Concept Unlearning via Cross-Attention Activation Projection for Diffusion Models
PURE is a method to erase target concepts from text-to-image diffusion models without retraining by editing cross-attention weights. It achieves this by building 'forget' and 'retain' bases from model activations during denoising, rather than text embeddings. This approach is superior because activation features better capture the concept evidence used in image generation.