NovelAI Model tags (from finetune.py)
Welp, the NovelAI Model got leaked. Had some experiments with it and can say it's interesting.
There exists a finetune.py
script in the stable-diffusion-private.tar.zst
file which I checked and made some observations on the tag processing.
The following notes are meant to be my observations and may not represent the final model expectations of prompt inputs.
Notes
All ,
are removed, So things such as: Test, test2
are transformed to Test test2
.
All _
are replaced with "" (space key) as well. So
testa_testb
are transformed to testa testb
Artist tags according to the finetune.py script exists, so you can try using those.
The tag: "nsfw" is taken from rating:questionable
& rating:explicit
. Enables those kinds of Lewd artworks.
there is no rating
tags used.
"How about brackets?" Well those are included. So like "STest Tag (Something!)
" are kept as is.
Japanese tags are also kept. so stuff like futanari
should be there too.
These tags: bad anatomy
, bad hands
, 6+girls
and all the related numbering tags, 3boys
as well as artist:
, copyright:
, character:
are kept.
There appears to be no skipping of images. So these tags should exist:
"comic", "panels", "everyone", "sample watermark", "text focus", "tagme", "jpeg artifacts"
Cases are kept. So all lower case.
Tags not used for training
Do not use the following tags since they don't exist in the set:
absurdres
highres
translation request
translated
commentary
commentary request
commentary typo
character request
bad id
bad link
bad pixiv id
bad twitter id
bad tumblr id
bad deviantart id
bad nicoseiga id
md5 mismatch
cosplay request
artist request
wide image
author request
artist name
Misc
Feel free to put in the FAQ.