Thread about tags for various "AI" image generators

It’s not yet formalized what overarching tag(s) should be used to group these kinds of images. There had previously been a discussion between me (Background Pony) and PUBLIQclopAccountant in the General Tag Discussion thread, beginning at the linked post and in this thread (and a little in the thisponydoesnotexist thread, mixed with philosophical debate) without really reaching any firm conclusion.
machine learning generated is the tag used to group these kinds of works.

Please give some feedback!

The only tags implying anything at the moment:
artist:thisponydoesnotexist implies neural network only.
dall·e mini and nightcafe ai imply ai generated only.
Several more have been updated.

I think it might be good to invent a tag to identify when there’s ai generated imagery PLUS human-artist generated imagery in a post, for cases where someone’s redrawn something, like in >>2718269 - because it’s not sufficient to search or filter with a query like ai generated, artist:* because people may use artist tags on pure ai generated imagery because the process of writing a prompt and curating the results is arguably an artistic process.

Suggestions?

Audio:

Except for uberduck and so-vits-svc, these tags are used both for output of the software, and for a few images relating to/fanart for those projects. None of them have any implications set up.

Visual:

Video:

Please reply with more if you know about them.

midjourney has appeared

I’ve done a bit of cleanup on various tags competing with “ai generated” so they’re restricted to in-universe things instead of referring to image creation, and aliased “ai generated” to “machine learning generated”.

dall-e 2 alias-> dall·e 2 and migrate the category and implication

dall·e mini and all its aliases -> craiyon

waifu labs imply-> machine learning generated

disco diffusion imply-> machine learning generated

stable diffusion imply-> machine learning generated

“This pony does not exist” uses an artist: type tag. I tried to follow this example and created artist:stable diffusion. Then someone removed the artist: prefix.

So what is the current consensus? Should AIs be tagged with the artist: prefix, without it, or even with an ai: prefix or similar?

EDIT: And perhaps we should have separate tags for ML-generated sound and ML-generated imagery?

@Background Pony #AA1D
I’m not a consensus, take from my comments what you will:
There was a proposal to use an ai: prefix, which I disagreed with instead recommending no prefix for tags identifying software tools. I also separately suggested removing the artist: prefix to empower searching and filtering around it.

thisponydoesnotexist is a little different from other AI art tools in that, AFAIK, if it hasn’t changed since I looked at it, all the images it serves are pre-rendered and you’re just looking through the bulk collection, not actually changing or producing anything. In that sense, thisponydoesnotexist could be thought of as the artist: (or, the programmer of the generation software project instead) or instead it could be thought of as a fanwork/series, which also has been suggested and is what I think I agree with the most.

tl;dr I’d just call them all without the artist: prefix, or any other prefix, and categorize as appropriate. Not even sure where it came from to label “thisponydoesnotexist” as the artist.

purplesmart.ai imply-> stable diffusion, machine learning generated

I see that machine learning assisted is a thing. Does it apply to cases where someone made a very small edit to the AI image or only when the human contribution is substantial?

An example of the former is this pic, where only the mouth was slightly edited:

A less extreme example is this pic, where someone cleaned up the AI pic:

@Background Pony #AA1D
IMO, “substantial”. Whatever that means.

I’d make a first try to define that as there being concrete enough portions of the image that you could state, “X part of the image was neither generated by an AI nor handmade for the purpose of correcting AI artifacts/errors, and that part is not a trifling part like an earring.”

How to judge that line in terms of significant hands-on manipulation, cutting/compositing, creation of underlying images fed into “img2img” functions, etc. I’m still quite uncertain about.

@Background Pony #5ACC
Then how should we tag slightly corrected AI images? It’s useful to tell them apart from raw AI images in order not to overestimate the abilities of a given AI.

@Background Pony #AA1D
edit, editor:*, and write in the description what has been done. Other than that, I don’t know.

Since we require the image uploader to tag the artist and provide a source, should we require them to tag the machine learning system which generated the image and provide the prompt, respectively? This is of course about machine learning generated, not machine learning assisted images.

novel ai alias-> novelai (the service’s name is written without a space)
novelai imply-> machine learning generated, stable diffusion

Done.

Another model I guess
anything pony: sd is magic imply-> stable diffusion, machine learning generated

Done.

pinegraph imply-> machine learning generated

Done.

imgcreator imply-> machine learning generated

@Background Pony #AA1D
I agree that the prompt should be required where available. However, I don’t want images banned because someone went scraping Discord or Twitter for AI images that did not have prompts attached.

@PUBLIQclopAccountant
The artist tag will always be much more important than that information, but we don’t delete images because they don’t have artist tags. We just ask for them, or try to find them.

If someone is scraping Discord or Twitter for AI images and tagging them badly, they’re more likely to be deleted for Rule #3.

pony diffusion v4 <-> pony diffusion <-> ponydiffusion

all: imply-> stable diffusion, machine learning generated