That "ethical" argument is just silly. No one cares what lewd images anyone gns under any given name tag.

It's simply fear of copyright law suits, because people (butthurt artists) don't understand how diffusion works.

No artists at all. The trainer is very "pro-artist".

@Xenobreeder Some tags do work, asura works for instance, there's a list online that someone generated via brute force of about 100 artists so far. Either the data wasn't pruned well or the creators had let artist tags remain that didn't opt out.

@tofit42156 Have a link to that list anywhere?

@Ada321 Sorry for the delayed response don't check this site much, here's the last list I saw it's a few days old, it might be expanded.

https://files.catbox.moe/mezb4a.txt

I'm using standart automatic1111 setup and I managed to copy-paste params from the Millie pic from AstriliteHeart (author) and it generated exactly as the one from him.

@dextermorgan What is your webui version? Used fresh installed 1.7.0 with stock setttings,and something went wrong.

https://i.ibb.co/JzdHLWV/00096-4245040871.png

IMO the quality of turbo is kinda meh. If it's not too hard a turbo version of it might be neat.

Artist tags make for bad training. It puts all the training in artist "buckets" so models that use them are bad without artist tags. I shouldn't need a knowledge of furry artists just to generate a pony.

No offense, and I'm unrelated to the training of this model, but it's something to keep in mind. Furry models shoot themselves in the foot a lot with their over-reliance on artist tags to generate anything sensible.

@Torgo This. The 1.5 furry models can't generate anything remotely good without knowing a specific artist tag. And honestly, I don't want to generate in a real artist's style, because it hurts the artists. Those who attack real artists when they complain about getting into the datasets always use the argument "real people train the same way, bruh", but the real people never draw in someone else's style, that's plagiarism.

Instead of artist tags, we should have more specific style tags, like "cell shading", "gradient shading", "speed paint", "western", "anime", "sumi-e", "pencil", etc.

@Torgo @daxivi6026 

Firstly, over-reliance on artist tags is still a far better option than having no artist tags at all. It allows for much greater quality and style control by enabling you to blend different artistic styles to create new ones. This not only lets you generate new visuals you've never seen before, but also enables you to leverage the artist's strengths to enhance your work. For instance, if an artist excels at drawing tentacles, using the artist's name will undoubtedly result in nicer and more consistent tentacles. To a lesser extent, Nai3 also relies on artist tags. Are you going to criticize the most advanced and powerful anime model of the moment simply because it also incorporates artist tags? The most powerful 1.5 model for anime right now (easyfluff) uses artist tags. The most powerful SDXL model for anime right now (Nai3) uses artist tags. Yet you dare tell me that artist tags are bad.

Go to any pro-artist forum and tell them that your AI model, trained on millions of copyrighted artworks, is now "ethical" because it's not using artist tags. I'm sure it will go well; it certainly worked for Midjourney! Frankly, I have no idea where people got this silly idea from, especially when you can simply use a LORA to reintegrate an artist's style into the model. In the end, the only thing you're accomplishing is creating a subpar model to cater to people who don't care about it and will continue to dislike it anyway.

@Nonom This has nothing to do with the artists, it's just bad training.

You can do both actually, tag general style along with artist name. That way you can specify the look without it being neutral soulless mess that previous sdxl models were before.

Without artist tags it's just another neutered sdxl model, with them it's local NAIv3, SOTA basically.

@Nonom Your use of the word "cater" sounds like you're accusing someone of bowing to the woke mob, when in fact the artists created the anime models, and the work done by the AI community is like 1% at best, writing code and processing the data. Not including the tags is simple human decency, not "catering".

And you're right, you can still use Loras for artists' styles. I use them too, although I wish I didn't have to. It's like piracy. I know it's bad, but I still use torrents. Yet I don't pretend to be a liberator of progress, and I don't think that the creators of the content I pirate are some kind of parasites holding me back.

Besides, the argument of the artists' tags as the buckets that lock away the styles still stands. NAI3 is great, but you still have to use the artist tags with it. EF can't even generate anything passable without the artist tags. This is not a good thing. I shouldn't be required to know over 99 artists to get what I want out of SD. And I don't want to just copycat, I want a tool I can use in production. I want to tell the AI to give me 0.5mm thicker line art and cell shading and a little more anime or a little less, not to copy the artist_name's style. Maybe it's okay to refine what you've prompted with a specific artist tag if you want to nudge your generation in a certain direction, but the current approach of starting with the artist tag and then just throwing in the booru tags and praying to the god of randomness is terrible. Danbooru/Gelbooru is a shitty source of tags. It lacks quality/anatomy- and general style- tags, the same way e621 does. The booru tags are too abstract, and they lack nuances. The already dismal situation is made worse by the fact that the pony author has additionally purged some of the non-artist tags because of his subjective bias.

@daxivi6026 "I don't want to generate in a real artist's style, because it hurts the artists" Yeah, jerking off to a generated image that will be forever forgotten on the bottom of your hard drive hurts the artist, sure man. I am sick of how stupid people with guilt complex are. @Nonom is right, you are still hurting the artists by using amalgamation of all their styles as they say. Or to put it more correctly, by appeasing the complete strangers on the internet who hate AI, regardless non-artists or artists, you will achieve nothing. Either stop using SD or stop larping as SD user, OR stop listening to twitterheads if you still have a drop of common sense.

On the side note, style has never been a subject to copyright, not is it a plagiarism, but it is considered good manners to credit and give thanks to the artist you was inspired by. And SD users would've been much more comfortable crediting them if not for the twitter crowd gossiping, picking to pieces and regurgitating their hate of the AI.

@Torgo Publicly available SD models are more of an instrument where you want more precise control rather than a service like dall-e or midjouney where housewives can enter a few words and get good-looking image with no specifics. You can't get precise control without artist tags, maybe in the future if someone will train a model that can describe each component of every artist style with incredible precision, which may be interesting but difficult thing to do.

@clex I agree, but I shouldn't need to know furry artists just to get basic operation of a model. I don't know any furry artists I like that have a tag (or they have so few public images that the tag makes things worse), so this the first pony model I can actually use.

The problem I have with that training type (artist tag in front), is that furries pushed heavily for it, and are too used to it now. And thus, we never get out of the cycle, since they will demand the artist tags because that's what they are used to. And while you can use style LoRAs for replicating someone's style, you can't use any LoRAs to fix a model trained with artist tags in the front. You clearly realize how bad it is without them, while non-furry checkpoints don't need them to produce okay results and still admit style LoRAs if you want them.

And, just to be on the same page, because some people seem to be assuming, I'm not speaking for "the artists". I am one, I know my work has been used in models (I had a tag in this model before, incidentally), and I feel honored, not threatened. The others only seem to talk about fame, money or control, never mentioning how creating art is fun. We clearly don't see eye to eye, so I'm not willing to get into online slapfights for their sake. To me, the issue is entirely technical.

I do hope someone with the right hardware figures it out someday so we can have artist tags without them being a requisite.

Just to be perfectly clear: The problem is that they are required in all big furry models, not optional.

@Torgo EasyFluff v11 is perfectly capable of getting good results without artist tags. I'm not sure why you're making such a broad generalization. Outputs just come out looking kind of generic, which is also the case for this model.

@SlugCat Apologies, English is like my fifth language.

But in the "big furry model" category, going by top download order in civitai, the ones that don't demand tags are a minority or highly specialized (anime monster girls, hyperrealistic furry, kemono, eevees only no items final destination, etc). The message is already a wall of text without "except..." and "on average..." or stuffy wording like "top percentile...", etc.

I've used some that work okay but aren't quite compatible with fluffyrock or indigo LoRAs. Since many LoRAs ask for those big models, which are the ones that demand tags to work...you get the idea.

You also missed my hoping for someone to figure out a way to satisfy everyone. I am not against the artist tags being there, I'm only against it when it's the only way for the model to not output deformed or garbage generations. The point is, again, that I shouldn't need to know artist tags for basic operation to get anything comparable to the other non-furry models. This is still a furry model only issue.