Kohaku-Delta-Alpha

This is an experimental, academical model, which was trained from kohaku-delta-rev-1, with LoKr method (LyCORIS), with 6M images in total. 8x3090x1.5 months, 8640 3090 hour = about $4000-worth GPU hours were spent for this model training.

The models is not trained for any aesthetic themes, nor specific default style or special datasets.

Some special techniques were introduced, however, I'd call this as "Dataset is all you need" and "Training compute is also you need".

I'm collecting budget for H100 server rental / equipment : ko-fi