Toward Trustworthy Portrait Editing: Evaluation of Demographic Misrepresentation in I2I Models
A new research paper highlights significant demographic misrepresentation issues in instruction-guided image-to-image (I2I) editing models. The study identifies two failure modes: soft erasure of requested edits and stereotype replacement with unrequested demographic attributes. Across 5,040 edited portraits, the research found pervasive and demographically uneven identity preservation failures, with a notable tendency for outputs to exhibit skin lightening, particularly affecting Indian and Black source portraits. AI
IMPACT Highlights critical trustworthiness failures in generative editing systems, potentially reinforcing representational disparities and shaping AI-mediated self-representation.