3D-anchored Light Probes (3DLP)

3DLP is a real-world benchmark dataset for studying whether image editing models can reproduce lighting changes in a physically consistent way. Each scene contains a visible light probe: a spherical bulb mounted on a stand, captured once with the light turned on and once with it turned off.

The dataset is released as anonymous review material for the accompanying paper, Do Image Editing Models Understand Lighting? It is intended to support evaluation of light transport in image editing: shadows, highlights, attenuation, and interactions with reflective, transparent, and metallic materials.

What Is Included

The release contains 1,000 paired captures across anonymized real-world locations and scenes.

• main: full-resolution data for benchmark use.
• preview: flat, side-by-side JPEG previews for quick visual inspection.

The location and scene names are anonymized. Identifiers such as location_3, scene_4, and view_1 are stable dataset IDs and do not encode real place names.

Capture Notes

The main subset was captured with an emphasis on radiometric quality rather than speed. Each HDR image is built from two 14-bit, 9-exposure brackets with 1 EV spacing; the two brackets are averaged (stacking) to reduce noise before HDR merging. Including capture, bracketing, and repeated measurements, each of the 1,000 on/off pairs took roughly 5 minutes to record. The images are then corrected for lens distortion and vignetting and stored as linear TIFF files.

The TIFF files retain the camera-captured EXIF metadata. This includes exposure information useful for recovering the capture conditions and for aligning the on/off measurements in physically motivated evaluations. Across a viewpoint, focus and aperture were kept fixed, and five on/off pairs were recorded with a shared base exposure setup.

Dataset Layout

.
+-- main/
|   +-- metadata.jsonl
|   +-- images/
|       +-- location_<i>/
|           +-- scene_<j>/
|               +-- view_<k>/
|                   +-- images_real/
|                   |   +-- images_<n>_on.tiff
|                   |   +-- images_<n>_off.tiff
|                   +-- images_labels/
|                       +-- <label_name>/
|                           +-- image_<n>_<label_name>.png
+-- preview/
    +-- metadata.jsonl
    +-- images/
        +-- location_<i>_scene_<j>_view_<k>_image_<n>_on_off.jpg

In main, each metadata row points to one HDR on/off pair and the corresponding label masks. The real images are stored as TIFFs. Label masks cover regions relevant for detailed lighting analysis, including light bulbs, windows, mirror surfaces, transparent surfaces, metallic surfaces, cast shadows, specular highlights, and ambient light effects.

In preview, each JPEG concatenates the tonemapped off/on pair horizontally for fast visual browsing. The left half is the light-off image; the right half is the light-on image.

Downloading Only The Preview

The preview subset is useful for checking the contents without downloading the full HDR release. These files are compressed JPEG images and are intended only for quick visual inspection of the dataset diversity. They are not meant for scientific experiments or quantitative evaluation, and they may contain JPEG compression artifacts. Use the TIFF files in main for benchmark experiments and analysis.

Using the Hugging Face CLI:

hf download AnonymousAuthorsXYZ/3DLP \
  --type dataset \
  --include "README.md" \
  --include "preview/**" \
  --local-dir 3dlp_preview

Using Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="AnonymousAuthorsXYZ/3DLP",
    repo_type="dataset",
    allow_patterns=["README.md", "preview/**"],
    local_dir="3dlp_preview",
)

To access the full benchmark data, download main/** as well.

Metadata

main/metadata.jsonl contains one row per image pair. Important fields include:

• image_on_file_name: HDR image with the light probe turned on.
• image_off_file_name: HDR image with the light probe turned off.
• ambient_light: coarse lighting condition for the scene.
• *_file_name: optional segmentation masks for annotated lighting and material regions.

preview/metadata.jsonl contains one row per preview image:

• image_file_name: path to the fused preview JPEG.
• location, scene, view: anonymized hierarchical identifiers.
• image_index: pair index within the view.

Intended Use

3DLP is designed for evaluating lighting edits against measured real-world ground truth. A typical use case is to give an image editing model the light-off image and ask it to turn the visible light probe on, or give it the light-on image and ask it to turn the probe off. The generated result can then be compared with the corresponding captured target image.