Minki Hong

Minki Hong

My research interests include generative AI, with a particular focus on social norms, bias mitigation, fairness, and inclusivity.

Evaluating Demographic Misrepresentation in Image-to-Image Portrait Editing

Published in Under Review, 2026

We are excited to share our latest research on demographic-conditioned failures in instruction-guided image-to-image (I2I) portrait editing. While demographic bias in text-to-image (T2I) generation is well studied, our work is the first to systematically examine and formalize how identical edit instructions can yield systematically different outcomes across subject demographics in open-weight I2I editors.

I2I Bias Evaluation Framework
Figure 1. Comparison of I2I portrait editing across diverse demographics. Given identity-agnostic prompts (e.g., "senior executive," "senior military strategist"), baseline methods often produce stereotype replacements—changing race or gender toward majority demographics (red boxes). Our prompt-level mitigation preserves the original identity across all demographic groups (green boxes), demonstrating that identity preservation failures disproportionately affect minority subjects.

In this work, we: 1️⃣ Identify Two Failure Modes: We define and characterize Soft Erasure, where requested edits are silently weakened or ignored, and Stereotype Replacement, where edits introduce unrequested, stereotype-consistent attributes (e.g., skin lightening, gender change).

2️⃣ Introduce a Controlled Benchmark: We create a systematic evaluation framework using 84 factorially sampled FairFace portraits spanning race, gender, and age, paired with diagnostic prompts—yielding 5,640 edited images across three open-weight I2I editors.

3️⃣ Propose Prompt-Level Mitigation: We demonstrate that a prompt-level identity-preserving control can substantially reduce demographic change for minority groups without model updates, revealing asymmetric identity priors in current editors.

Our findings establish identity preservation as a central and demographically uneven failure mode in I2I editing, motivating the development of demographic-robust editing systems.