#generated artwork

[dalle2]

“The FULL Quaker Oats painting”

Created with DALL·E 2, a new AI system by OpenAI that creates realistic images and art from a description in natural language.

Sharing is caring!

Substack: dalle.substack.com

Twitter: @Dalle2AI

The heading of this post was used to generate the image, src

[gay-i]

[lexart-io]

GENERATIVE AI ART 〄 LEXart.io

New IG! Follow me ⇝ instagram.com/lexmartin.ai

[tangibletechnomancy]

How To Use AI To Fake A Scandal For Fun, Profit, and Clout

Or, I Just Saw People I Know To Be Reasonable Fall For A Fake "Ripoff" And Now I'm Going To Gently Demonstrate What Really Happened

So, we all know what people say about AI. It's just an automatic collage machine, it's stealing your data (as if the rest of the mainstream internet isn't - seriously, we should be using that knee-jerk disgust response to demand better internet privacy laws rather than try to beef up copyright so that compliance has to come at the beginning rather than the end of the process and you can be sued on suspicion of referencing, but I digress...), it can't create anything novel, some people go so far as to claim it's not even synthesizing anything, but just acting as a search engine and returning something run through a filter and "proving" it by "searching" for their own art and "finding" it.

And those are blatant lies.

The thing is, the reason AI is such a breakthrough - and the reason we memed with it so hard when DALL-E Mini and DALL-E 2 first dropped - is because it CAN create novel output. Because it CAN visualize the absurd ideas that no one has ever posted to the internet before. In fact, it would be a bigger breakthrough in computer science if we DID come up with an automatic collage machine - something that knows where to cut out a part of one image and paste it onto another, then smooth out the lighting and colors to make them fairly consistent, to make it look like what we would recognize as an image we're asking for? That would make the denoising algorithm on steroids that a diffusion model is look like child's play.

But, unlike the posts that claim that they're just acting as a collage maker at best and a search engine at worst, I'm not going to ask you to take my word for it (and stick a pin in this point, we'll come back to it later). I'm going to ask you to go to Simple Stable (or Craiyon, or the Karlo demo, if Google Colab feels too complicated for you - or if you like, do all of the above) and throw in a shitpost prompt or two. Ask for a velociraptor carousel pony ridden by a bunny. Ask for Godzilla fighting a wacky waving inflatable arm flailing tube man. Ask for an oil painting of a capybara wearing an ornate princess gown. Shitpost with it like we did before these myths took hold.

Now take your favorite result(s) and reverse image search them. Did you get anything remotely similar to your generated image? Probably not!

So then, how did someone end up getting a near perfect recreation of their work? Was that just some kind of wacky, one-in-a-million coincidence?

Well - oh no, look at that, I asked it for a simplistic character drawing and it happened to me too, it just returned a drawing of mine that I never even uploaded, and it's the worst drawing I've done since the fifth grade even just to embarrass me! Oh no, what happened, did they change things right under my nose, has digital surveillance gotten even WORSE?? Look, see, here's the original on the left, compare it to the output on the right - scary!! They're training on the contents of your computer in real time now, aaaagh!!

Except, of course, for the fact that the entire paragraph above was a lie and I did this on purpose in a way no one could possibly recreate from a text prompt, even with a perfect description.

How?

See, some models have this nifty little function called img2img. It can be used for anything from guiding the composition of your final image with a roughly drawn layout, to turning a building into a dragon...to post-processing of a hand-drawn image, to blatantly fucking lying about how AI works.

I took 5 minutes out of my day to crudely draw a character. I uploaded the image to this post. I saved the post as a draft. I stuck the image URL in the init_image field in Simple Stable, cranked the init strength up to 0.8, cleared all text prompts, and ran it. It did exactly what I told it to and tried to lightly refine the image I gave it.

If you see someone claiming that an AI stole their image with this kind of "proof", and the image they're comparing is not ITSELF a parody of an extremely well-known piece such as the Mona Lisa, or just so extremely generic that the level of similarity could be a coincidence (you/your favorite artist do/es not own the rule of thirds or basic fantasy creatures, just to name one family of example I've seen), this is what happened.

So from here you must realize that it is deeply insidious that posts that make these claims usually imply or even outright state that you should NOT try to recreate this but instead just take their word for it, stressing ~DON'T FEED THE MACHINE~. It's always some claim about "ohhh, the more you use them, the more they learn, I made a SACRIFICE so you don't have to" - but txt2img functions can't use your interaction to learn jack shit. There's no new information in a text prompt for them TO learn. Most img2img models can't learn from your input either, for that matter! I still recommend being careful about corporate img2img toys - we know that Facebook, for instance, is happy to try and beef up facial recognition for the WORST possible reasons - but if you're worried about your privacy and data harvesting, any given txt2img model is one of the least worrying things on the internet today.

So do be careful with your privacy online, and PLEASE use your very understandable knee-jerk horror response to how much extremely personal content can be found in training databases as a call to DEMAND better privacy laws ("do not track" should not be just for show ffs) and compliance with security protocols in fields that deal with very private information (COMMON CRAWL DOESN'T GO FAR OUT OF ITS WAY, IT SHOULD NEVER HAVE BEEN ABLE TO GET ANY MEDICAL IMAGES THE PATIENTS DIDN'T SHARE THEMSELVES HOLY SHIT, SOME HOSPITAL WORKERS AND/OR MEDICAL COMMUNICATIONS DEVELOPERS BETTER BE GETTING FIRED AND/OR SUED) - but don't just believe a convenient and easy-to-disprove lie because it aligns with that feeling.

[da-sanityismyvanity]

Visit my DeviantArt profile for my full gallery, new images added every week

[vestaignis]

Кои освещенные солнцем. Koi carp are illuminated by the sun.

[stablediffusion]

Cute Pinups (workflow in comments)

Give us a follow on Twitter: @StableDiffusion

h/t itsnotrealx

[oldion]

Oil painting of a woman on a steampunk ship through a flooded city. I was inspired by Mortal Engines and Dishonored.

[catkiinson]

50’s Furbies.

[auradecristal]

[fakefrootart]

Biker future club

[dalle2]

Seinfeld episode where Kramer installs a toilet in Jerry's doorway as a shared building amenity, "The Convenience", season 2, 1991

Generated by OpenAI's DALL·E 2

Sharing is caring!

Substack: dalle.substack.com

Twitter: @Dalle2AI

The heading of this post was used to generate the image, src

[gay-i]

[lexart-io]

GENERATIVE AI ART 〄 LEXart.io

[adnerwbell]

Some randomly arranged drawings from a project I released on fxhash.xyz called ‘Borders’.   8 layers to create one image, pulling from a pool of 100 layers. Some of my favourite and most playful work from 2022 :) 

[da-sanityismyvanity]

Visit my DeviantArt profile for 6000+ images and my Redbubble store for clothing, prints and more, new images added every week