Evaluation of Concept Erasing for Artistic Styles in Diffusion Models

Artists have been raising the alarm about unfair and opaque practices of generative AI models for nearly two years. In a technological attempt to answer these concerns, concept erasing has promised to give artists more agency in determining whether new generative AI models get to clone their signature artistic styles. The idea behind this method is simple: take an already trained AI model, choose an artist's style to remove from the model, and apply minimal changes to suppress that concept. Simple integration into existing pipelines makes it a particularly appealing solution. However, the efficacy of this method has so far been examined for only one artist's style at a time. This paper, thus, examines concept erasing at scale, simulating the hypothetical application of this technology for simultaneously erasing many artists. It finds that the method does not reliably erase the concepts in cases with more than one artist's style, highlighting a gap in the technology and the literature around it.