https://arxiv.org/abs/2210.00939
Demos:
For SDXL model, try to tune down everything if you are getting blurry/overfried/distorted results (or lower base resolution), SDXL is more sensitive than SD1.5
Bilinear interpolation: Attention mask is interpolated using bilinear method, resulting sharper image
Attention target: Choose the block Attention mask would apply to, dynamic means depending on noise sigma value (maybe broken on SDXL, only use dynamic on SD1.5)
Base resolution: Change attention resolution scaling, set it to 0 to ignore this setting
Smooth Vectors: inspired by Smooth Energy Guidance, we smoothed and applied median blur on QKV vectors first to make latents cleaner, set it to 0 to disable it
Perturbed Vectors: inspired by Perturbed-Attention Guidance, keep it at low scale (around 0.05~0.2)