---
pretty_name: 3DLP Ablations
license: mit
tags:
  - light transport
  - image editing
language:
  - en
viewer: true
configs:
  - config_name: ambient_light_ablation
    default: true
    metadata_filenames:
      - metadata.jsonl
    data_files:
      - split: eval
        path:
          - "ambient_light_ablation/images/**"
          - "ambient_light_ablation/metadata.jsonl"
  - config_name: light_type_ablation
    metadata_filenames:
      - metadata.jsonl
    data_files:
      - split: eval
        path:
          - "light_type_ablation/images/**"
          - "light_type_ablation/metadata.jsonl"
---

# 3DLP Ablations

This companion dataset contains 450 real HDR image pairs from
the ambient-light and light-probe ablations of the anonymous 3D-anchored Light
Probes (3DLP) release. Each pair captures the same scene and camera view with
the target light probe switched on and off. The dataset accompanies the paper
*Do Image Editing Models Understand Lighting?*

## What Is Included

### Ambient-light ablation

`ambient_light_ablation` contains 300 complete image pairs. This
controlled experiment studies whether image-editing models can disentangle the
illumination caused by the target probe from the scene's existing ambient
lighting. The target remains the spherical probe used by the main 3DLP
benchmark, while the ambient sources are placed behind the camera so that the
requested on/off edit remains unambiguous.

The subset covers six ambient-light arrangements of increasing complexity:

- a point-like spherical source;
- a 60 cm x 120 cm area light;
- one directional lamp;
- two directional lamps;
- one complex basket lamp; and
- two complex basket lamps.

These conditions are recorded across multiple indoor scenes and viewpoints,
allowing the interaction between ambient illumination, scene geometry,
shadows, highlights, and the target probe's light transport to be compared.
This experiment is described in the supplemental section *Ambient Light
Complexity Evaluation*.

### Light-probe-type ablation

`light_type_ablation` contains 150 complete image pairs captured with six
different target-probe geometries and emission patterns:

- spherical;
- shaded;
- light bar;
- directional;
- daylight; and
- basket.

This subset motivates the spherical probe used in the main benchmark. For a
turn-on edit, more complex lamp geometry also requires the model to infer a
more complex, spatially varying emission pattern. For a turn-off edit, that
pattern is already visible in the input image. The experiment is described in
the main paper section *Light Probe Ablation*.

Each metadata row describes one complete real light-on/light-off pair. All real
TIFF files present in the source ablation folders are included; generated model
outputs are intentionally excluded. One light-off TIFF in the ambient-light
subset has no corresponding light-on capture; it is included in the image tree
but is not listed as a complete metadata sample.

## Dataset Layout

```text
.
+-- 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
```

Location, scene, view, and image identifiers are anonymized. Their numeric IDs
are stable across the two subsets and do not encode real place names.

## Metadata

Both subsets use `metadata.jsonl`, with one JSON object per complete captured
pair:

- `image_on_file_name`, `image_off_file_name`: real HDR TIFF captures.
- `Ambient_Light`: the ambient-light arrangement or probe-light type.
- `u`, `v`: probe coordinates retained from the source metadata.
- `annotations`: optional shadow and highlight annotations.

## Intended Use

These real TIFF captures support the ablation experiments reported with 3DLP.