Version 2V2 was trained on more than 30x as many images as V1, so its understanding of Hu Tao's appearance is much better. Unfortunately it seems that training Flux models on Danbooru tags leads to some weird prompting issues, so I will look into using more natural tagging for V3.