V2 is out - it fixes many stability and consistency issues from V1. It also improves the quality, with more faithful art style to the anime. The model is good enough to exit the prototype stage.

There are still two issues with this model:

1. The model doesn't learn her hair color consistently, so it is advised to include the color in your prompt - e.g. "has light blue hair".

2. The model doesn't learn the interface headset - the white dots in Rei's head. That is because many training data are of Rei in casual clothes, and didn't include the headset.

I might revisit the dataset to see if I can address these two issues.

The trigger word is ppayanamirei, to avoid picking up the model bias of Ayanami Rei.