IMPROVING DIFFUSION MODELS WITH PERTURBED-ATTENTION

Summary from the TLDR AI newsletter

Perturbed-Attention Guidance (PAG) is an approach to enhancing the quality of images generated by diffusion models without additional training or external modules. By creatively manipulating self-attention mechanisms within the model, PAG significantly improves the structure and fidelity of both unconditional and conditional samples.