Overall
This Lora is for the model from Ghost Rule. The training dataset comes from Project Diva Future Tone. The base model is NAI. Showcases indicates a lora named "test1", that is GhostRule_v2.0, I have renamed it.
Trigger Word
There is no trigger word for this model, but recommanded to use with
twintails,very long hair,(black dress:1.1),neck ribbon,elbow gloves, black gloves,black thighhighs,blue hair,purple hair,purple eyes,
(In version 1.1 those words are not necessary, since v1.1 is overfitting.)
Weight
0.6-0.8 depends on which checkpoint you used.
About Version 2.0
In version 2.0 I use SVD script to avoid the lora overfit training data 3D style from project diva future tone, now if you use 0.6 weight this lora could performs checkpoint orignal style very well. This lora also support the other characters, there is an example showcase :) You could also only use the clothes related tag.
About Version 1.1
Since version 1 training data has messy background, I did preprocessing with those and delete them, then I use the clear data trained 1.1, the model has more accrucy but it overfits the training data (Maybe the messy background did some job to avoid overfitting LOL). I trying to slove and that becomes the realeased model. If you like the Project Diva Future Tone model then go ahead to use version 1.1, it has better background adjustment and stylistic reproduction however it could not show the checkpoint style efficiently.
Description
Clean the training dataset, version 1 would produce messy background, now the problem has been fixed. V 1.1 has a drawback which is a little bit overfitting training data ;(
FAQ
Comments (2)
You should change the order of versions, currently the older v1.0 is displayed as default.
Nah becasue 1.1 overfits the training data, if you need the style from Diva FT you could use v1.1, otherwise that is not recommanded. I am still trying to train a model without overfitting. After that I will release a new version and rearrange the order.