Search papers, labs, and topics across Lattice.
OrthoEraser addresses the problem of collateral damage in concept erasure for text-to-image models by using sparse autoencoders (SAE) to disentangle sensitive and benign features. It introduces an analytical gradient orthogonalization strategy that projects erasure vectors onto the null space of coupled neurons, thereby decoupling sensitive concepts from critical benign subspaces. Experiments demonstrate that OrthoEraser achieves higher erasure precision and preserves the generative manifold better than state-of-the-art baselines.
Concept erasure in text-to-image models no longer needs to be a blunt instrument: OrthoEraser precisely removes harmful content while preserving image quality by analytically orthogonalizing the erasure process.
Text-to-image (T2I) models face significant safety risks from adversarial induction, yet current concept erasure methods often cause collateral damage to benign attributes when suppressing selected neurons entirely. This occurs because sensitive and benign semantics exhibit non-orthogonal superposition, sharing activation subspaces where their respective vectors are inherently entangled. To address this issue, we propose OrthoEraser, which leverages sparse autoencoders (SAE) to achieve high-resolution feature disentanglement and subsequently redefines erasure as an analytical orthogonalization projection that preserves the benign manifold's invariance. OrthoEraser first employs SAE to decompose dense activations and segregate sensitive neurons. It then uses coupled neuron detection to identify non-sensitive features vulnerable to intervention. The key novelty lies in an analytical gradient orthogonalization strategy that projects erasure vectors onto the null space of the coupled neurons. This orthogonally decouples the sensitive concepts from the identified critical benign subspace, effectively preserving non-sensitive semantics. Experimental results on safety demonstrate that OrthoEraser achieves high erasure precision, effectively removing harmful content while preserving the integrity of the generative manifold, and significantly outperforming SOTA baselines. This paper contains results of unsafe models.
