LightYourFace-160K

LightYourFace-160K provides physically rendered face fill-light components with six-dimensional lighting control. It accompanies Light Up Your Face: A Physically Consistent Dataset and Diffusion Model for Face Fill-Light Enhancement, ICML 2026.

Project · Code · Paper · Model

Dataset files

Archive	Lighting components	Metadata records	Edited inputs
train.tar.gz	165,419	165,419	—
LYF-Val.tar.gz	3,006	3,006	—
LYF-EditVal.tar.gz	3,006	3,006	1,002

Each archive contains PNG files and metadata.jsonl directly at its root. Training records refer to FFHQ source portraits; validation records refer to CelebA-Test source portraits. Source portraits are obtained separately from the original datasets under their respective terms. LYF-EditVal additionally includes edited input portraits for evaluating fill-light enhancement under underexposure and side lighting.

The PNG named {source_id}_{variant}.png stores the rendered fill-light component. In LYF-EditVal, {source_id}.png stores the edited input. The two validation sets share lighting components and physical parameters while using different input illumination settings.

Download

hf download GOBUNU/LightYourFace-160K --repo-type dataset --local-dir LightYourFace-160K
mkdir -p train LYF-Val LYF-EditVal
tar -xzf LightYourFace-160K/train.tar.gz -C train
tar -xzf LightYourFace-160K/LYF-Val.tar.gz -C LYF-Val
tar -xzf LightYourFace-160K/LYF-EditVal.tar.gz -C LYF-EditVal

Metadata and lighting control

Each JSONL record identifies its source portrait, lighting variant, component filename, and physical lighting parameters. LYF-EditVal records also provide the edited input filename. Filenames are relative to the directory where the archive is extracted.

Parameter	Unit	Model normalization
disk_diameter_px	px	diameter / 200 - 1
plane_distance_px	px	distance / 1000 - 1
lamp_offset_x	px	x / 3000
lamp_offset_y	px	y / 3000
half_peak_angle_deg	degrees	angle / 40 - 1
color_temp_K	K	temperature / 8000 - 1

The training loader samples an input scale r in [0.2, 0.4] and constructs the input and target in sRGB as r * source and r * source + 0.6 * component. This release preserves the lighting component so users can construct training pairs with the original source portraits. The code repository provides the model inference interface and lighting parameter normalization.

Citation

@inproceedings{gong2026lightface,
  title={Light Up Your Face: A Physically Consistent Dataset and Diffusion Model for Face Fill-Light Enhancement},
  author={Gong, Jue and Zhou, Zihan and Wang, Jingkai and Liu, Xiaohong and Zhang, Yulun and Yang, Xiaokang},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2026}
}