nishitanand/image-relighting-diffusion-data

Learning Illumination Control in Diffusion Models — Dataset (HF)

Public data and evaluation assets for Learning Illumination Control in Diffusion Models (ReALM-GEN @ ICLR 2026).

Code	github.com/nishitanand/image-relighting-diffusion
Model weights	huggingface.co/nishitanand/sd-image-relighting-model
Paper	arxiv.org/abs/2604.24877
Project site	nishitanand.github.io/relighting-diffusion-website

---

Download (CLI)

Install the Hugging Face CLI (pip install -U "huggingface_hub[cli]"), then:

huggingface-cli download nishitanand/image-relighting-diffusion-data \
  --repo-type dataset \
  --local-dir ./image-relighting-diffusion-data

You can also browse files on the dataset page and download subsets manually.

---

Contents

Training & test tensors / metadata

Folder	Description
data-train/	Synthetic degraded inputs + paired metadata for SD1.5 fine-tuning
data-test/	Held-out test split for quantitative evaluation
data_hf_train/	(Optional) Pre-sharded datasets format for faster dataloader startup
data-val/	(Optional) Extra split folder if you mirror the paper’s three-way split on disk

OOD qualitative pack

Path	Description
qualitative_comparison/selected-64/	64 face crops for the paper’s out-of-distribution qualitative evaluation
qualitative_comparison/ood_test_64.csv	64 rows — one editing_instruction per image (paper Figure 6 qualitative protocol)
qualitative_comparison/ood-64-results/	(Optional) Archived run outputs + ood_results.json

Paths in ood_test_64.csv are relative to the qualitative_comparison/ directory (e.g. selected-64/img000-lat349.png).

Optional evaluation bundles

Path	Description
evaluation/evaluation_results_comparison/	Saved comparisons (our model vs SD1.5 baseline) + JSON
evaluation/baseline_sdxl_long_descriptions/	SDXL baseline outputs
evaluation/baseline_flux_long_descriptions/	FLUX baseline outputs

---

Using with the GitHub code

1. Clone image-relighting-diffusion.
2. Download this dataset to e.g. ./hf_dataset (command above).
3. Training: point --data_dir (or symlink) at hf_dataset/data_hf_train or rebuild triplets from CSVs in the code repo — see the GitHub README “Download prebuilt data” and “Full pipeline”.
4. OOD: copy hf_dataset/qualitative_comparison/selected-64/ and ood_test_64.csv into the clone’s qualitative_comparison/ next to process_ood_test.py.
5. Quantitative eval: CSVs in the code repo use paths relative to the repository root; keep the same relative layout or rewrite prefixes.

FFHQ originals are not part of this dataset; obtain FFHQ under its license from NVlabs/ffhq-dataset and cite the StyleGAN / FFHQ paper and NVIDIA terms as required.

---

Citation (BibTeX)

@article{anand2026learning,
  title={Learning Illumination Control in Diffusion Models},
  author={Anand, Nishit and Suri, Manan and Metzler, Christopher and Manocha, Dinesh and Duraiswami, Ramani},
  journal={arXiv preprint arXiv:2604.24877},
  year={2026},
  note={ReALM-GEN @ ICLR 2026}
}

---

License

This dataset bundle on Hugging Face is released under CC BY-NC-SA 4.0. See LICENSE in this repository.

FFHQ. Curated training splits in the paper trace to Flickr-Faces-HQ (FFHQ). Individual FFHQ images were published on Flickr under licenses such as CC BY 2.0, CC BY-NC 2.0, and public-domain marks; the FFHQ dataset distribution (metadata, scripts, documentation) is provided by NVIDIA under CC BY-NC-SA 4.0. See NVlabs/ffhq-dataset.

Cite arXiv:2604.24877 when publishing results built on this bundle.