- Joined
- Jul 26, 2017
At this point why not save everything as a low quality JPEG so it has artifacting? It seems like that would take less time and have a similar effect.
AI Art Seething General
- Thread starter The Thinking Abacus
- Start date
- Joined
- Aug 28, 2015
It's not as performative, duh.
- Joined
- Jan 28, 2018
Mostly experimented with finetuning, improving the writing of llms and loras until recently, even though it's an old example, making a model consistently believe it is 1989 is still my fondest memory in this sphere, because the technical execution went relatively smoothly and the dataset was tiny for its effect. I left the AI sphere for a bit for other hobbies and I am still catching up. A month in AI time is like a year in other things. I went back to emacs recently because python is horrible and all llm related frameworks are a rats nest of poor decision making by developers and lisp is great for complicated decision trees and emacs/elisp is great for working with text. I'm now in the process of teaching a group of smaller LLMs to work together as one brain to "win" simple text games written in lisp. I'm really interested in that interesting sweet spot of problem solving LLMs can reach that is not easy to implement with more "conventional" programming approaches. I feel it could make for very "lifelike" videogame AI. That said, LLMs soon won't be the cutting edge anymore, multimodal is the future.
There is some pixel art I generated and posted before. For anyone with a passing interest in pixel art, it's easy to see how these pictures could be fixed to not look AI made anymore. It is a pity about Stability, but there are currently very interesting things brewing, some of them coming from China who are far less concerned with censorship when it gets in the way of efficency, funnily enough. Especially imagegen is nowhere close to hitting the ceiling. I was thinking about how to replicate 80s high color pixel art that contrary to this weird autistic current pixel art with its billion drawing rules usually was the result of digitized pictures being traced over with a very limited color palette, then LLMs got more interesting.
It started way before photoshop. I got video hardware lying around here which would give Amigas (who were limited to maximum 32 colors out of a 12 bit palette, 64 if you stretch things with halfbrite, so not really enough for any kind of pleasing digital art that is not pixel art) millions of colors to be ran with bespoke graphics software, specifically made for the hardware. There was quite a bit of hardware like this because people clearly saw the potential for "computer art" and animation and these programs were less "abstract" than modern software, still heavily leaning on the physical process of painting to ease people into working with computers. (which how a lot of computer software was and still really is, leaning on physical world concepts the 20-something twitter artists haven't even experienced because they were already obsolete when they still shit their diapers) It'd still take 5-10 years from such rather expensive and for all intents impractical to use hardware expansions until digital art was actually anywhere close to common place and absolutely nobody remembers the pioneer stuff now. That's why it's naive to think things have peaked, thinks didn't peak with that not-that-useful hardware either. The ideas and concepts are formed. The potential is clear. It's an engineering/scaling problem now. People that think technological advances that have already proven useful to this degree will just suddenly "disappear" and the world will just turn back a few years are simply delusional.
std::string
kiwifarms.net
- Joined
- May 17, 2018
AI doesn't care about artifacts.
Glaze "works" by adding guided noise to the image that tricks the neural network training on the image into following the wrong path.
YumeNoHero
kiwifarms.net
- Joined
- Feb 12, 2021
Not sure if this thread is proper for this but I'll ask anyway:
Most of the AI art I saw that are not just straight up text2img, they used some sort of rough sketch before letting AI handle the rest. How much can the AI do if I feed it a 3D render scene or a 2D image representation of it and tell the AI to generate something? I am only aware of the usual image generation and haven't caught up with the recent developments of generative AI yet.
Most of the AI art I saw that are not just straight up text2img, they used some sort of rough sketch before letting AI handle the rest. How much can the AI do if I feed it a 3D render scene or a 2D image representation of it and tell the AI to generate something? I am only aware of the usual image generation and haven't caught up with the recent developments of generative AI yet.
- Joined
- Dec 12, 2022
Read up on the various models for Controlnet here:
https://github.com/lllyasviel/ControlNet?tab=readme-ov-file#controlnet-with-canny-edge
The examples alone should give you an idea. But yes, from depth maps to really shitty MS paint scribles, those can all work to guide the generation. Hell, there's an openpose plugin for Powerpoint. Posing skeletons in Powerpoint.
Is $13,500/month a high server cost for a website with 300k users? Because it only takes 1% of 300k users to pay $5 a month to cover those costs. I think the website will remain pretty stable for the time being. Or is expecting 1% of users to donate too generous? What I'm confused about is if it was $2,200 per month when it was only 100k users, how did it blow up by a magnitude of six in server cost when the userbase only tripled? It seems like server costs don't grow linearly with user growth.
- Joined
- Apr 20, 2021
I'd happily take AI artwork existing over them not existing. I mean look how gorgeous this one looks.
The world already has as much protective measures as possible, but human created art never died despite the advent of AI. If not we could experiment with AI and use it as reference to better our artistic abilities like PewDiePie has. It's a win - win.
The world already has as much protective measures as possible, but human created art never died despite the advent of AI. If not we could experiment with AI and use it as reference to better our artistic abilities like PewDiePie has. It's a win - win.
- Joined
- Apr 10, 2022
I'm no server expert and would love to have @Null make a more qualified guess, like he did during Gabs server misadventures. For a website that only hosts images and gifs, paying $13,500 a months, just for the storage part, seems like they're getting shamelessly scammed, by whatever cloud host they're using.
I bet they could buy all the servers and infrastructure they could ever need and much more, for $13 - 25 k one time and just have that running in a data center somewhere, next to no monthly costs.
Unless you're making some crazy livestreaming project, adding users should not be this much of a problem. I'd kill to have some internal analysis of this, to see if the website has some crazy flaws, or if the Glaze thing being ran on the images, is the root cause, of all their problems.
- Joined
- Jan 28, 2018
Maybe you think too complicated. Maybe they're just embellishing the costs to get people to pay more and pocket the difference. Artoids are natural born scammers.
- Joined
- Apr 10, 2022
I don't think so. Zhang Jingna is Forbes 30 Under 30 Asia list, friends of the journos, currently suing Google and many other companies on the side, levels of rich.
Therefore I can easily believe someone at Amazon (or similar scam host), just quotes her the worse possible hosting plan, that she pays without questioning and asks the community if they want to contribute, without much care, if profit is made.
Big happening - maybe
Illinois Senate passes artificial intelligence protections for artists
"The Illinois Senate approved a bill Friday that would allow artists to sue entities that replicate their work through artificial intelligence without their consent."FULL ARTICLE HERE
Last edited by a moderator:
- Joined
- Jan 28, 2018
Quite possible, I have no idea who that person is.
but it mostly doesn't? If you don't basically break the model by overfitting, you can't really replicate specific art of an artist, even if you try very hard. The model will always generate something novel. I don't think you can protect non-specific styles like this, otherwise brb I'll get sunsets and 2D drawings of women copyrighted.
Conspiracy Mawile
kiwifarms.net
- Joined
- Nov 20, 2023
My guess why most professional artists are bothered and freaked out about AI is not because it can generate media of "better" quality than what they can produce. The main selling point of AI models is (in my amateur opinion) their speed. One artist needs like 2 hours for a nice commission, AI can make a bajillion images in that time. And now it's not your artistic abilities, techniques and talent you have to use to compete with other artists in the marketplace, it's your programming skills and the hardware you can afford.
If I had studied drawing for six years or so and than be replaced by some nerd who doesn't know how to draw a straight line and care about color theory but can code really well and owns the latest hightech I'd also be quite pissed.
In the end, you have to sell your art to make money out of it, but I guess what will be sold is not the AI art, but the AI program and who wants to spend 50 bucks to commission some furry artist when you can just use the latest AI model to do it in 3 seconds?
If I had studied drawing for six years or so and than be replaced by some nerd who doesn't know how to draw a straight line and care about color theory but can code really well and owns the latest hightech I'd also be quite pissed.
In the end, you have to sell your art to make money out of it, but I guess what will be sold is not the AI art, but the AI program and who wants to spend 50 bucks to commission some furry artist when you can just use the latest AI model to do it in 3 seconds?
If you are a digital artist who used any software to do your work, your work is probably already reliant on the work of some nerd who doesn't know how to draw a straight line and cares about color theory but can code well and owns the latest high tech to begin with.
- Joined
- Oct 6, 2016
I wonder if they get mad at the machines mixing paint or making pencils.
Conspiracy Mawile
kiwifarms.net
- Joined
- Nov 20, 2023
Good point, humans could have mixed that paint or made those pencils and their jobs got taken away by machines, we should abolish them too.
std::string
kiwifarms.net
- Joined
- May 17, 2018
Sounds about right if a bunch of the new users migrated all their art over and they're running Glaze on everything that gets uploaded. The number of images they're hosting would correlate more directly with costs than the number of users.
If you want something specific you'd still have to commission your favorite furry porn artist. I don't think this affects commissions at all, actually. The people dropping a tenner on runpod.io to blast out thousands of furry porn images aren't really your client base.
Golly Bo-Bolly
kiwifarms.net
- Joined
- Oct 5, 2023
Either these people are getting scammed by their ISP, DNS, Glaze, or these people don't know how to manage an image site. Cara hasn't even broke a million users yet.