Here are three sets of wildcards with some of my successful prompting combinations from the last year. I also added a bunch of language from the danbooru wiki. V2 tries to be better with using actual danbooru tags with the underscores.
Description
FAQ
Comments (10)
Please upload results here, since wildcards don't auto-populate like other models
crazy timing i was just working on wild cards i created about 4000 of them but this is good too!
Thanks for the share, always appreciated. ^^
Tho i have a question after looking at tags you've put into your txt files. Many "tags" aren't Danbooru tags so those being correctly interpreted by the text encoder is either checkpoint dependent or relying on Illustrious capacity to deal with natural language.
In both cases this isn't secure nor accurate at all. If a checkpoint hasn't been trained using those non-Danbooru tags then those tags won't have any effect and their impact on the picture generated will only be the result of the prompt weight change because of them.
And if counting on Illustrious capacity to deal with natural language then this is always a bad bet for as far as Illustrious isn't that good at dealing with natural language.
So what was your angle exactly? Because if there is not a solid logic behind those non-Danbooru tags then it would be better to clean those text files and remove those tags and try to replace them with actual active Danbooru tags.
Oh, I copied most of the tags over from tha danbooru wiki. some of the additions are just things that have worked in my images in the past. maybe I misunderstood. I reworded the description to more precisely reflect my process.
@KaosEngine Yeah, no problem. Tho has mentioned if those aren't trained tags then you d better remove them and replace them with active danbooru tags (not deprecated tags). This is the full active Danbooru tags list: https://danbooru.donmai.us/tags?commit=Search&search%5Bhide_empty%5D=yes&search%5Bis_deprecated%5D=no&search%5Border%5D=date
@antarek thanks! I'm working on revisions. V2 will use primarily accurate current tags. I've gone through the camera file and brought it into compliance. Poses might have more creativity, but the emotion list was largely built off the wiki, so proper tags shouldn't be too hard to find
@antarek Hopefully v2 does the danbooru thing better
@KaosEngine This is much better. 👍
@antarek thanks for the constructive criticism!
You're welcome. And for anyone using Danbooru tags on ILXL checkpoints, which you should be doing if you want accuracy instead of rolling the dice, remember that your checkpoint’s understanding of Danbooru tags is entirely dependent on the specific database version it was trained on.
The real problem is that about 99% of released checkpoints are merges. And most people have absolutely no clue what datasets the original, base-trained models used before being merged. Then those merged checkpoints get merged again, and again, and again… so you end up with a giant pile of models built on top of other merges whose training data is basically a black box.
You can imagine how messed up the situation has become when it comes to knowing what training data any given checkpoint is actually based on today. So to have somewhat of an idea regarding the dataset version that a merged checkpoint knows you can test using character tags going from the most recent Danbooru addition to the oldest one. When your checkpoint ll render a character that matches the one you asked him to generate using those character specific tags then you ll know on wich dataset your checkpoint has been trained on.