I wanted to ask you guys a 100% newbie question, regarding AI-generated art…
Hey everyone.
I’ve seen the rise of AI-generated art becoming more and more “convincing”, sometimes photorealistic, sometimes with adorable manga vibes, sometimes looking like it’s inspired by a precise artist or to the contrary looking unique…
And, damn, that’s impressive. Just one year ago I’d have thrown the same joke as everybody, about the hands, about how the drawings seem hollow and soul-less, but nowadays, seeing as it borrows actual inspiration from actual people, and (huge airquotes) “simply” recombines it, I’m not pessimistic anymore. There will be greatness, eventually.
Which leads me to the question I’m asking myself, as I have very little knowledge of how the whole thing works, save vague general notions. Could I ask for a, well, a clarification, from those of you who are into that?
To be super blunt: is it something that can be done, to feed hundreds of drawings by one precise artist to a piece of AI software, have the software munch out those drawings, and then, bam, it will spit out, on demand, art looking like it’s been drawn by the same artist?
If the answer is “yes”, I’m definitely trying that with Gunma Kisaragi, and asking for all kinds of art, from softcore to hardcore
Of course, I suspect it’s far from being so easy.
Maybe a hundred images isn’t enough, and it needs a thousand times more materials. Or maybe the images have to be perfectly cropped for the software to know what to do with it. Or maybe the images require that a human types a precise description of what’s in each image. Or maybe it demands an incredibly powerful PC and the electricity budget of South Africa. Who knows.
But well, here’s for the question. I could totally go for generating art by my favourite artists like that, if it were doable, huhu.
Thanks if those of you who know the stuff can answer! ^___^
*
A quick temporary answer: thank you very much for your answers already guys!
It gave me some googling to do, I’ll answer later when I have smarter (no, scratch it, less dumb) answers to make!
My understanding is that it is possible but will require a machine with a significant amount of power. Might want to check out some of the online AI sites if you want to experiment. Many of them don’t allow explicit images. Maybe try out Stable Diffusion if you’re curious. https://stablediffusionweb.com/#demo
I tried setting up one of these model on my local machine and it’s still quite cumbersome.
if you check out the YouTube video from the corridor crew Rock Paper scissors Anime they do a nice job showing the potential of AI generated art and a tutorial of how they did it as well it should answer some of the questions you have here’s the link to the video
https://www.youtube.com/watch?v=_9LX9HSQkWo&t=276s
The way it works is you train the model on “All hentai”. I.e., everything you’ve ever posted. Then you give it instructions to “draw like a specific artist a story about X” and it’ll draw something now. As long as “all hentai” is tagged with the author’s name.
Some artist names are “recognized” by certain models but I doubt Gunman Kisaragi is among them. A LoRA can be loaded together with a larger model in Stable Diffusion to imitate an art style (the quality and fit of LoRAs varies just as models can vary). With a very shallow search I could not find a LoRA for Gunma Kisaragi.
A LoRA can be trained with <200 images, but each image still requires labels/tags to pair with them (more details in the link), and probably 8 GB+ of VRAM at the minimum. In general, you’ll have an easier time with Windows/Linux and NVIDIA GPU. Cloud computing services are another option, free within some restrictions, but I have not explored that option at all.
I have not done training myself, so I cannot offer tech support, and there may be better guides out there. https://rentry.org/lora_train#training-loras
Even if your PC or cloud can’t train a LoRA, it may be good enough to generate images with Stable Diffusion. My AMD RX 580 4GB can generate a 512×768 image in 2~3 minutes. There are plenty of guides out there for installing and using SD with a graphical interface.
CivitAI and HuggingFace are two sites where you can find models and LoRAs, including NSFW ones. Unsurprisingly, there are a lot of NSFW Anime models.
Le fait de voler le travail d’artistes pour le faire “digérer” par ces IA est à la fois incroyablement insultant et immoral. Je m’étonne que pratiquement tout le monde trouve ça normal de se servir et de piller sans vergogne pour produire de vagues dégueulis de pixels…
Que voilà une diatribe à la fois inutile, ampoulée et bien vénère pour rien.
1- Insultant? Non. Apprécier un artiste au point de vouloir générer des images qui rappellent son style, c’est en soi une forme d’admiration. Tu veux qu’on parle de tous les détournements de la Joconde à fins publicitaires ou humoristiques?
2- Immoral? Par pitié, les leçons de morale à deux balles, il suffit de se baisser pour les ramasser. Si tu vas par là, tous les pseudo musiciens qui font du samplig sont à écorcher vifs en place publique vu qu’ils font du pèze littéralement avec la musique des autres. Je ne vois pas d’exécutions avoir lieu…
3- “se servir et piller”? Bon déjà, si tu viens télécharger ici des copies de manga et doujins, par définition, les auteurs et leurs ayant-droits pensent déjà ça de toi. Hypocrisie very much?
4- “de vagues dégueulis”? Mon pauvre, l’argument tient de moins en moins, car les IA s’affinent de plus en plus et les promptages sont de mieux en mieux maîtrisés. On commence déjà à voir des résultats bluffants.
Le débat aurait une quelconque valeur ici si Oliver voulait plagier un artiste et revendre le résultat comme étant son travail à lui. Là, oui, ce serait un piratage sans vergogne. Seulement voilà, il s’agit ici d’un intérêt purement personnel, pour son plaisir à lui. Il se dit “et si Gunma avait dessiné une page où il se passe ceci et celà, ça donnerait quoi?”. Enfourner le truc dans une IA pour voir une approximation de ce concept n’a rien de répréhensible en soi.
Le vrai débat n’est pas dans la moralité ou l’immoralité de l’utilisation des IA à fin d’imitation. Par ailleurs, l’imlitation et l’émulation n’ont pas attendu les IA…
Que des artistes puissent geindre quand un quelconque zozo plagie leur style et poste dsur DevArt ou ailleurs en se vantant de leur “oeuvre”, c’est totalement légitime, surtout s’ils ne créditent pas l’artiste imité. Jeter des cailloux à quiconque a envie de faire mumuse avec l’outil pour leur propre distraction, c’est juste de la bien-penseance à deux ronds.
Tu te contentes de répéter le discours bien rodé des corporations qui développent ces soft, mec (oui, ils ne sont pas créés dans un garage).
Je n’ai rien contre l’IA en tant que tel . Mais en l’état actuel, les IA génératives sont du VOL pur et simple. C’est peut-^tre un joujou pour toi, mais pour les artistes il s’agit de leur vie, de leur oeuvre, de leur salaire.
La bien-pensance te fais la bise.
Et voilà, lopokonpri. Le souci c’est encore et toujours ce que tu fais de ce que l’IA a créé. Histoire de te rendre la chose compréhensible, je te renvoie à ce vieux proverbe informatique, sans cesse vérifié:
“le problème se situe entre la chaise et le bureau”.
L’IA générative n’est pas du vol. Utiliser une IA générative pour pomper le style de l’artiste et faire du fric avec, oui. Par conséquent, tu devrais plutôt aller pourfendre tes moulins à vent sur des plateformes où cette pratique a cours. Je pense que de plus en plus, on en a assez des offusqués professionnels.
*Mic drop*
You obviously subscribe to the concept of creator’s rights, in perpetuity, per the EU.
This is a double edged sword: Sure, artists benefit by being able to enjoy the monetary benefits of their creations (and sometimes an artist only has one great creation in them, so they obviously want to be able to make a living off it), but being able to control the derivative uses of your work in perpetuity is exactly what the corporations want and have been using to extend rights and control of works to the point where you DON’T own anything, you only LEASE it from them.
For some idea of really good stuff being done, search for Redoxa – AI collab – he’s taking his old 3D stuff and reworking it using the AI and it is really photo realistic.
The real issue I foresee coming down the line, is for work akin to some of the manga here – it will skirt/cross the line into being deemed illegal b/c it’s too real looking.
A quick temporary answer: thank you very much for your answers already guys!
It gave me some googling to do, I’ll answer later when I have smarter (no, scratch it, less dumb) answers to make!
Good evening. I feel like I’m going to be confused but I’ll try to give you my opinion. Frankly it’s true that it’s impressive but go to PIXIV and look. The images all look alike. They can be a lot of differences but you don’t feel the paw that a designer can have. I have never seen an image having the style of… On YOUTUBe you have “what if starwars was a dark fantasy film from the 80s”, then robocop, then warhammer, then etc etc at the end you have the impression of always seeing the same thing.. An artist that I sponsor tried but he decided to use AI to make backgrounds in which they put his characters. We all said “phew” So for the moment AI has a “manga style” or “hyperrealistic style” but does not yet know how to copy or take over an artist and I want to say fortunately. Otherwise where is the point of drawing and having a style. Look at you it’s stories of KISARAGI me it would be this late MORIYA MAKOTO. But does the drawing also make the story? A drawing is a line, a pressure, a quality of gesture, a choice of color, a sensitivity. Try to find AROMA SENSEI’s latest “RAVEN” drawing. No machine can draw that. And if so, then we’re finished the devil will have won our souls (it’s an architect who tells you that) For the budget and the machine try to find the artist “ZUZU” on Pixiv. He has to do at least 30 images a day and I don’t think he has a nuclear power plant at home.
True that, nothing beats the real thing. A machine can only provide an approximation and can only work with quantifiable data, which rules out artistic sense. Just yesterday evening I was watching a streaming by an artist who was drawing a page for his next doujin and it’s really a captivating thing to behold. I don’t want that to ever disappear
TRY this when the server is not too busy.:
https://pornpen.ai/make…bye
At this time its not entirely possible to do what you ask for Oliver but giving how fast AI art nowadays progresses I doubt it you will need to wait much longer to do what you want, in this case with Gunma Kisaragi art. I mean, even some of the artists recognize this idea to ‘help’ them create more content, like cuboon who seems keen on the idea (so he would use his art to generate more art etc.. like, the idea you’re asking Oliver but artist himself would do that with their own creations.. and if going by Twitter it seems he already started it?):
https://twitter.com/EmblemOfAcumen/status/1632671597416058880/photo/1
Anyway, the last year AI art made so much evolution – I remember ppl posting on some forum some shitty AI generated art and one year later you see really good stuff. I can only imagine what the future holds (and no, I’m not saying artists will be left aside but I think its going in the direction that ANYONE will be able to create art so its bonkers). Yeah, talent is talent and whatnot but I really didn’t expect one year ago that I will be seeing so much good AI generated art today. And its been only one year! One year!
I will try to do short and concise.
Yes it is possible.
Those AIs are like some autistic genius children. May look dumb but are highly specialized in one field.
It mathematically simply copy the inner working of a brain synapses to learn.
So, the same you would teach a child, you show it something and give it a name so it remembers it.
For AI images, that’s giving it labeled images. So you would have to find a lot of art from someone or multiple people and manually label everything.
Then come the teaching part where the AI will mildly sucks at it at first then became better then will stop trying to guess the answer but will rather pick an answer amongst what it knows. That’s over learning, means it’s too much, as if you always picked the same pics of a blue car to show to a child and saying this is a car for tens of year. The first time he sees a car not blue, then it’s not a car.
The ideal would be to have bland templates of simple drawings labeled with poses, direction, ambiances, humor, acts, etc then try to overwrite this with an artist style from different angles.
I don’t even want to guess how much work it would be to design something like this but that would be impressive. Impressive and scary like “hey, you want to read a comics drawn by this artist dead 70 years ago with new stories every day taken from existing novels? No problem”
Your question about Stable Diffusion opened my mind for some possibilities. But since I’m total noob about it, I started looking for some help, tutorials etc. So I found one (it was Lora), after some troubles I managed to install everything (which is also not that easy) just to realise, my GPU is shit, and I can only dream about making some Stable Diffusion Content on my PC. Also the whole prompt text to image is something you have to learn, if you don’t want to create some horrible pictures. So I started looking for some options to try it out via cloud system. Tried huggin face, but they don’t have nsfw content. I also tried google colab, but it was also not good solution. So finally I found one, that works via cloud system GPU, has nsfw content enabled (hell, nsfw is their main reason to make this site), has huge amount of model systems. They also have great tutorials, how to properly prompt image, how to edit it, improve it. They even have prompt for noobs that helps you add proper text to prompt, so pictures could look better and desire your expectations. They also help you on reddit, as well as they have assitants on the site. The only issue is, they have subscrition, so you have to pay for it every month, if you want to use it.In my country the cost for month is about 12 USD. You can also test it for 7 days for free, but without nsfw content.
https://www.piratediffusion.com/
In the end I wanted to tell you, I’m using hentairules for so many years, I wanted to thank you for all those years. We came to pretty poor times, when the best ones stopped making hentai and doujins. I miss Kisaragu just as you are, I suppose. But yeah, just wanted to thank you for all those years.
Oliver, I think you’re right. As long as you train the AI with Gunma drawings and representations of what it should draw. But will that be available to us?
Going by rule 34, someone is already working on it: “I’ll have my AI make custom porn and sell it as a service.” Then, “I’ll add NFTs.” Then, “But if the NFT points to my system, will we get sued for using others’ art?”
So I’d say, “Not yet, but soon,”and it’ll be messy.
@oliver
Please check it the discord “unstable diffusion” they have essentially already done what you’d like to varying degrees of success and even p provide resources and instructions to help you do it.