RISE: Red-teaming via Iterative Strategy Evolution for Modern Text-to-Image Models
This work finds that judges widely used in prior T2I red-teaming work are unreliable under vague unsafe-content targets, and defines strict category-specific success criteria and calibrate strong VLM judges against human labels to solve the exploration problem.
Dmitrii Kharlapenko, Sergei Bratchikov, Konstantin Korolev et al.
· 0 citations