This is a redo with the current kohya_ss (v22.6.0 at time of writing) since the settings I used with the old version don't produce the same result. Training data and notes are included for those that are interested. The training run used gradient accumulation = 4, so the 500 steps are similar to 2000 steps without GA.
This is a redo with the current kohya_ss (v22.6.0 at time of writing) since the settings I used with the old version don't produce the same result. Training data and notes are included for those that are interested. The training run used gradient accumulation = 4, so the 500 steps are similar to 2000 steps without GA.