Layers for concepts

#2
by Ertugrul - opened

Hey, targeting single transformer blocks works well for characters or styles, but when it comes to concepts (especially those unknown to pretraining, I guess), do you have any clues about what modules to target when it comes to more complex concepts? Because the current setup breaks everything for me and just overtrains the stuff outside the concept.

Great work anyways!

9 and 25 usually work well for concepts, don't use a very high lr, and start with a small amount of steps

Ertugrul changed discussion status to closed

Sign up or log in to comment