Graphs of merged Weight?
Hello. I was wondering about something like the graph of merge percentages (Converted weights) on the page. Can you tell me how this was output?
Thank you for asking.
The coverted weights are not actually "converted" values. They are actually raw values calculated on how much I wanted to reflect each model with each layer, and I converted the values so that I can use them in Merge Block Weighted tab on SD-webui(the values you see in Formula box).
How I calculated the values, if you ask, is somewhat complex. I basically read a bunch of articles on "NovelAI 5ch wiki" to try and figure out which layer affects each attribute(outlines, brightness/saturation, facial details, etc.) of the generated images. And based on what I've found, I just made up a 26-row attribute-to-layer conversion matrix, scored attributes of each model(on a scale of 1 to 5) by comparing sample images, and converted the scores to layer weights using the matrix. The whole process was a little bit too messy to give you some more details as now. I'll try to make some documentations of what I've tried if you ask.
How I made the graph images, is rather simple; I just Microsoft Excel-ed them.
Thank you. This is very helpful. If possible, it would be more understandable if you could share the detailed creation process.
(I know it is a lot of work to re-edit a memo in a way that others can understand it. I will wait patiently.)