Back to feed
arXiv cs.AI·

Orthogonal Concept Erasure for Diffusion Models

Signal
78
Hype
15
In three linesOrthogonal Concept Erasure (OCE) proposes an editing method to remove undesired concepts from diffusion models using multiplicative orthogonal transformations. Unlike existing additive approaches, OCE preserves neuron magnitude and angular geometry while precisely erasing concepts. The approach erases up to 100 concepts in 4.3 seconds.
Read source
Your take?
PapersAI safetyAlignment

Summary generated by Claude — human-verified