>>62713642Yeah, HCP-Diffusion is very user-unfriendly. I wanted to try pivotal tuning method too, and i also gave up.
You can instead try
https://github.com/Nerogar/OneTrainer for masked training.
It's easy to use, it also supports
insanely slow but interesting RL method from
https://align-prop.github.io/Only issue with this trainer is maximum token lenght is fixed at 75, long tags will be cut off