Brief

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

Researchers have developed a new framework to evaluate cultural bias in generative image models, focusing on both text-to-image generation and image-to-image editing. Their study, conducted across six countries and using a detailed schema, found that models often default to Global-North, modern depictions and that iterative editing can degrade cultural accuracy. The models tend to apply superficial changes rather than contextually appropriate ones, highlighting the unreliability of culture-sensitive edits in current systems. The researchers have released their data, prompts, and evaluation protocols to promote reproducibility and further research. AI