MATH · IN · MODELS

OCE erases diffusion concepts via a shared orthogonal rotation preserving norms and angles

measured in 1 paper

Sun et al. edit a diffusion model's cross-attention key/value weights in place via a single shared orthogonal matrix P solved in closed form from an orthogonal Procrustes problem [sun-etal-2026-orthogonal-concept-erasure-for-diffusion-models] Because P is orthogonal, neuron magnitudes and pairwise inter-neuron angles are preserved exactly, an algebraic guarantee rather than a measured statistic [sun-etal-2026-orthogonal-concept-erasure-for-diffusion-models] An ablation shows the design matters: additive scaling has negligible effect, layer-wise shared rotation gives a clean semantic shift, and neuron-wise rotation degrades image quality [sun-etal-2026-orthogonal-concept-erasure-for-diffusion-models] It erases object classes, styles, up to 100 celebrities simultaneously, and NSFW content on SD v1.4, transferring zero-shot to the DiT-architecture FLUX.1 [sun-etal-2026-orthogonal-concept-erasure-for-diffusion-models]

Context

a single shared orthogonal (rotation) matrix as a minimal-distortion weight-editing intervention, exact algebraic preservation of neuron norms and pairwise inter-neuron angles under an orthogonal edit, orthogonal-complement subspace projection for multi-concept erasure

Papers

Orthogonal Concept Erasure for Diffusion Models — Sun, Yuhao, Yu, Lingyun, Xu, Haoxiang, Miao, Fengyuan, Xu, Zhuoer, Xie, Hongtao2026 · arXiv:2605.28902