Another experiment using different settings and dataset’s proportions.
25 renders of Ashe model by Anonymous Frankenstein. However, can generate Ashe with normal proportions too.
2 official artworks/concept arts, 13 unnoficial ones focusing Ashe’s face.
No NSFW stuff was used during the training, based on NAI leak model.
//
1) As long as “detailed face” tag is in your promt, it shouldn’t create Lovecraftian abominations.
2) It was trained using “swimsuit” tag in most cases, however, it does slightly better without it
3) It “can” work with NAI (aka Animefull; not recommended to use with this lora), other anime-looking models, 3D ones and realistic models (would look like an unreal render most of the times). However, semi-realistic ones like JAM are still the way to go.
4) High-res fix isn’t needed this time
5) Was trained on clip skip 2, works well on clip skip 1 as well
6) “gigantic breasts” tag would cause a different outfit and, well, you can probably guess what else. Not using it somehow causes a classic-looking outfit (was present like 2 times out of 40 images). No idea how it works but it is what it is