When image-generating AI models memorize and reproduce training examples, they risk leaking private or copyrighted data. Most existing fixes weaken the AI's understanding of the prompt too broadly, hurting the quality of the output.
Researchers discovered that this memorization is caused by just a tiny number of attention heads, components that decide which parts of the prompt to focus on, in the early layers of the model. By measuring which heads exert overly strong control, they were able to pinpoint the memorization problem.
The team developed a targeted method that only suppresses the problematic heads, leaving the rest of the model's prompt understanding intact. This approach reduces memorization while preserving image quality, achieving a better balance than previous methods.
