V3: Trained it with twice the dataset, from 50 to 100, and used a lower learning rate. Added a tag to help the model out. It should be able to pretty accurately and flexibly work in various poses. A huge improvement over V2. V2: trained it with a few more repetitions/epochs. It should be better able to replicate the look with prompting, but it's somewhat inflexible from other perspectives. I'll need to expand the dataset a bit to fix that, but this should be an improvement over V1.V1: So this is my first attempt at a character LORA. From what I can tell, it's pretty undertrained; it can replicate things to a decent extent, but you're going to need to increase the strength over 1 to around 1.2 or so. It's been trained on joycaption captions, so you will also need to mostly describe what you actually want if you want it to replicate the character well at the moment.

Ino Yamanaka

Tried retagging the dataset and trained it for longer; this model should give more accurate results when prompted, but I've found that it's lacking a bit when it comes to flexibility as a result. I'll need to expand the dataset a bit but overall I think it's easier to work with than v1.

Files with SHA256: 08435a20b9e65c5bd9353db927267b97634d5778745f846d04d5370756fe2999