Danbooru/e621 autocomplete tag lists incl. aliases (+ Krita AI support)版本1st Dec 2024 (ID: 1116599)

These tag list(s) contain tags + their aliases (active & deleted + chars/artists/franchise tags) from Danbooru and/or e621 and are for use with sd-webui-tagcomplete extension and pythongosssss/ComfyUI-Custom-Scripts autocomplete feature as well as Krita AI Diffusion and Stability Matrix.

There are 3 main versions. All versions have tags with less than 50 posts associated with them deleted from the lists because those usually aren't in datasets.

1. Danbooru (DBR) - contains only Danbooru tags (best for anime models and Pony/Illustrious models without NoobAI knowledge, unless you don't plan on using e6 tags)

2. e621 (E6) - contains only e621 tags (best for e6 oriented models/usage)

3. DBRE6 - contains a merge of both the Danbooru tags and e621 tags (best for NoobAI and Pony based models and if you plan on using e621 stuff)

There's also Krita AI Diffusion compatible tag lists inside the corresponding folder.

Note: In e621, e.g. the tag long_hair is aliased to very_long_hair, while on Danbooru, they are separate tags. This may cause suggestions to show both long_hair and very_long_hair when using very_long_hair and combined list. This shouldn't really cause any issues but it could be a tiny bit confusing or annoying if this somehow bother you.
If you are using Krita AI Diffusion you won't need the merged list since it lets you select both separate lists to be used for autocomplete.
If you are using the sd-webui-tagcomplete extension then you don't need this because it comes with the 22.12 lists including the merged one by default now

The script with which I made these tag lists is open source on my GitHub. The repo also contains guaranteed automatic updates happening every 2 months in case I don't upload manually. (not working rn because me stupido and the code is slightly out of date because im too lazy to update it but still works for whats uploaded here

Artist/character/series name wildcard functionality coming soon (probably)

Small note: according to the HF repo, NoobAI EPS AND Vpred 1.0 had a cutoff roughly before 23rd Oct 2024 meaning if you are using NoobAI eps/vpred 1.0 based models it'd be best if you choose the older tag lists closer to that date, and for other models like FluffyRock its probably best to choose older lists than those uploaded here. Your decision to make

Some trivial info: NoobAI's e621 dataset is only up to date until early 2024 meaning older tag lists for e621 might be better but idk

The Danbooru tag scraping part was borrowed from BetaDoggo over here (many thanks)
Also a thank you to e621 for having every list easily accessible (danbooru has them easily accessible too but e6 was just easier to work with for me in this case.
also pandas dataframes are kinda cool.)

描述:

Created on 1st December 2024 (Upd. reason since it's out of schedule: NoobAI EPS 1.1 release)

Minimum post count threshold is 40 (-10 from last version since noobai seems to know more than expected)
Characters/Artist/franchine names + Active and deleted alias tags are included (pending alias tags from e621 are excluded)

Next update will include a major script update and remove useless tags like conditional_dnp and banned_artist etc.

If there are any performance problem with the huge combined list (even though there shouldn't be any) please tell me in the comments. I'm sure there's things that can be optimized for the list, I already got some rather trivial things in mind. I might also lower the post count threshold if noobai knows those tags too and there aren't any performance problems.

After this version there will be some changes to the tag lists and script mainly when I got time:

• deleted alias tags will be ignored if they exist as a full tag (don't think thats handled yet or is a real issue anyway)

• slight changes to the file names to make it easier to differentiate between different versions of self made tag lists using the script on GitHub

As always, you can find the tag lists on GitHub.

If you are using just the E621 tag list I recommend using the oldest version of it since there doesn't seem to be a model trained on up to date e621 dataset

I will also probably do some work to be able to create artist wildcards based on post threshold

训练词语: