I'll never understand the allowance of AI images.

Surprisingly, I am not against it, and I believe it is a natural state of evolution in art, whether we accept it or not.

The problem is, in its actual form, it is just tracing, which is unacceptable. I mean, there do exist AI that are able to do real "imaginative" work inspired by what's existing without showing much the references (and that is just what most of us artists do in fact!!!!), but they are so complex, resource-needing and above all not available on the market as they are prototypes (and as such would be incredibly costy also), that what we have today are only copy-paste machines with some degree of randomness.

But we will in the near future have much better AI available than that I am certain - that's exactly what happened with perspnal computers, mobiles, online translators, 3D softwares and so on.........

Seeing AI art as being similar to tracing shows a fundamental misunderstanding of how it actually works. You should study the mechanics of it much more in depth.

I do very much acknowledge I don't know much about AI image creation, however what I do know and you can't pull it off from me is that I saw many pieces of AI "art" that were basic tracing of existing works. MANY.
I know that there are several kind of programs and some actually produce very, how should I put it, "innovative" results, as in "extremely uncanny". These are however not the kind of AI I am talking about here (I admit I never precised that in the first place).

Now if you can enlighten me slightly about this I would be extremely grateful to you, really.

But once again, I am talking about what I am looking at, it probably makes me deliver some wrong theories about the process, although I still want to know how an AI generated "original" pic can look almost exactly as a copy of an existing work with just meaningless details and colors changed. I don't care much about the physics/mathematics of it then, it's just plainly unacceptable.

How AI creates art: https://jalammar.github.io/illustra.....ble-diffusion/

Thank you for the link, it is far too technical though and I don't really see concrete examples of the process, I am absolutely alien to such sentences as "The timestep is transformed into a time step embedding vector, and that’s what gets used in the layers" - wich is basically what composes all this article, it doesn't translate at all into what's happening in an image in my mind.

I can try to explain it plainer.

The AI has multiple components, one is the text classifier which allows you to turn language into inputs the image generator understands. That's not really relevant to what's contentious here, so I won't go deeply into that.

But in short, the text classifier takes text and outputs a set of vectors (which are basically a list of decimal numbers, mostly in range of 0.0 to 1.0). This set of vectors describes the image in terms of "how much of X" is in the image. If there is a red ball in the image, then the amount of "ball-ness" is high, close to 1.0, and "red-ness" is high, but the amount of say, "box-ness" is relatively low. There are millions of parameters here that go into the image generator. (This is gross simplification, but general idea is that)

The image generator also has multiple stages, it is so called autoencoder, which means it encodes inputs into what's called "latent space" does a transformation on that compact latent space, and then decodes that latent space back into outputs. Again, latent space is just a set of vectors (list of numbers).

Now as illustration of an autoencoder, I could feed one an image, it would encode it into a compressed latent space description of the image, then it would decode it into an image that would look mostly the same as the source image (but not identical). How well the autoencoder works depends on the training it has received. The training allows it to compress the image better into the latent space, and then expand the latent space better back into the image.

The training data is not stored into the model, and you can't extract any of the training data out of the model. But the training data influences how the model does things. If the model has seen certain kinds of brush strokes a lot, then it will output that kind of brush strokes. And so on. The more it has seen certain kind of feature, the more it wants to output similar features when that input parameter is high.

Now, in our case our input is not an image, but instead we start with literal random noise (random values) because we want to create a new novel image. And we don't start with image of noise, but rather we skip the encoder and directly feed the decoder latent space noise inputs. So we start with completely random state of the "brain". "This image has 0.8 bucket, 0.2 ball, 0.7 fox, 0.5 box, 0.3 red, 0.9 blue..."

The next step is the "diffusion" step (this is where Stable Diffusion gets its name). We have another NN that is trained to guess how much noise has been applied to an image (or latent space in this case), so it can say things like this has 10% noise, or 50%, or whatever. But that can also be run in reverse and it can be told to remove X % of noise from the input.

We use that to remove a bit of the noise from the latent space, then we use the text input vectors to select the result that guides it toward those vectors. And repeat until we're happy.

This converges surprisingly fast into an image (because the latent space is very information dense) and each subsequent step refines it. This video from the linked page shows it working step by step:

https://jalammar.github.io/images/s.....-all-loop.webm

I'm cool with it if the AI uses pictures of real human and animal anatomy as a base, not images that are already a human-created synthesis of those things.

Considering how much concern there was over AI-generated art, it was a foolish decision by DA to allow it on their platform as they. While they did do damage control really fast in automatically making all submissions opted out, its still a dumb move and one that will hurt their reputation even further.

My words exactly

It's very romantic. The besieged artists' last stand against the AI barbarians. I understand artists don't want their art used in the AI databases ; but I think this is a very transient battle : In a few years AI will be so pervasive that it won't matter.

This reminds me of the young internet. Companies and inventor would try to prevent people from saving pictures to their computers. By using flash or java applets and whatnot. Well it didn't work, and nobody in their right mind nowadays even dreams of stopping the copy of pictures published on the Internet.

This struggle against AI will be just the same in my opinion. The horse is out of the barn already.

That is indeed exactly what I replied  ty_hanson above if you fancy to have a look!

I stopped using DA when they forced that extremely ugly UI change on users a year or two ago.
Seems like a crappy management on that site.

I only mentioned it and did not insist, but that was indeed another major awakening about their philosophy...

It has apparently already been repealed due to user feedback. You now have to opt-in to have your images used in datasets.

Personally, I feel DA is a better website than FA, but FA has a more established (and stubborn) community.

I've seen the first pages of rants from users in that Twitter feed, yes, as I said they were fast to react, but if you look closely at some comments there might be many.... Exceptions? Glitches? Roundabouts?

Also, one pertinent thing I saw from a video on Youtube is that their AI does not gather pieces of art ANYMORE, meaning that not only it still keeps what has been collected between the moment they officially made that switch on until they back-pedaled due to the imense backlash, but this AI actually was already collecting art far before that without anybody being really aware!!!

I am curious to hear the reasons you say you still feel DA is a better website than FA, it is true that there are so many aspects to be considered to create an opinion, as for myself I am essentially looking at how the staff cares about their artists. And nobody will make me change the idea that while FA is artist-oriented, DA is money-oriented, and this is a MAJOR difference to me, and actually my main focus.

The main difference bewtween the two is : FA allows most hentai adult pics. DA says they don't but if you look - they allow soft core porn. The problem is DA ahs too many obsessive child "artists" who draw like they're in 3rd grade yet they are aloowed to post their crappy work.

It always made me chuckle, ALWAYS, that "Deviant"Art was actually the LESS deviant site of all - since the beginning the banner is deceptive of the product.
Also to complete what you said, yes, DA doesn't allow hentai pics, you will get axed quite quickly if you upload such a drawng, even if it is artistic and extremely innocent. Buuuuuuuuuut it is perfectly OK to upload real photographs of spread genitals, yeah, this has obviously a lot more artistic value for whatever reason I just don't want to hear about. Well, actually I regretably know - the artistic value lies in the money you can't afford to give them but those pornographs generously shove in.

That's just something I actually found out all by myself, by the way, my own real experience.

They didn't even allow art showing erections. That's very puritanical.

well I dont have much uploaded there, it was always such a pain to upload there with the 50 different catagorys and if it wasnt in the right one they would delete your stuff.

They would do that on top of the rest??!!!! I didn't even know, that's just... Insane

Even better are all those people who abandoned FA for Twitter. I simply don't understand the mentality of choosing a single service that seems to be the most popular, and using it exclusively.

Sooner or later, people will have to learn that "loyalty" doesn't pay. If a site feels they are in a position to screw artists over, they WILL do so. Your best bet is to diversify so if any one service goes to hell, you have options. I had no trouble deleting my gallery on DA years ago, since I post my art on a whole bunch of different sites.

Ha ha, don't even push me on the Twitter subject!!!!!!
It's been a place of shitstorm even before the Trump/wokeattitude era. Yes, you got exposition there, I reckon, but you also get unleashed all the worse humanity can produce. I bet nobody there doesn't have a list of at least ten accounts they had to ban from their pages - and you are yourself in a ban list of at least ten people without even knowing why! And I don't even speak about the "moderation".
On top of all that, Twitter isn't an art site from the beginning, it's great to expose your boring life, not your wonderful creations!!

I guess I have a very fair reason to delate my DeviantArt account now. I never liked this website and it was never of much use to me.
But... so, that's the best thing to do?

This is all your own will. You know what this site is about. You decide whether you stay, accepting all the risks, or leave, and potentially deprive you of some exposition and friends there.
What I'm saying is just that it is suicidal to exclusively get your gallery there, since it's the less artist-friendly art site I ever saw.

Agreed - and there's no guarantee that your gallery hasn't already been scraped and sold off

True, exactly as I told  DOPR5 above......

I like FA rather than dA and Twitter. Also, e621.