You Don't Need All That Attention: Surgical Memorization Mitigation in Text-to-Image Diffusion Models
Researchers have developed a new framework called GUARD to mitigate memorization in text-to-image diffusion models. This method adjusts the image denoising process during inference to steer generations away from specific training data while maintaining prompt alignment. GUARD's approach involves selectively attenuating cross-attention based on statistical analysis, offering a robust, per-prompt solution that improves memorization mitigation without sacrificing image quality. AI
IMPACT Offers a novel approach to address privacy and copyright concerns in generative AI by preventing verbatim reproduction of training data.