V2 is out! Hikari is also out of prototype phase, as this new model is trained with network dim=4 and offers much better stability and consistency. Just compare the showcase between v1 and v2 and it should be pretty obvious.

One thing to note is that I trained this model using 512x512 resolution. To my surprise, I have not noticed any negative effect to the model outcome. For those of you who are figuring out the ideal resolution for Flux anime LoRAs, this could be a data point.