Datasetpublic

gbmpi/genlit_dataset

GenLit Dataset This is the official synthetic dataset for GenLit: Reformulating Single-Image Relighting as Video Generation, published at SIGGRAPH Asia 2025. Paper: ACM Digital Library | arXiv Project Page: genlit.is.tue.mpg.de Code: GitHub Dataset Description GenLit reformulates single-image relighting as a video generation task, where the scene remains static while a point light moves through 3D space. This dataset contains synthetic image sequences rendered… See the full description on the dataset page: https://huggingface.co/datasets/gbmpi/genlit_dataset.

sourceHugging Faceotherupdated 8mo agoView on Hugging Face

1likes300downloads

Dataset Card

GenLit Dataset

This is the official synthetic dataset for GenLit: Reformulating Single-Image Relighting as Video Generation, published at SIGGRAPH Asia 2025.

—Paper: ACM Digital Library | arXiv
—Project Page: genlit.is.tue.mpg.de
—Code: GitHub

Dataset Description

GenLit reformulates single-image relighting as a video generation task, where the scene remains static while a point light moves through 3D space. This dataset contains synthetic image sequences rendered with Blender, showing objects under varying lighting conditions controlled by a moving point light source.

Each sample consists of 14 frames depicting the same 3D scene with different point light positions, along with comprehensive metadata to reproduce the scene.

Dataset Statistics

Split	Size
`train`	~342 GB
`test`	~31 GB

Dataset Structure

Each sample contains:

Field	Type	Description
`images`	`Sequence[Image]`	14 rendered frames showing lighting variation
`frame_idxs`	`Sequence[string]`	Frame identifiers
`object_id`	`string`	Unique identifier for the 3D object
`rotation`	`string`	Object rotation variant
`seq_id`	`string`	Sequence identifier
`metadata`	`dict`	Scene configuration (see below)

Metadata Fields

Field	Type	Description
`objects`	`list`	List of objects with id, dataset source, translation, rotation, scale
`scene_id`	`string`	Trajectory identifier
`point_light_trajectory`	`Array2D[14, 3]`	3D positions of the point light for each frame
`point_light_intensity`	`Sequence[float]`	Point light intensity per frame
`env_light_intensity`	`Sequence[float]`	Environment/ambient light intensity per frame
`floor_texture`	`dict`	Floor texture name and rotation
`hdri`	`dict`	HDRI environment map name and rotation
`camera_matrix`	`Array2D[3, 4]`	Camera projection matrix
`image_ids`	`Sequence[string]`	Image identifiers

<!-- ## Usage

python

from datasets import load_dataset

# Load the test split
dataset = load_dataset("gbmpi/genlit_dataset", split="test")

# Access a sample
sample = dataset[0]
images = sample["images"]  # List of 14 PIL Images
light_positions = sample["metadata"]["point_light_trajectory"]  # [14, 3] array

Alternative Download

For bulk downloads, you can use our download script: https://genlit.is.tue.mpg.de/download.php

Citation

bibtex

@inproceedings{bharadwaj2025genlit,
  title={GenLit: Reformulating Single-Image Relighting as Video Generation},
  author={Bharadwaj, Shrisha and Feng, Haiwen and Becherini, Giorgio and Abrevaya, Victoria Fernandez and Black, Michael J.},
  booktitle={SIGGRAPH Asia 2025 Conference Papers},
  year={2025},
  publisher={ACM},
  doi={10.1145/3757377.3763970}
}

License

This dataset is released under the Max Planck Institute for Intelligent Systems Non-Commercial Research License.

Before using this data, you must read and accept the license conditions at: https://genlit.is.tue.mpg.de/license.html

Key Terms

—Permitted: Non-commercial scientific research, education, and artistic projects
—Prohibited: Commercial use, pornographic/military/surveillance applications, redistribution
—Attribution: You must cite the SIGGRAPH Asia 2025 paper in any publications

For commercial licensing inquiries, contact: ps-license@tue.mpg.de

Contact

For questions about the dataset, please visit the project page or open an issue on the GitHub repository.