ContextBias Benchmark Evaluates Stereotypical Bias in Text-to-Image Models
August 29, 2026
ContextBias introduces ContextBench, a framework comprising 92 roles and 1,656 prompts to measure how visual stereotypes persist when professional roles are moved into semantically unrelated contexts. The study evaluates four state-of-the-art models across 66,240 generated images.
HOW THIS AFFECTS YOU
●
researcherYou can use this benchmark to isolate whether model bias is tied to specific prompts or inherent concept associations.
●
policyThis provides a controlled method for auditing generative models for societal bias.