Trigger word is G3NSHIN.
- V3 -
Pretty much the same as V2, but with a bigger dataset (more characters, different poses, objects, backgrounds, animals, NPCs, etc).
I've included 1~3 images of each female playable character, but unless your model is already familiar with the characters, you won't be able to generate them accurately without a character lora.
Adding "3D" and increasing strength will help with the style, but may make poses more stiff and background details weird.
In most of the example pictures I've used strength 1.2 and face detailer.
- V2 -
Remade the LORA with a completely new dataset. This new dataset includes way more locations, characters, angles, and poses.
This time the only trigger is "g3nshin".
If the style isn't strong enough I recommend bumping the strength up to something like 1.1 or 1.2.
- V1 -
I wanted to see what it would look like if I trained a LoRA with a bunch of Genshin screenshots. The results are... OK, I guess.
All screenshots were taken in Natlan, with the game on max settings, and a somewhat big variety of human characters/NPCs in different poses.
A few female NPCs showed up more than once in the dataset so I tagged them as "NatlanEnemy" and "NatlanNPC". I don't recommend trying to summon them with those tags since they barely show up in the dataset, but if you want to try it anyway I've included a few pictures below of what they look like.