Closed-Form Concept Erasure via Double Projections

TL;DR

This paper introduces a simple, analytical linear transformation method for concept erasure in generative models, which is efficient, interpretable, and matches or exceeds existing techniques.

Contribution

It proposes a closed-form, training-free approach for concept removal that is deterministic, geometrically interpretable, and preserves non-target concepts better.

Findings

Matches or surpasses state-of-the-art erasure performance

Preserves non-target concepts more faithfully

Requires only a few seconds to apply

Abstract

While modern generative models such as diffusion-based architectures have enabled impressive creative capabilities, they also raise important safety and ethical risks. These concerns have led to growing interest in concept erasure, the process of removing unwanted concepts from model representations. Existing approaches often achieve strong erasure performance but rely on iterative optimization and may inadvertently distort unrelated concepts. In this work, we present a simple yet principled alternative: a linear transformation framework that achieves concept erasure analytically, without any training. Our method adapts a pretrained model through two sequential, closed-form steps: first, computing a proxy projection of the target concept, and second, applying a constrained transformation within the left null space of known concept directions. This design yields a deterministic and geometrically interpretable procedure for safe, efficient, and theory-grounded concept removal. Across a wide range of experiments, including object and style erasure on multiple Stable Diffusion variants and the flow-matching model (FLUX), our approach matches or surpasses the performance of state-of-the-art methods while preserving non-target concepts more faithfully. Requiring only a few seconds to apply, it offers a lightweight and drop-in tool for controlled model editing, advancing the goal of safer and more responsible generative models.