Topic: [Feature] The tagging system is outdated, we need algorithms for automatic tagging

Posted under Site Bug Reports & Feature Requests

Dear e621 Moderators,

As you may know it, e621.net is the most reliable furry imageboard on the internet because of its sophisticated and complex search function and blacklisting algorithm. But, it also has its flaws: it's completely manual, the Tag What You See rule attached to the methods of tagging has some space for gender lobbying, hereby getting images where the character's gender is wrongly tagged, get tucked into lots of people's blacklists and they can only watch and follow the content they would like to consume if they remove the tag that is wrongly associated with the picture/character from their blacklists, risking the danger of actually seeing images that actually fit the blacklisted tag's criteria (for example the Mikhaila Kirov gender controversy).

So for avoiding conflicts, death sentences and mass federation bans, I would like to recommend developing an automatic filtering system, which would work just like Google TensorFlow: the user would upload a picture to the site, then the AI would start its work recognising patterns in an image that the programmers had previously taught to the algorithm, like general body and face shape recognition for accurate gender detection without the actual input of the programmer or the moderator (to avoid another controversy similar to the Mikhaila fiasco), accurate genital detection, species detection, artstyle detection and other detections to make it sure the picture gets every necessary tag possible for sharing it to a broader audience. And finally, after 20-30 seconds of image and tag processing, the image would be uploaded with (mostly) the right tags.

Although the robot-generated tags would be locked tags and would be tied to the post, the user would still have the ability to change tags via directly contacting a moderator. When a moderator approves a tag change, he/she would change the old tag(s) with (a) new, still locked tag(s) and the changes would appear in just a few minutes. Even though it appears to be the same experience for the moderators, the catch is in the details: with this new, automatic tagging system, the moderators would only be granted with the permissions of editing tags when another user has contacted him/her about this issue, to also avoid gender lobbying and the abuse of the moderation tools.

In conclusion, developing an algorithm like this (in my honest opinion) would be hard and requires a lot of money, time and work put into this and the programmers will probably have to take down the site for a day or two when they are done with the development part and they will be applying the algorithm to control the entire site's tagging rules, but it would be extremely beneficial for the users of e621.net, for the Staff, for the Founder of e621 and for the CEO of the parent company because the Staff will only have to change tags when it's urgent, to prevent future controversies surrounding tagging.

I hope you take my feature request into consideration, find it useful and implement it because it would mean a lot both to the users and for the Staff, too.

Yours sincerely,

Louis (GRFurryStuff), a caring user of e621.net

Updated by Pup

I had hopes this would be more broad, but at the end of it, it's only about the gender tags. The reason that the blacklist works well on the site, and why finding things on the site works so well is because of manual taggers. People that search things out and include things where they should be, things that bots can't be programmed to do accurately. And tags which can be dealt with by bots already are dealt with by bots such as the 1:1 tag or the resolution tags.

-1 to this idea, if it's only purpose is to search for gender ruling.

Updated by anonymous

This is the worst idea so far, and I genuinely hope this is just a joke.

Updated by anonymous

Reminder that google photos tagged my meatballs as flowers and when searching for my cat it doesn't find all of the images with cat in it.

And this is still real life, not fantasy content which can be drawn in multiple ways, this is also why there's many furry porn tumblrs which are fully unaffected by the porn ban because tumblr image recognizion does not see porn in the drawn furry porn images unless style is extremely realistic.

On top of this we would need to start paying google (it's not free service) instead of relying on free work labour.

Updated by anonymous

Wow... -1. There's no such thing as a flawless tag system, and our tagging system works better than algorithms would for this site. Unless you can guarantee that it's flawless, 100% of the time...

Updated by anonymous

-1
I think you know how well YouTube's algorithm works. Making one for e6 wouldn't be any better than that.

Updated by anonymous

Well... At least I shared my opinion. Even though nobody likes it, it still somewhat fills me with joy that I still requested a feature. Just lock this thread and call it a Denied request. It's all for the better.

Updated by anonymous

GRFurryStuff said:
Well... At least I shared my opinion. Even though nobody likes it, it still somewhat fills me with joy that I still requested a feature. Just lock this thread and call it a Denied request. It's all for the better.

It's not that we don't like it, it's just that it wouldn't work at all with current technology.

Updated by anonymous

GRFurryStuff said:
Well... At least I shared my opinion. Even though nobody likes it, it still somewhat fills me with joy that I still requested a feature. Just lock this thread and call it a Denied request. It's all for the better.

I agree with @Pyke. You did a good thing. :) For my own part, I think we'd just spend more time fixing mis-labeled tags than we would benefiting from the AI. It might actually cause more arguments than fewer.

Updated by anonymous

GRFurryStuff said:
Well... At least I shared my opinion. Even though nobody likes it, it still somewhat fills me with joy that I still requested a feature. Just lock this thread and call it a Denied request. It's all for the better.

Even if something isn't as good of an idea as it might seem on paper, bringing it up can still be beneficial if only to put it out in the open where it can be shot full of holes instead of lurking in the shadows as a "maybe".

Updated by anonymous

-1

Unrealistic, algorithms are not really made for anything other then objective repetitive simple tasks, gender/sex in visual art certainly isnt one of them. Complicated by art styles, deferring detail, image quality, depicted species, forms(feral,anthro,eta.),character number and so on. If there are already regular disputes by people as to what gender a character is in the art then algorithms are certainly not going fair any better, and you wanna give the algorithm locking permissions too, that is just a disaster waiting to happen.

Updated by anonymous

Clawstripe said:
Even if something isn't as good of an idea as it might seem on paper, bringing it up can still be beneficial if only to put it out in the open where it can be shot full of holes instead of lurking in the shadows as a "maybe".

The idea of "tag suggestions from an artificial neural network" has been brought up recently, actually. forum #275999. There was one particular user who seemed to have some experience in AI, but I haven't heard whether they've made any progress.

I do think the idea of generating tag suggestions is a valid one, and the state of AI technology seems promising, but I would not agree at all with such a system being used as an authority. The tech is nowhere near that point yet.

Setting aside the issue of accuracy, however, the best-case scenario would be the AI accomplishing what TWYS already aims for: tagging an image based on the data actually present in the image, not based on outsider knowledge (which has been the center of all the sex-identity arguments).

Updated by anonymous

Replacing all tagging with automated tagging is a bad idea.

Assisting tagging with automated tagging could work and has been brought up repeatedly. But you will need to check back in at least 5 years before you can expect good results. e621 or more likely the users themselves might want dedicated hardware for accelerating it.

Updated by anonymous

Pup

Privileged

There's things like Deep Danbooru which works quite well for anime, but anime characters still wouldn't be as varied as furry characters.

I tried testing something with tensorflow, and it wasn't too bad though it ended up mis-tagging breasts/butt, and was correct for 90% of anthro/feral, though it fell apart when characters were kneeling.

The other thing that I found to be a major problem when teaching a network is that it expects you to say if an image contains something or not, but just because an image isn't tagged doesn't mean it's not there. There's still a lot of posts in here, -zero_pictured -solo -duo -group, so you can't even feed it all images and have it tag character count.

There's also tags like patreon, which is what I was originally trying to teach it, which often get mis-tagged as it's just for the logo and depictions of the site, not content that's from patreon or a patreon URL.

Updated by anonymous

  • 1