Klein training doesn't seem to work well

[pflky]

Tried different schedulers, learning rates, alpha, rank, etc.....can't get any likeness at even 100 steps per training image, even went as far as 200 steps. I'm training on 9B base, to use the LoRA on the distilled 9B model. Either Klein has some very specific requirements in training that I'm missing, or there's something wrong with how its implemented.

To add, the concept I'm training on is a different variety of the same type of object, not a person or style. I know this type of training can be a little more difficult, since it's not about exactly recreating a specific subject, but about creating a variety of a specific object.

I've trained this same dataset on previous models like SDXL and Z image just fine.

[dxqb]

join our discord. in #training-results there are multiple people posting 9B results and they seem to be happy with them. maybe they can help