2 Words about big model:
1) you can click on "Training Images Download", there non training images, it's wildcard file with trigger words;
2) 21k steps x4 batch size(80k steps is scare if you don't know what it have like 16 subsets and small learning rates) -> don't use it with power 1.0, use it with 0.7 or even less. Yes it not so much flexible, it more about character Details.
Train LoRa is time and resource expensive, you can support my future works here:
it would be great if you could show some of the results of your work as a post/review
//info will be updated later(on monday i think, after i start to upload other characters)