JueGiroux/SheddingLightBenchmark

Shedding Light Benchmark

Data for the Shedding Light benchmark, which evaluates how accurately generative image models reproduce real-world illumination when inpainting light probes into photographs.

Project page | Paper | Code

Dataset structure

├── gt/                        Ground-truth data (~5 MB)
│   ├── real_probe_light_direction.json
│   ├── canonical_scenes.txt
│   └── excluded_scenes.txt
├── masks/                     GT probe masks (~120 MB)
│   └── <scene>/
│       ├── dir_N_matte_mask.png
│       └── dir_N_mirror_mask.png
├── precomputed/               Light direction estimates for 16 models (~95 MB)
│   └── <model>/
│       ├── <model>_probe_light_directions.json
│       └── <model>_chrome_pos_probe_light_directions.json
└── inpainted_images/          Full inpainted outputs (~122 GB)
    ├── FLUX.1-Fill-dev/<scene>/dir_N_inpainted.png
    ├── Z-Image-Turbo/<scene>/dir_N_inpainted.png
    └── stable-diffusion-2-inpainting/<scene>/dir_N_inpainted.png

<model> is the identifier listed under Evaluated models below.

Downloading subsets

Use the benchmark CLI to download only what you need:

pip install shedding-light-benchmark
shedding-light download --subsets gt masks          # just GT + masks (~125 MB)
shedding-light download --subsets precomputed        # add precomputed results (~95 MB)
shedding-light download --subsets inpainted_images   # add inpainted images (~122 GB)
shedding-light download                              # everything

Source data

Scene images

The benchmark scenes come from the Multilum dataset (Murmann et al., 2019), which provides 1,016 indoor scenes each photographed under 25 calibrated lighting directions. Multilum images are not redistributed here — download them from the Multilum project page.

Masks

The masks/ subset contains binary masks (PNG, same resolution as the Multilum images at MIP 2: 1500×1000) marking where the ground-truth matte and mirror chrome probes appear in each scene and lighting direction. These masks define the inpainting regions used by all evaluated models.

Inpainted images

The inpainted_images/ subset contains the full 1500×1000 inpainted outputs (seed 608313082563526) from three open-weight models. Each scene directory has 25 images (dir_0_inpainted.png through dir_24_inpainted.png), one per Multilum lighting direction.

Directory	Model	Organisation	License	Weights
FLUX.1-Fill-dev/	FLUX.1 Fill [dev]	Black Forest Labs	FLUX-1-dev Non-Commercial	HF
Z-Image-Turbo/	Z-Image-Turbo	Alibaba Tongyi-MAI	Apache 2.0	HF
stable-diffusion-2-inpainting/	Stable Diffusion 2 Inpainting	Stability AI	CreativeML Open RAIL++-M	HF

These images were produced by applying the models to Multilum scenes using the probe masks from the masks/ subset in this dataset. The inpainting prompt, parameters, and pipeline code are documented in the benchmark repository.

Note: The inpainted images inherit the license terms of both the generating model and the source data. The FLUX.1 Fill [dev] outputs are subject to the FLUX-1-dev Non-Commercial License.

Precomputed results

The precomputed/ subset contains the light directions estimated by the benchmark pipeline from each model's inpainted probes, for all 16 models evaluated in the paper. Each model directory holds two JSON files:

File	Content
<model>_probe_light_directions.json	Estimate from the diffuse probe inpainted at the matte-probe position (the main benchmark result)
<model>_chrome_pos_probe_light_directions.json	Estimate from the diffuse probe inpainted at the chrome-probe position

Both are lists of records {"scene", "direction_id", "direction_lumiere_x", "direction_lumiere_y", "direction_lumiere_z"} using the same convention as gt/real_probe_light_direction.json, so the angular error of any model is the angle between its estimate and the ground-truth vector for the same (scene, direction_id). Open-weight models were run with seed 608313082563526, the same seed as the inpainted_images/ subset; closed models do not expose a comparable seed and were run once. Entries are missing where the pipeline could not detect a usable probe. These files reproduce the light-direction figures of the paper with shedding-light figures.

Evaluated models

Identifier	Model	Organisation	Category	Availability
stable-diffusion-xl-1.0-inpainting-0.1	SDXL 1.0 Inpainting 0.1	Stability AI / Diffusers	Inpainting	Open
stable-diffusion-2-inpainting	Stable Diffusion 2 Inpainting	Stability AI	Inpainting	Open
stable-diffusion-3.5	Stable Diffusion 3.5	Stability AI	Inpainting	Open
FLUX.1-Fill-dev	FLUX.1 Fill [dev]	Black Forest Labs	Inpainting	Open
FLUX.2-klein-base-4B	FLUX.2 [klein] 4B Base	Black Forest Labs	Editing	Open
FLUX.2-dev	FLUX.2 [dev]	Black Forest Labs	Editing	Open
FLUX.2-klein-base-9B	FLUX.2 [klein] 9B Base	Black Forest Labs	Editing	Open
Kolors-Inpainting	Kolors Inpainting	Kuaishou	Inpainting	Open
Z-Image-Turbo	Z-Image-Turbo	Alibaba Tongyi-MAI	Inpainting	Open
HiDream-E1-1	HiDream-E1.1	HiDream AI	Editing	Open
HunyuanImage-2.1	HunyuanImage 2.1	Tencent	Inpainting	Open
Qwen-Image-Edit-2509	Qwen-Image-Edit-2509	Alibaba	Editing	Open
Gemini-2.5-Flash-Image	Gemini 2.5 Flash Image (Nano Banana)	Google DeepMind	Editing	Closed
Gemini-3.1-Flash-Image	Gemini 3.1 Flash Image (Nano Banana 2)	Google DeepMind	Editing	Closed
Firefly-Image-5	Adobe Firefly Image Model 5	Adobe	Editing	Closed
Photoshop-Generative-Fill	Photoshop Generative Fill	Adobe	Inpainting	Closed

How the data connects

1. Multilum scenes (external) provide the input photographs
2. Masks (this dataset) define where probes are placed in each scene
3. Inpainted images (this dataset) are model outputs after filling the masked regions
4. Precomputed results (this dataset) are the extracted lighting estimates
5. GT (this dataset) provides the ground-truth light directions for comparison

To evaluate your own model, you need the Multilum images and the masks from this dataset. See the benchmark code for the full pipeline.

Citation

@article{giroux2026shedding,
  title     = {Shedding Light: A Benchmark for Evaluating Lighting
               Understanding in Generative Image Models},
  author    = {Giroux, Justine and Hilliard, Jack Oliver and
               Hold-Geoffroy, Yannick and Vazquez-Corral, Javier
               and Lalonde, Jean-Fran{\c{c}}ois},
  journal   = {ACM Transactions on Graphics (TOG)},
  volume    = {45},
  number    = {6},
  articleno = {227},
  numpages  = {25},
  year      = {2026},
  month     = dec,
  doi       = {10.1145/3842579},
}

References

• Multilum dataset: Murmann, L., Gharbi, M., Aittala, M., & Durand, F. (2019). A Multi-Illumination Dataset of Indoor Object Appearance. ICCV. Project page
• FLUX.1 Fill [dev]: Black Forest Labs. HuggingFace
• Z-Image-Turbo: Alibaba Tongyi-MAI. HuggingFace | arXiv:2511.22699
• Stable Diffusion 2 Inpainting: Stability AI. HuggingFace | arXiv:2112.10752