Welcome to Furry_irl, a community for furry memes, shitposts, and other relatable images or comics.
Community rules:
Happy 6/21 everyone! Keeping up the tradition of a yearly Esix for the big day.
This is kind borderline for furry_irl, but I think with people memeing about 6/21 a bunch and with how small of a window it is, I'd prefer it here over furry. What are your tag(s)? >:3
hand_holding, hug, kissing, rating:s
Trust me, those are the only tags you'll find in my search history UwU
duckduckgo how to delete e6 history
Ooh I search rating:s all the time! Though I do it with a dash before it first...
Now that you mention it, most of my searches contain at least one tag that makes -rating:s unnecessary :3c
IMO: they need to add an Ai tagging feature to E621. Ai is basically purpose built for image tagging. Then users could upvote/downvote accurate tags the Ai suggested. If the tag is highly upvoted, then it could be added to the image. Welcome to my Ted talk.
Thanks for listening to my rant
I don't have experience with models, but if there is a compatible free software model (E6ing is FOSS, though not GPL) which suggests tags with a reasonable degree of accuracy, sure. Given the way images are uploaded, if such a model suggested existing tags for all new submissions the uploader could validate, that'd likely be a great jumping off point instead of only a handful of tags and a +tagme. Maybe these suggestions could also exist within the tag editing page for others to validate in the future as well.
Could use a system similar to DeepBooru. Commonly used with the FOSS Stable Diffusion tool for suggesting tags for input images used in training. Basically auto tagging so SD knows what is in the images its being trained off of. Could likely implement something like that.
I love the idea of suggested tags. I'm not super familiar with Ai models, but it seems like using their current data to train something wouldn't be too hard. They already have a bunch of images with accurate tags.
TLDR >!AI is not a solution from moral, practical and economy standpoint!< . Of course I am open to other's standpoint so correct me if I am wrong anywhere. :3
while big AI models are REALLY power hungry. Mi****oft restarted whole nuclear power plant to manage energy demand.
Training models is very enegry hungry. But using models can be ran on consumer graphics card. I suppose you can argue to include the training cost to using models, but the models are going to be trained rather you use them or not, so the point is just moot.
I don't entirely agree, but your points are valid. I in no way believe E621 is non-profitable. If it wasn't making money, it would go away. Second, it should 100% be an opt in system. It should also remain proprietary so nobody could use it outside E621. Getting artist on board wouldn't be too hard if you made it clear that its explicitly for image tagging and not for generating spoofs of their art. Make it up front and clear that this is only for enhancing e621 and not stealing from them. I'll just skip over the training point because there doesn't seem to be a flaw in that logic. It will be expensive to train. I do dispute the accuracy numbers. Since this Ai would only suggest tags, it doesn't need to be 90-100% accurate to be helpful. It just needs to put ideas for tags in the uploaders mind. As more tags are added to images, the Ai could also gradually improve.
AI is controversial morally and resource hungry. Why use an AI model and have users check it when the users could just add the tags themselves?
Because they don't. Many images get uploaded to E621 with few tags, only enough to allow the upload. The other problem is that people may not be aware of the tags they can add. E621 allows for a lot of very specific tags. If a character is wearing a swimsuit, you could add that as a tag. You could also be way more specific and say what kind of swimsuit and what color. The goal should be to maximize the number of accurate tags per image so its easier to find what you're looking for. Yes, Ai can be very resource hungry to train, but thats the trade off. I see no moral issue with training an Ai to identify commonalities between images and quantify those common elements. We aren't using this Ai to replace artists, only help volunteers who tag their images.
Speaking of specific tags, one I often search is "Dolphin Shorts" (basically shorts with usually a single stripe down the side) and many images that feature them don't have that tag. Probably because most people don't know that is what they are called.
Exactly this! Most people would just add "shorts" and not specify what kind. I try to be as explicit as possible. I'll colors, styles, textures, and whatever else I can think of or that gets suggested.
I don't know how to even add tags without having uploaded something myself... 😕
If you have an E621 account, click the "edit" button under the image. There will be a few boxes where you can enter in meta data about the image. This can include extra sources, general tags, descriptions, and parent posts.
Ah... It's right next to the comment button; I assumed it was to edit your comment 🤣
How would you train such a model given that there isn't much public domain furry art?
Thats probably the biggest challenge. E621 would have to make it an opt-in program. I haven't read their ToS so it's possible they already reserve the right to use the images uploaded to their site for any purpose they want. That wouldn't be the right thing to do though.
fish order:score -animated