SyncLight Dataset

Base data for SyncLight: Single-Edit Multi-View Relighting (NeurIPS 2026) · Project page

SyncLight edits the light sources in one view of a scene (switching a light on or off, or changing its intensity or colour) and propagates that edit to other views of the same scene. This repository contains the data it was trained and evaluated on, released as one image per light source:

Source	Content	Base files
Infinigen	procedurally generated rooms (bathroom, bedroom, dining room, kitchen, living room)	linear HDR renders (EXR, float32 RGB), 1280×720
BlenderKit	artist-made indoor and outdoor scenes	linear HDR renders (EXR, float32 RGB), 1280×720
Real	multi-view photographs of real rooms	RAW photographs (DNG / CR2), one per light

Light is additive, so a weighted sum of the per-light images gives the scene under any combination of lights, intensities and colours. Instead of a fixed set of image pairs, this repository provides the per-light images and a script that renders as many relighting pairs as you want, with the colour palette, edit types, tone mapping and size under your control. The default configs reproduce the distribution and scale of the pairs used in the paper (about 1 M training pairs).

Statistics

Source	Split	Scenes	Per-light images	Size	Pairs (default config)	Rendered images (default config)
infinigen	train	59	1,777	18.0 GB	920,878	1,617,594
infinigen	test	9	664	6.9 GB	29,516	51,572
blenderkit	train	37	607	6.5 GB	44,420	74,950
blenderkit	test	3	48	0.5 GB	3,222	5,496
real	train	2	14	0.2 GB	12,756	22,700
real	test	4	36	0.7 GB	1,712	3,058
total		114	3,146	32.8 GB	1,012,504	1,775,370

A pair is one entry of transforms.json: two views of a scene before and after one light edit, plus the lightmap. Pair counts depend on the generation config; the numbers above are for scripts/configs/<source>.yaml, which match the paper.

Layout

infinigen/{train,test}/<room>_<seed>.tar      one tar per scene
blenderkit/{train,test}/<scene>.tar
real/{train,test}/<scene>.tar
scripts/                                      generator, configs, lightmap reader, extraction helper
metadata/manifest.jsonl                       one line per tar: file counts, size, sha256
metadata/skipped_scenes.jsonl                 scenes left out and why

Inside a scene tar:

<scene>/rendered/camera_<rig>_<cam>_<light>_<vis>.exr      one light alone, linear radiance, channels R, G, B (float32)
<scene>/rendered/camera_<rig>_<cam>_<light>_light_map.png  where that light is in the image
<scene>/rendered/camera_transforms.json                    camera poses and intrinsics per rig
<scene>/captured/scene<N>_<cam>_<light>_<vis>.{DNG,CR2}    real split: one RAW photo per light
<scene>/captured/..._light_map.png                         real split: light positions

<vis> says from where a light can be seen: camvis from this camera, rigvis only from other cameras of the rig, nonvis / ambient from no camera (ambient or window light). Cameras of one rig look at the same part of the scene from different viewpoints. Only camvis lights can be edited, because the edit is drawn on them in the reference view.

The EXRs keep only the final image (Combined pass) of the original Blender renders, without any change to the pixel values. The Blender scene files are not included: Infinigen scenes can be regenerated with Infinigen, and BlenderKit assets are covered by BlenderKit's own licences.

Quick start

pip install -U huggingface_hub numpy opencv-python pyyaml scipy pillow OpenEXR rawpy

# 1. download and extract (all sources, or e.g. "blenderkit test")
hf download davidserra9/synclight --repo-type dataset --include "scripts/*" --local-dir synclight
bash synclight/scripts/extract.sh ./synclight_raw blenderkit test

# 2. render pairs with the paper's settings (or your own)
python synclight/scripts/synclight_generate.py --config synclight/scripts/configs/blenderkit.yaml \
    --input ./synclight_raw/blenderkit/test --output ./synclight_pairs/blenderkit/test --workers 8

Each scene folder then contains the images, the lightmaps and a transforms.json listing every pair:

{"input_guide": "cam_0_1_00000_0_guide.png", "target_guide": "cam_0_1_00000_1_guide.png",
 "input_image": "cam_0_2_00000_0.png",       "target": "cam_0_2_00000_1.png",
 "lightmap": "cam_0_1_00000_0_to_1_lightmap.png",
 "relative_rotation": [...], "relative_translation": [...], "intrinsics": {...}}

• Guides (input_guide / target_guide): the reference view before and after the edit.
• Images (input_image / target): another view of the same rig, before and after.
• Lightmap: the edit, drawn in the reference view. Each lit pixel holds [activation, L, a, b]: activation −1 turns the light off, 1 sets it to intensity (L+1)/2 and CIELAB colour (a, b)·128.
• Format: lightmaps are 16-bit PNGs; read them with scripts/lightmap_io.py.
• Real split: captures are not calibrated, so the pose fields are zeros.

See scripts/README.md for how the generator works, every parameter (number of pairs, colour palettes from neutral to wild, edit types, tone-mapping operators, exposure, resolution), and the paper-equivalent settings.

Paper splits

The test folders hold the held-out scenes used for evaluation; no scene appears in both splits. The default configs reproduce the paper's sampling rules exactly, and the renderer is bit-exact with the one used for the paper. The sampling in the original runs was not seeded, though, so generated pairs are statistically equivalent to the paper's, not identical.

Citation

@inproceedings{serrano2026synclight,
  title     = {SyncLight: Single-Edit Multi-View Relighting},
  author    = {Serrano-Lozano, David and Bhattad, Anand and Herranz, Luis and Lalonde, Jean-Fran{\c{c}}ois and Vazquez-Corral, Javier},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}