Well, in the case of this LoRA, I gathered my dataset of around 100 images and then trained it with the default CivitAI LoRA trainer. Of course, this is not a truly optimal way of making a style LoRA because Klein is really stubborn in terms of learning something that isn't realism. The fact is, the editing feature would work best if we had the possibility of using a target and control dataset: one that shows the style you want to achieve and another that represents realism (before and after just like for training Flux Kontext). That would work leagues better than what I'm doing here, but since I cannot do that, and training on AI Toolkit becomes longer or more expensive, I chose this more aggressive approach for training Klein, which consists of high repeats and smaller, less descriptive captions.

Even then, I'm still playing around with it, so the moment I get my results back from another style I'm currently training, I will share my findings here. Hope this helps you.

Also i forgot to say i will be releasing an article if all goes as planned that way people can try a tweak some of the settings i use and maybe even get better results.