Leveraging diffusion and Flow Matching Models for demographic bias mitigation of facial attribute classifiers

Download PDF
Open Webpage

Sreeraj Ramachandran, Ajita Rattani

Published: 01 Jan 2025, Last Modified: 12 Nov 2025Neurocomputing 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•Introduced a classification diffusion model (CDM) for bias mitigation and fairness.•Proposes uncertainty-based test-time rejection using outputs of diffusion model.•Proposed a classification conditional flow-matching (CCFM) model for faster inference.•Developed 2nd-order and single-step solvers for faster CCFM evaluation.•Evaluated on multiple facial attribute datasets, mitigating gender and race bias.