In order to understand block weights, we need to understand Stable Diffusion U-Net architecture first.
U-Net
If you’ve ever tried to research the topic, you definitely stumbled upon this U-shaped diagram:
U-Net as seen from a 100 miles away
An input image starts it’s way at the top left and after a transforming journey over a U-path is reborn at the top right.
I read a ton of articles on the subject in the last few days and the most accessible one for those of us who get easily discouraged by complex math is this one:
You can’t spell Diffusion without U
Understanding the core building blocks of Stable Diffusion
If you are not entirely sure what U-Net is, I encourage you to read it in full. It’s relatively short and totally worth it from start to finish.
I especially liked how
used the telephone game as an example to describe the way U-Net blocks communicate with each other.