Penance said:
The overlap of people who are "AI enthusiasts" and people who actually know anything about computer systems is almost nonexistent.
Somehow people managed to do that just fine in the span of over a decade before AI tagging existed on Danbooru!
If you think that "AI enthusiasts" will not think of using autotaggers locally on their computers or simply will not use workspaces on huggingface, then you are wrong.
Think about "Why "AI enthusiasts" found your site?"
Using an autotagger on your site does not play a major role for the "AI enthusiast" in terms of creating automatic tags for images, since it is not the autotagger that plays an important role here, but the banal convenience of using tags already existing on the site and learning new tags that the user did not know about (but to whom your autotagger offers new tags). For me, as a user, it was simply convenient to publish content and tag it, since your site and autotagger solved several problems:
- It offered tags from which you could choose the right ones. Very quickly and for a large amount of content.
- It offered "technical" tags.
- The user chooses the tags suggested by your SITE, and does not submit tags from some other autotaggers that he found on the Internet. Newbies in AI will probably use bad (inappropriate) autotagers, which is why they will upload images with poor quality tags to your site, and not what your site suggests.
- It was possible to more conveniently and quickly add new tags manually in the text field itself (for example, with characters that have appeared recently, which the autotager does not know about) - this is also convenient in combination with the autotagger.
- Thanks to the autotagger, it was possible to learn new tags, although not immediately and not all of them. It was easier and faster for newbies to learn new tags, although they make a lot of mistakes and are often lazy (for this, this toy can be taken away from them, and not punish all users of the site).
- It is easier for honest users and moderators to add good tags (by performing all actions on the site) or to add tags to existing pictures.
Now, those who used your site to create more or less high-quality tags for their pictures will be forced to do all this locally or on other sites with a similar system. This will simply lead to an outflow of people who uploaded content to your site. But the worst thing for you is that now "AI enthusiasts" instead of using your convenient interface will use local tag generators, and, which is very likely, will not even check these tags with a huge probability in relation to your site and upload these terrible tags without paying attention to reputation and all that.
I see from the posts that even with the autotagger disabled, posts with a huge list of tags of dubious quality appear on the site. And as the site grows, this trend will only get worse, since newbies will use all this with less care.
The problem is not that your autotagger allows people to "do tags automatically and not care about quality", but that autotaggers appeared at all. Danbooru could exist without autotags for years, because there was no autotagger and people had to do everything manually, but the number of tags on average for each picture on Danbooru is no more than 30 pieces. In addition, artists did not make so much content to worry about writing tags, unlike "AI enthusiasts". AI enthusiasts, in connection with the appearance of autotaggers, are in dire need of high-quality autotagers, and they will look for them / buy / train them to solve their problem. "Somehow people managed to do that just fine in the span of over a decade" - does not work when automation appears, available to everyone. On AIbooru, instead of fighting with people who use autotaggers, it would be better to retrain the autotagger Danbooru based on the new neural networks, since it has not been updated for over two years (or implement existing ones, such as wd-eva02-large-tagger-v3, if the tags match). And also limit the number of suggested tags for images to 30, with a match percentage of 65+. This will greatly improve the quality of image tagging, people will make fewer mistakes (with a small number of suggested tags, it will be easier for them to remove unnecessary ones). At the same time, moderators can be given access to extended tags (100+ tags with a match percentage of 30+), so that they add suitable tags.
If you look at the autotagger wd-eva02-large-tagger-v3, it currently has colossal accuracy in identifying tags. It is very surprising how accurately it produces results. But tags from wd-eva02-large-tagger-v3 and other similar autotagers do not always match aibooru tags, and this is already a problem for aibooru if someone uses these tags when publishing content on aibooru, not paying attention to the problems with the autotagger.
Remember one thing - when you force people to use external tools (in this case, local deployment autotagers), you will get the sores and problems of external tools that do not match your data and content sets.
The same applies to Danbooru - it will suffer the same disease due to the spread of external free autotagers. Regular artists are already starting to use this, too, although they themselves are against the development of "AI tools" (and it's fun to watch).
And don't take my messages as criticism or requests to return the autotagger and checkboxes. I don't care what happens to this function. I, like other users, will always find other sites or programs where it is convenient and easier for us to work with content, where it is easier to distribute it.
But take my message as one of the serious reviews that tries to show you that there is no "black and white" and there is a trend that you need to adapt to.
I will simply hope for the best, that future autotaggers will give almost perfect tags, so that your sites will not have problems with those people who will use external autotaggers; so that they do not harm you.