Prototype-Guided Concept Erasure in Diffusion Models

TL;DR

This paper introduces a prototype-guided method for erasing broad concepts in diffusion models by leveraging embedding clustering to identify and remove concept representations, improving reliability and safety in image generation.

Contribution

It proposes a novel approach using embedding clustering to identify concept prototypes for more effective broad concept erasure in diffusion models.

Findings

Enhanced erasure of broad concepts like 'sexual' or 'violent'

Preserved image quality after concept removal

Outperformed existing methods on multiple benchmarks

Abstract

Concept erasure is extensively utilized in image generation to prevent text-to-image models from generating undesired content. Existing methods can effectively erase narrow concepts that are specific and concrete, such as distinct intellectual properties (e.g. Pikachu) or recognizable characters (e.g. Elon Musk). However, their performance degrades on broad concepts such as ``sexual'' or ``violent'', whose wide scope and multi-faceted nature make them difficult to erase reliably. To overcome this limitation, we exploit the model's intrinsic embedding geometry to identify latent embeddings that encode a given concept. By clustering these embeddings, we derive a set of concept prototypes that summarize the model's internal representations of the concept, and employ them as negative conditioning signals during inference to achieve precise and reliable erasure. Extensive experiments across multiple benchmarks show that our approach achieves substantially more reliable removal of broad concepts while preserving overall image quality, marking a step towards safer and more controllable image generation.