Regarding a method to avoid AI-generated style
This is my personal opinion; I believe AI-generated content should be trained.
Because even if you try to avoid removing AI-generated content from the training set, it's difficult to completely prevent it from mixing into the training set.
Wouldn't it be better to explicitly let the model understand the style of AI generation and then stop it in the negative prompts?
Wouldn't this make it easier to avoid AI-generated style image generation?
I think you want this thread: https://huggingface.co/circlestone-labs/Anima/discussions/83
Yes, I saw that post, but my content isn't entirely duplicated.
Explicitly defining the AI's generated style (training more AI-generated styles to label AI-generated content)
and refining the scoring criteria are two different approaches.
As far as I know, Anima's training set is:
"Generally, all sources used for anime images prohibit the use of AI-generated content."
Furthermore, in fact, in Anima, I used AI-generated content as a negative indicator, but it didn't suppress AI styles like other models.
I believe that prohibiting AI-generated content leads to the AI not explicitly defining its generated style (even though there are still overlooked AI-generated images mixed in).
But in the current situation, the emergence of AI styles is an objective reality.
This:
Wouldn't it be better to explicitly let the model understand the style of AI generation and then stop it in the negative prompts?
Was brought up in that thread, someone said they'd train a lora to test this out. I decided not to reply initially to let the person train their lora, but I might as well reply now: someone did that a while ago and it does work for Illustrious, at least to some extent - https://civitai.com/models/1452858 (NSFW)
If you're hoping to get the model to not mess up fingers or the other errors typically associated with AI, that's not happening in this way.
I think that drawing the wrong hand, and other AI-related errors, don't only occur in AI-generated works, but also in human-generated works.
However, as long as the tags generated by AI aren't dropped during training, I think it should be possible to have relatively good control over them.
I think that drawing the wrong hand, and other AI-related errors, don't only occur in AI-generated works, but also in human-generated works.
The vast majority of AI generated errors do not occur in human works - no 6 fingers, no 3 legs, no holes or incoherent details, distortions, concept bleed or other stuff. The only notable error I can think of that both humans and AI do is flipped hands/feet, though humans do those very rarely, certainly not enough to influence a model. Danbooru has a tag for "bad hands" - this means badly proportioned hands, usually with spider leg-like fingers, not 6 fingers or fingers melting together. More than that, e621 has tags for finger count (though it's mostly biased around 4 fingers). The model with the obfuscated artist tags still screws up fingers.
You are not prompting out most AI errors. They happen because the model/architecture is too weak, not because it's specifically learned to make 6 fingers from humans. And making the model/architecture better is no trivial ordeal.
no 6 fingers
Artists make this mistake decently often thus there is a tag for it on danbooru as well, "extra digits", of course a few of those uploads are intentional but not most of them
I don't think <437 images is decently often, no way it's enough to affect the model in any way other than maybe making them promptable (assuming "digits" doesn't mildly confuse it anyways like "cross" does in cross-eyed).
You never intentionally add bad data to an AI model with hope you can rely on imperfect CFG to negate this bad data. This applies to LLM, Image model, anything else.
"Garbage in garbage out" method is especially relevant for AI.
I checked some of the "extra digits" images on booru and honestly its just stupidity I think? Like they count the skin next to the pinky as an extra digit/finger. I understand, that this can be confusing and some images indeed show an extra finger that shouldn't be there but that tag is badly managed on booru. If anything it's often just slight perspective error, at most.