' dataset-shards.txt) > one_shard.txt
python download.py one_shard.txt --output-dir ./luxremix --split test \
--unpack
# 4. Inspect a downloaded scene
python examples/quickstart.py ./luxremix/<scene_id>/
# 5. Iterate the reference PyTorch dataloader
python examples/training_loop.py ./luxremix/For the full dataset (~9.7 TB), run step 3 against the full
dataset-shards.txt (drop --unpack if you'd rather keep tars on disk).
USAGE.md covers every flag.
| Purpose | Files |
|---|---|
| Public docs | README.md, LICENSE.md, CITATION.cff, CONTRIBUTING.md, CODE_OF_CONDUCT.md, docs/{DATACARD,FORMAT,USAGE,ACKNOWLEDGMENTS}.md |
| Download | download.py (HTTPS, parallel, optional unpack) |
| Reference dataloader | dataset.py (PyTorch IterableDataset), examples/quickstart.py, examples/training_loop.py |
| Perspective regeneration | tools/generate_test_sv.py, tools/generate_test_mv.py, tools/erp_to_perspective.py, data/mask_strategies_*.json, data/camera_params_352.json |
| Visualisation | tools/generate_vis_grid.py |
| Manifests | data/shards_training.json, data/shards_test.json |
The LuxRemix dataset and accompanying code are released under the LuxRemix Dataset License Agreement, derived from the Aria Synthetic Environments Dataset License Agreement (see LICENSE.md for the full text). Use is restricted to non-commercial research; redistribution of the data is prohibited without prior written permission from Meta.
@InProceedings{LuxRemix,
author = {Ruofan Liang and Norman M{\"u}ller and Ethan Weber and Duncan Zauss and Nandita Vijaykumar and Peter Kontschieder and Christian Richardt},
title = {{LuxRemix}: Lighting Decomposition and Remixing for Indoor Scenes},
booktitle = {CVPR},
year = {2026},
arxiv = {2601.15283},
url = {https://luxremix.github.io},
}(Also available as machine-readable CITATION.cff.)
LuxRemix builds on the Aria Synthetic Environments scenes, env maps from Poly Haven, the Infinigen procedural-asset library, and the AgX color science by EaryChow. See ACKNOWLEDGMENTS.md for detailed attributions.