The world is being quietly rearranged by people who write very long documents.


The title they went with Forget Many, Forget Right: Scalable and Precise Concept Unlearning in Diffusion Models Noisy translates that to

AI image models can now forget copyrighted styles and artists — without forgetting how to paint


Researchers built a faster, simpler method to remove specific concepts from AI image generators without degrading their overall ability to create images. The technique works at scale — removing many concepts at once — and needs no extra data or models to do it, making it practical for companies to implement copyright protections.
Until now, removing a concept from a diffusion model (say, the style of a particular artist) either broke other similar concepts or required expensive workarounds. This method solves both problems by identifying exactly which parameters in the model control a specific concept and adjusting only those. The result is a technical answer to a real legal problem: if an image model trained on an artist's work without permission, rights holders now have a plausible way to demand the company remove it — without retraining the entire model from scratch.
Watch whether any major image model company (Midjourney, Stability AI, Adobe) actually deploys this technique in response to copyright takedown requests, and whether it holds up when artists test whether their style actually disappeared.

If you insist
Read the original →