I wanted to see how a LoRA trained using 3D model images would fair so here is the result. Red bangaa from FFXII.
V2.0 Trying some new stuff. Figured out how to get the model in Blender so I added some more images and a visor version. Tried out prodigy optimizer and cosine w/restart. V1.0 still has a certain charm to it, but I think V2.0 has a better accuracy and pose variation. .6 to .8 variation seemed to get best results.
Main Trigger: bangaa
Optional: bottomwear, cuff \(restraint\),
Optional V2.0: visor
V1.0trained on yiffymix with the recommended VAE
V2.0 no VAE
Yiffymix 3.4, no VAE, Additional images from Blender, Prodigy, cosine w/restarts, bangaa with visor