Our work on understanding and characterizing concept erasure methods in text-to-image models has been accepted at NeurIPS 2025.