Tipo
LORA
Data di pubblicazione
2024-07-15
Modello Base
Pony
Parola Chiave
source_anime
Copia
Introduzione alla versione
Retrained against my Tweaked Base Pony XL checkpoint (to allow for some subtle quality improvements) on a dataset with around 200 more images than V3's. The whole dataset was also recaptioned from the ground up using the same combined "Florence 2 Large detailed captions + WD-VIT-3 Booru tags" approach I've been using elsewhere recently.
This version no longer uses a unique trigger, instead being trained directly against 'source_anime'. Using '2d' in the positive and '3d, cgi' in the negative will also tend to strengthen the effect. Recommended strength is around 0.7 - 0.8.
