>>93803707>>93803997Retrained with the following settings:
>--learning_rate 1.0 --text_encoder_lr 0.15 --max_train_steps=800>--optimizer_type "prodigyplus.prodigy_plus_schedulefree.ProdigyPlusScheduleFree" --optimizer_args use_orthograd=True use_speed=True split_groups_mean=False factored=False weight_decay=0.1>--max_grad_norm=0 --scale_weight_norms=2 Looks cleaner. Something like this is probably a good starting point if someone wants to try it.
