Posted under Tags
Does anyone know how often the autotagger is updated?
I read that it was trained on Danbooru in 2022, are there plans to train a new model?
There are a lot of tags that are currently really difficult or even impossible to find with the AI now, and I think an update would be really beneficial for finding these images
I’d even volunteer to help train the new model with my 4090, if that’d help make a difference
https://github.com/danbooru/autotagger/releases
Training is very easy if you have a good, reliable, properly tagged dataset, and time. Since classifier models such as auto taggers actually take a very low resolution image to do the training compared to other models and the fact you will probably not train a whole model from scratch compared to my last assumptions your training on a 4090 should not take more than a few hours.
For further tagging in addition to the oblivious ones I add, I use Pixai with a small userscript that allows me to send the current image to an API and get it evaluated at 0.75 probability (compared to the model's default 0.35), afterwards I can click on the tags to add them to the list, combined with BrokenEagle's ValidateTagInput it works really well except a few occurrences such as not being able to tag footwear correctly (since it's around 8 months old, it tags with deprecated ones) but it's really easy to fix and very often I find tagging them myself anyway.
Updated by ggraphy
If AI tags gets an update, is there any chance artist tags could be included? artist request is at 135k posts and I feel like on average, it's also disproportionately higher quality art (at least these days). If you restricted it to artists with at least 100 posts for example, how many artists would that be?
I'm sure it wouldn't be a foolproof way to guess artists, but it'd at least help in tip-of-my-tongue situations where you recognize a style but can't remember a name.
wavedash said in forum #436519:
If AI tags gets an update, is there any chance artist tags could be included? artist request is at 135k posts and I feel like on average, it's also disproportionately higher quality art (at least these days). If you restricted it to artists with at least 100 posts for example, how many artists would that be?
I'm sure it wouldn't be a foolproof way to guess artists, but it'd at least help in tip-of-my-tongue situations where you recognize a style but can't remember a name.
I believe this would cause a lot of issues and would not be possible since the models usually get a very low resolution image and unless the art style is extremely distinctive it will spat out random artists for most images.
