3D-anchored Light Probes (3DLP)

Do Image Editing Models Understand Lighting?

Tim Küchler^1,3,*  ·  Johann-Friedrich Feiden^1,*  ·  Matthias Nießner²  ·  Carsten Rother¹

¹ Heidelberg University  ·  ² Technical University of Munich  ·  ³ Zuse School ELIZA
^* Equal contribution

3DLP is a real-world benchmark dataset for studying whether image editing models reproduce lighting changes in a physically consistent way. Each pair captures the same scene and camera view with a visible light probe turned off and on, providing measured ground truth for shadows, highlights, light falloff, and interactions with different materials.

🔑 What's Included

The release contains 1,000 HDR image pairs for the main benchmark and 479 additional pairs for the ablation experiments, alongside JPEG previews of the main benchmark.

Subset	Samples	Contents
main	1,000 pairs	Full-resolution, linear HDR TIFF captures with material and lighting-effect masks.
preview	1,000 previews	Tonemapped, side-by-side JPEGs of the main pairs: light off on the left, light on on the right.
ambient_light_ablation	329 pairs	A spherical target probe under six ambient-light arrangements, including area, directional, point, and complex lamps.
light_type_ablation	150 pairs	Six target-probe types: spherical, shaded, light bar, directional, daylight, and basket lamps.

📷 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 brackets are averaged to reduce noise, then merged into an HDR image. Including capture, bracketing, and repeated measurements, each on/off pair took roughly five minutes to record. Images are corrected for lens distortion and vignetting and stored as linear TIFFs.

The TIFFs retain camera-captured EXIF metadata, including exposure information. Focus and aperture were kept fixed, and five on/off pairs were recorded per viewpoint with a shared base exposure setup.

🗂️ Dataset Layout

3DLP/
├── 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
├── ambient_light_ablation/
│   ├── metadata.jsonl
│   └── images/location_<i>/scene_<j>/view_<k>/
│       └── images_real/<ambient_condition>/
│           ├── images_<n>_on.tiff
│           └── images_<n>_off.tiff
└── light_type_ablation/
    ├── metadata.jsonl
    └── images/location_<i>/scene_<j>/view_<k>/
        └── images_real/<light_type>/
            ├── images_<n>_on.tiff
            └── images_<n>_off.tiff

The masks in main identify light bulbs, windows, mirrors, transparent and metallic surfaces, cast shadows, specular highlights, and ambient lighting effects. The ablation subsets contain real on/off captures and their pair metadata.

📝 Metadata

Each subset has a metadata.jsonl file with one JSON object per sample. File paths are relative to the directory containing that metadata file.

Main benchmark — main/metadata.jsonl:

• image_on_file_name, image_off_file_name: the paired HDR captures.
• ambient_light: the scene's ambient lighting condition.
• *_file_name: paths to optional segmentation masks for lighting and material regions.
• u, v: optional probe coordinates.

Previews — preview/metadata.jsonl:

• image_file_name: the side-by-side preview JPEG.
• location, scene, view, image_index: identifiers for the corresponding main pair.

Ablations — ambient_light_ablation/metadata.jsonl and light_type_ablation/metadata.jsonl:

• image_on_file_name, image_off_file_name: the paired HDR captures.
• Ambient_Light: the ambient-light arrangement or target-probe type. This field uses the capitalization shown here in both ablation subsets.
• u, v: probe coordinates retained from the capture metadata.
• annotations: optional shadow and highlight annotations.

🧷 Citation

If you use 3DLP in your research, please cite our paper:

@misc{kuchler2026imageeditingmodelsunderstand,
  title={Do Image Editing Models Understand Lighting?},
  author={Tim K{\"u}chler and Johann-Friedrich Feiden and Matthias Nie{\ss}ner and Carsten Rother},
  year={2026},
  eprint={2606.26738},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2606.26738}
}

📄 License

This dataset is released under the MIT License. Commercial and non-commercial use, modification, and redistribution are permitted, provided the copyright and permission notices are retained.