Representation Steering via Null-space Alignment for Concept Unlearning in Diffusion Models
Abstract
The rapid deployment of Text-to-Image (T2I) diffusion models has raised urgent concerns about the generation of unsafe, copyrighted, and offensive content. Recently, steering-based concept unlearning methods have emerged as an efficient alternative to retraining, suppressing undesired concepts by injecting learned direction vectors into model representations. Despite their efficiency, these approaches rely on unconstrained directional updates that indiscriminately alter model behavior, often leading to over-suppression, cross-concept interference, and degradation of non-target generation quality. To address this issue, we propose RESET, a theoretically grounded framework for Representation steering via null-Space alignment in diffusion models. Moving beyond unconstrained heuristic steering, we steer representations under a null-space constraint to avoid interference with retained concepts. Specifically, we construct a preservation subspace induced by benign concept representations and restrict parameter updates to its orthogonal null space, thereby preventing interference with retained concepts. Within this constrained space, we derive a closed-form solution that selectively suppresses unsafe concepts while preserving general generation capability. Extensive experiments across object, style and identity benchmarks show that RESET consistently outperforms state-of-the-art methods, achieving more precise concept removal, improved robustness against red-teaming attacks, and superior preservation of unrelated content fidelity.