Sorry it took so long to answer. I don't want to make any promises I can't keep.

Making a dataset for Shinozuka Yuuji model will require a lot of cropping and retouching, I will try to make it a bit later, right now I still have many ongoing projects.

It's called animefull-final-pruned-fp16. You can find it on huggingface.

For this one I've used dim and alpha both at 128 and unet at 0.00005

Repeats and epochs - you'll it have to decide for yourself. They won't affect the end result but rather create more points where you'll be able to see if your LoRa is undercooked, fine or overcooked. I'm not a pro by any means but I recommend to use 3-4 repeats if you're working with 100 images dataset, or 5-6 for 60-70. What's more important is the steps. From my experience, styles usually solidify at 800-1000 steps and after that point they keep adding features. So train it for around 3000 steps and choose the best result.

Anyway, your dataset will always be the most important part of the training. Weed out textboxes and signatures, crop blank spaces, don't use variations of the same image, prioritize quality over quantity, etc.

Hope I was able to help.