Character LORA model for Tachikawa Mimi in Digimon, trained on NAI.

There are few fanarts of her so I decided to train it, and the results are surprisingly good.

The dataset and training info is provided and you can download them, if you want to train yourself :)

Update: After some experimenting, I found a good workflow for training character who don't have a lot of good dataset or the datasets are of old style, it's a bit like Polymerase Chain Reaction (PCR):

1. Trained on 50 original dataset all tagged with 1girl, xxx (with or without class image) on NAI for an overfitting LORA model (called beta model).

2. The result LORA model doesn't perform good on original model, but does excellent job on merge models, and the features of the character are maintained. Therefore use prompt 1girl, xxx to generate a lot of pics and select 50 good ones, called augmented images. These can be treated as old characters drawn in modern style. You could also train from them like above to get a new LORA model, but loses some generality (called gamma model).

3. Put both 50 original dataset and 50 augmented images to a new dataset and tag all of them with wd-tagger, and train them with or without class image. The resulting LORA model works good on original and other models, and it keeps some generality (called delta model).

4. You can repeat the process so that the dataset is augmented from 50 to 100, 200, 400....，so even if we have few images with old style, we can still produce good results and keep some generality.