Towards Inclusive Biometrics: Synthetic Generation of Vitiligo Faces and Their Impact on Face Image Quality
A. Dörsch · A. Laguna Liang · C. Rathgeb · C. Busch
This is the official repository of the paper: Towards Inclusive Biometrics: Synthetic Generation of Vitiligo Faces and Their Impact on Face Image Quality
- HDA-Synthetic Vitiligo Faces (SynVF) dataset is being made available
This research work contributes to the development of more fair biometric systems by introducing HDA-Synthetic Vitiligo Faces (SynVF), the first publicly available synthetic face dataset incorporating real vitiligo patterns of varying severity applied to synthetic identities
| Dataset | Images | Description |
|---|---|---|
| SynVF | 1,803 | Synthetic vitiligo facial dataset containing 143 synthetic identities, each with a reference (non-vitiligo) image and multiple depigmentation variants derived from real vitiligo patterns. |
Note: As mentioned in our paper, not all vitiligo patterns were transferred to all individuals due to post-processing artifacts. More details about SynVF can be found in the associated paper
Recent studies indicate that Face Image Quality Assessment (FIQA) algorithms exhibit demographic biases, caused by physiological characteristics, particular skin pigmentation. In this work, we explore the potential impact of vitiligo, an autoimmune skin disease causing depigmentation patterns, on FIQA algorithms and associated biometric fairness. We introduce HDA-Synthetic Vitiligo Faces (SynVF), the first publicly available synthetic face dataset incorporating real vitiligo patterns of varying severity applied to synthetic identities. We utilize the Open Source Face Image Quality (OFIQ) framework and investigate the effect of vitiligo across pre-selected quality measures. We observe a concerning degradation of the Unified Quality Score (UQS) with increasing depigmentation, indicating that vitiligo tends to be misinterpreted as a quality defect rather than a depigmentation disease. By prompting additional synthetic vitiligo face images using OpenAI's 4o image generation model, we observed very low UQS scores on average, confirming that this bias trend generalizes beyond our proposed dataset. To mitigate the observed bias, we fine-tuned a Data-efficient Image Transformer (DeiT) for vitiligo classification and integrated its predicted probability into our proposed lightweight fairness-aware UQS adjustment that compensates for quality degradation. The classifier demonstrates strong generalization on real vitiligo faces, highlighting that synthetic data can be effectively used to support and improve the fairness of biometric systems.
Please contact André Dörsch (andre.doersch -at- h-da.de) to request access.
The database folder contains the following subfolders:
- base_images: Post-processed subset of selected synthetic reference images from HDA-SynChildFace
- mask(N)_upscale: Vitiligo mask (N) transferred to references from base_images
The vitiligo severity level associated to a vitiligo mask is listed below for the following masks included in SynVF (implementation and classification of severity levels can be found in the paper):
- Subtle depigmentation
mask3,mask5,mask6,mask10,mask22 - Noticeable depigmentation
mask1,mask4,mask9,mask13,mask16 - Prominent depigmentation
mask7,mask17,mask24, - Extensive depigmentation
mask12,mask18,mask23,mask25,mask26
This repository and associated model files are provided exclusively for academic research use.
If you use this repository and found it useful for your research, please consider citing this paper:
@inproceedings{Doersch-VitiligoFaces-WACV-2026,
Author = {A. D{\"o}rsch and A. Laguna Liang and C. Rathgeb and C. Busch},
Title = {Towards Inclusive Biometrics: Synthetic Generation of Vitiligo Faces and Their Impact on Face Image Quality},
Booktitle = {Proc. of the {IEEE/CVF} Winter Conf. on Applications of Computer Vision {(WACV)} Workshops},
Month = {March},
Year = {2026},
Pages = {915-923},
}
As identities from SynVF are based on HDA-SynChildFace, please further cite:
@article{Falkenberg-HDASynChildFaces-Frontiers-2024,
title = {Child Face Recognition at Scale: Synthetic Data Generation and Performance Benchmark},
author = {M. Falkenberg and A. B. Ottsen and M. Ibsen and C. Rathgeb},
year = {2024},
journal = {Frontiers in Signal Processing}
}