Perturbed-Attention Guidance (PAG) stands as a new sampling guidance method designed to enhance the quality of samples without additional training or external modules. PAG operates by strategically perturbing the self-attention mechanism within the model, allowing for the correction of noisy components while preserving locally aligned structures in the generated samples. This perturbation is carefully applied to specific layers of the model, ensuring an improvement in sample quality without significant deviations from the original samples.

The implementation of PAG involves configuring various parameters such as guidance scales and specific layers for perturbed self-attention. Integrated with existing models like ADM, Stable Diffusion, and text-to-3D generation, PAG showcases enhancements in sample quality through human evaluations and downstream tasks.