Data Bias Mitigation and Evaluation Framework for Diffusion-based Generative Face Models

Hao Yu, Margrit Betke, Sarah Adel Bargal; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 300-310

Recent advancements in generative face models have facilitated the creation of high-quality face images. However, these models often suffer from a bias towards lighter-skinned faces due to the predominance of such faces in existing face datasets. To address this issue, we propose a comprehensive framework for bias mitigation and evaluation for generative face models. Our framework integrates underrepresented facial data, including large-scale datasets of Black faces, and provides standardized procedures for evaluating fairness. We demonstrate the effectiveness of the proposed framework by incorporating a curated high-quality face dataset of Black individuals, KenyanFaceHQ, into diffusion-based face generation models. Results show that including KenyanFaceHQ significantly mitigates racial and skin-tone bias in the trained models, while preserving the high quality of the generated images. Furthermore, we show that our framework can also be applied to gender classification, and leads to accuracy improvements ranging from 0.6 to 8.6 percentage points across various ethnic and gender subgroups. Our results highlight the framework's effectiveness in evaluating and reducing bias of both discriminative and generative face models.