README.md

training:

dataset = images-damian2 last 2 text encoder layers unfrozen 30 epochs at 2e-6, then 20 epochs at 1e-6 cond_dropout= default