A small, pointless experiment: training only the first text encoder, only the second, both encoders, only unet, unet+embeddings, unet+both text encoders+reg images.
All loras trained on base ponyXL
Conclusions?
- for pony - train unet-only or te2-only. te2 trains better than te1
- for realistic mixes - train unet
nude example:
https://files.catbox.moe/ko15z4.jpgblacked:
https://files.catbox.moe/obtx2r.jpg2girl:
https://files.catbox.moe/iwelsx.jpg 