Why don't indie devs use AI-generated images as art?

[Üstad]

https://www.jesserockwell.com/biology-of-vesarion3

https://www.jesserockwell.com/vesarion-3-ai-project

Using Midjourney for his work. Artstyle and the work of imaginary alien planet is consistent.

 

[V17]

Interactive SD in Photoshop. Pretty interesting, basically a variation on photobashing, only interactive, stylized and less detailed.

 

[0wca]

Goddamnit and here I am lurking for hours on end for royalty-free or atribution art for my indie TTRPG. Never heard of this before.

Can someone paste the websites where it's truly public domain or free for commercial use?

I already got all the artwork but I'll definitely exchange some of the weaker art for this.


EDIT: This DOES bother me though...

https://kotaku.com/ai-art-dall-e-midjourney-stable-diffusion-copyright-1849388060

 

[Üstad]

Goddamnit and here I am lurking for hours on end for royalty-free or atribution art for my indie TTRPG. Never heard of this before.

Can someone paste the websites where it's truly public domain or free for commercial use?

I already got all the artwork but I'll definitely exchange some of the weaker art for this.

https://beta.dreamstudio.ai/

I think this one has released source code and you can generate your own images if you have capable GPU (4 GB video ram atleast). Other than that you can make account and use free trials that is renewed for every different account. You can just sign in with discord account or gmail. This is the one using Stability AI, the one that the video above using.

I can't open the youtube video at the moment, just in case if the video is dead I'm dropping the r*ddit link on here. It's the same video.

https://www.reddit.com/r/StableDiff...rstablediffusion_integrating_sd_in_photoshop/


Midjourney is more quality, has limited free trials but you can generate limitless amount of images if you pay for monthly subscription. I think the copyright laws are rather blurry on this one, because AI can't be owner of intellectual property. But I'm %99 sure they want to be credited for images you generated and all of the generated images are saved in their database. But they can't do shit, if you edit the images and merely use it as baseline for your work, not sure but just guessing though.

https://www.midjourney.com/home/

Here is the site. Create discord account and click the option "join beta". You can generate your images with going on "newbie" rooms (ie: newbie#47) and use the "/imagine" prompt without quote marks. It has 20 uses for every discord account.

I hope this helps a bit, but of course I strongly advise you to double check and ask around.

 

[Nathaniel3W]

I've been posting my adventures down in the Science subforum. Vanilla Stable Diffusion didn't really work for the kind of art I wanted. So I got Textual Inversion up and running on an Amazon Web Services instance. I messed around with it a couple days ago, but today I am really training it hard using some tips I got directly from the author of the TI paper. Just to give you a tiny insight into what it can do, this is what it created 10 minutes into training it on the Naoki Ikushima collection I put together:

I didn't give it a prompt. I didn't give it an image to flesh out or to improvise on. This is the training output so it can show me what it thinks the essence of the concept is.

After 15 minutes, this is what it thought Naoki Ikushima meant:

20 minutes:

Once again, this is just a byproduct of training so it can show me what it's learning. I'm not giving it any prompts to work with.

After that I trained it on a couple of specific video games. Right now I'm training it on Akihiko Yoshida. And pretty soon I think I'll be ready to start cranking out character designs and promotional art for my next game.

Edit: And here is where we are so far on Akihiko Yoshida.

I think Ikushima's style is just easier for the model to grasp for some reason. TI picked up on Ikushima right away. I'm over an hour into Yoshida, and I think it's just now starting to get the hang of it. Not too many iterations ago it was just drawing lots of boobs. One image was downright pornographic with one character's armor consisting of panties around her thighs.

 

[Davaris]

https://www.jesserockwell.com/biology-of-vesarion3

https://www.jesserockwell.com/vesarion-3-ai-project

Using Midjourney for his work. Artstyle and the work of imaginary alien planet is consistent.

I'm impressed. It reminds me of 1970s fantasy sci-fi.

There's a new kind of AI in development that makes use of human brain cells. It trains thousands of times faster than computer AI, and uses far less power. So in the future you will keep your artist in a jar, and feed it sugar water.

https://www.youtube.com/watch?v=FuzoLdrRX5Q

 

[V17]

Interactive SD in Photoshop. Pretty interesting, basically a variation on photobashing, only interactive, stylized and less detailed.

Well the dude uploaded it to youtube and then apparently deleted it for some reason that I'm sure is stupid.
Didn't want to link to reddit, but it seems to be the only place where the video is online: https://old.reddit.com/r/StableDiff...rstablediffusion_integrating_sd_in_photoshop/

 

[mkultra]

Why? Because good synthesized images began with Disco Diffusion ~early this year, a game takes longer to be released or even to be shown for the public so you won't see those games yet.
Also if they started using it perhaps 8 months ago it's already outdated as hell, i mean not just a little, completely outdated as in likely better to scrap all of it and start over to get better quality/art.. Every month this year has been a tremendous update in terms of consistency.

But resolution is very limited so that's an issue. There's good AI for upscaling, but 512x512 upscaled to perhaps 2-4K or so will still look upscaled. The models are trained at 512 pixels, if you go above it funky stuff will start to happen (extra limbs etc)

Some dwarves i did today. I make many hundreds, if not thousands of images a day (for my mods, or just for fun and various projects i do). It's much better now, but not perfect, obviously it only takes 2-3 seconds to fix most of the issues in photoshop. Also i prefer to render heads and then paste them in, the quality is tons better because it has focused on making just a head. These are completely unedited

You can also use an "init image" - so you can use an image of yourself or pose an action figure and tell the AI it's something else (e.g an orgre by Frank Frazetta, or whatever) and it will regenerate that into art which is un/recognizable (from 1-100%) from the original photo or art you usd. By doing this at perhaps 50% you will gain a ton of consistency in the quality (no extra limbs or other funky stuff), so if you're looking for a certain pose or whatever this is way easier than telling the "AI" to pose your character.

From my pathfinder mod (i'd say 50% of those are quite edited in photoshop)

https://www.nexusmods.com/pathfinderwrathoftherighteous/mods/353?tab=files

 

[mkultra]

Why not skip the middleman and just generate entire games?

It's already happening, but obviously you can only tell it to generate very simplistic games. But yes, normal language works, no programing required at all.

For 3D models its happening too, which will be really fantastic when it gets really good.

This is changing the entire industry, no doubt about that, fantastic art can be generated by anyone. A 2D indie game by just one dev has no excuse to look shit in 2-3 years time, they can all look like they hired the most expensive artists possible.

I can't wait until animation gets really good too, because we can already move around in these generated landscapes..

I did this animation early this year(so its outdated) https://www.youtube.com/watch?v=vjctazDDfOU&ab_channel=Vurt_NexusMods

 

[Gastrick]

Many of these images are unusable unless the dev knows how to redraw faces or hires someone to do it.

 

[Peachcurl]

Why not skip the middleman and just generate entire games?

It's already happening, but obviously you can only tell it to generate very simplistic games. But yes, normal language works, no programing required at all.

For 3D models its happening too, which will be really fantastic when it gets really good.

This is changing the entire industry, no doubt about that, fantastic art can be generated by anyone. A 2D indie game by just one dev has no excuse to look shit in 2-3 years time, they can all look like they hired the most expensive artists possible.

I can't wait until animation gets really good too, because we can already move around in these generated landscapes..

I did this animation early this year(so its outdated) https://www.youtube.com/watch?v=vjctazDDfOU&ab_channel=Vurt_NexusMods

Looks amazing. Well done.

 

[mkultra]

Many of these images are unusable unless the dev knows how to redraw faces or hires someone to do it.

what you do is generate faces on their own, or use an init image and they get perfect. then you colormatch in photoshop, paste it over the glitchy face. there's also face restoration AI's, they also work pretty well, but they only do photorealistic faces, not cartoony or painted.

 

[dbx]

text-to-image models is yesterday's stuff.
The next new thing is obviously text-to-nerf

 

[Davaris]

I wonder if AI generated art is public domain, will games made using this technique also be public domain? Their art assets probably will be, and art seems to be more valuable than code these days, when it comes to IP.

 

[mkultra]

I wonder if AI generated art is public domain, will games made using this technique also be public domain? Their art assets probably will be, and art seems to be more valuable than code these days, when it comes to IP.

A game is more than its art, also its unlikely that everything would be AI based or that someone could prove that it is.. usually you will spend a lot of time editing it, and if you do it's no longer just AI, it has a owner and a creator, and he/she will have a right to it.
But yeah for pure AI images yes, but it depends on what you're using too, there's not just one text to graphics API, and there's also not just one law (e.g US law) so it might get complicated.

Overall this is quite similar to sampling. For us who followed it from the start we see the same scepticism, same fears, same issues with copyright etc.. It really changed everything for music and so will this for graphics. Great looking graphics (for anything) will be a norm just like good sound quality is a norm, it's taken a long time for graphics to catch up.

Also this is of course just the start, what we have now lol.. yes, we can come back to it and have a good laugh in a couple of years. The first tracks made with samplers were incredible back then ("woah! those 8-bit drums sounds exactly like the real thing!"), now it's like a bad joke, this will be similar to that (but we don't see it yet because we don't have the future references, obviously...)

 

[Davaris]

I've seen people make similar arguments about copying copyrighted art, arguing if they change it by a certain amount, it no longer belongs to the original creator. I'm not a lawyer, but that argument didn't convince me. Your argument is similar to that.

Making music using cuts from other people's music used to be profitable, until the music studios took it to court, and made it unprofitable.

I think the big studios won't touch this with a barge pole, because of the legal issues. It will be fine for open source projects and free games. Also, enforcing copyright will get harder, because how can you prove you made something without using AI, unless you document every part of the development process?

On thinking about it, I wouldn't pay for a service where I can't sell what I made with it. On the other hand, I would use a free AI art generator for free projects, because its output isn't worth anything if it is taken to court.

 

[V17]

Making music using cuts from other people's music used to be profitable, until the music studios took it to court, and made it unprofitable.

That did not happen. Sampling was a legal issue almost from the beginning and while lawsuits happened and made things more difficult, they never stopped it from being used. It arguably became more normalized in the mainstream after that and certainly still is a firm part of some genres.

I think lawsuits of that sort are inevitable sooner or later here. Since the issue of how derivative the AI products are or how different that is from a living artist simply taking inspiration from another living artist has no clear answer, I suspect the case might come down to regulating what media are the AI companies allowed to use to train their models. Which has the potential of slowing down progress for a few years but hopefully wouldn't kill it either.

 

[gurugeorge]

Sampling was a legal issue almost from the beginning and while lawsuits happened and made things more difficult, they never stopped it from being used.

The real effect of the legal stuff around sampling was to dampen creativity. Certainly if you had the money, time and energy, and/or a record company/good lawyer behind you, you could probably afford to reach an agreement that let you use some snippet of someone else's music. But there was a whole field of music that was opening up in the early days of sampling that involved mangling recognizable chunks of other peoples music, and that was pretty much stopped by the legal people getting their shit together.

I suppose it depends on one's stance re. IP whether one thinks that's a good or bad thing, but it was a definite, notable effect. On the one hand, there's something a bit off about riding on the impact and recognizability of something people have already heard (or profiting off someone else's talent, if one wants to think of it that way), but if and when it's truly used in an artistic way, it's like another level of art, in which the signs floating around in society are re-purposed to artistic ends. And that's very cool, and the potentiality of it was lost.

 

[mkultra]

I've seen people make similar arguments about copying copyrighted art, arguing if they change it by a certain amount, it no longer belongs to the original creator. I'm not a lawyer, but that argument didn't convince me. Your argument is similar to that.

It works for music, that's why sampling is legal if you are being creative with it.. but this is different because it starts with belonging to no one (or everyone) and then it becomes yours because you made something creative with it. At least it's how i think it should work. You didn't rip off a specific image or a piece of artwork from someone, it's not how the AI generates images.

You can be very, very creative with AI, i have some AI stuff that i don't really want to share how i did, like one type of art where i'm mixing 3 specific artists with a specific prompt, took me around 2 weeks to get right. It creates very unique artwork. It's not much different from how someone comes up with a melody or perhaps a beautiful piece of music with a certain set of samples from various artists, its more than just stealing, there could be months or years of work behind it. Some of my favorite albums are created using only samples (Burial and DJ Shadow albums, for example) the effort behind those albums is tremendous and super creative.

With the AI i'm mainly using right now, yes you can sell it, but so can everyone else.. It's.. interesting, but not without issues i guess. Let's say i spend an hour fixing an image, using tools i've paid a lot of money for, perhaps adding some 3D artwork to it that i've created or similar + using photoshop that i've bought, it's still everyone's? I think the law would disagree.

 

[Davaris]

Consider you took images from a Bethesda game and spent a lot of time altering them and then put them in a commercial game. Do you think Bethesda would sue you? They would sue and win. I think the same would apply in reverse.

The thing that concerns me is you can't copyright rules for a game, because games aren't recognized by the law as useful inventions. So people can clone your game, but what slows them down is they must make their own art, because that can be copyrighted. However if you use AI generated art all they have to do is reverse engineer your pack file format, and use that to make new games that look exactly like yours. The only thing other than art you can own the rights to in your game, is if you trademark certain names, and the text you wrote.

I think they should let the choices people make when using AI be copyright, because choosing good from bad is a skill. But first we'll need smarter AI that can train and learn without scanning the entire Internet. When that happens they may successfully argue in court for copyright on AI generated art.

 

[mkultra]

Consider you took images from a Bethesda game and spent a lot of time altering them and then put them in a commercial game. Do you think Bethesda would sue you? They would sue and win. I think the same would apply in reverse.

The thing that concerns me is you can't copyright rules for a game, because games aren't recognized by the law as useful inventions. So people can clone your game, but what slows them down is they must make their own art, because that can be copyrighted. However if you use AI generated art all they have to do is reverse engineer your pack file format, and use that to make new games that look exactly like yours. The only thing other than art you can own the rights to in your game, is if you trademark certain names, and the text you wrote.

I think they should let the choices people make when using AI be copyright, because choosing good from bad is a skill. But first we'll need smarter AI that can train and learn without scanning the entire Internet. When that happens they may successfully argue in court for copyright on AI generated art.

No, the closest thing you are doing is an interpretation of their art, which is completely legal. It will have nothing in common with the original art apart from using it as a kind of base, artists does this all the time. LucasArts has been caught a few times tracing fanart and copying it to like 95%.. I've seen the comparisons, you can directly tell.
There's zero chance of Bethesda winning such a lawsuit.

Also it would likely be very hard to tell, unless you go around bragging about "stealing their screenshots from the web" and making your own game from it.

This is their screenshot, and me telling the AI to do various stuff with it,

I could also use like a ground texture from their game and change it similarly.. they would never know, even the original artist of it would not know it unless i'm really trying to make it look like perhaps 90% of their art (it's completely up to the user of how much of a change you want).

The AI learns similarly to how a human learns, by looking and repeating. It will repeat actors and well known people very easily because it's the people it has "seen". I doubt there would be a better way of learning really, it's a very human-like behavior you can say. If you only had seen celebrities and a limited set of art, it's very likely your own art would be very, very influenced by it.

The internet is a fantastic in that way, almost limitless influences of various art and photos. But if you want you could teach it with a specific set of art, perhaps your own, you would need many thousands of images though for a decent outcome. I'm sure we will see that too, trained for very specific types of art and used by studios.

 

[Davaris]

Ignore me. I don't know anything about the law. Ask a lawyer instead.

I just looked up public domain and it says, if you alter a public domain work significantly you can copyright your work.

https://www.quora.com/Can-my-work-be-copyrighted-if-it-is-based-on-a-public-domain-work?share=1

Christian Collovà
Copyright Lawyer, Entertainment Attorney5y
A work based on a public domain work can be copyrighted, provided that your own work (which is called a “derivative work”) passes the so-called “threshold of originality”. That means that your (new) work should be characterized by creativity and originality, that are the imprint of your authorship. Then this new work will deserve copyright protection.

I apologise to academics and fellow lawyers for the simplification that will follow, but this is a handy example I always use when I teach to my IP law students a first approach to derivative works, and the concept goes through well.

Think in layers. The lower layer is the work in the public domain. Think of any public domain work or character you may use for your derivative work. Now you add an additional layer that originates from your creativity (plot, script, other characters, situations, locations, etc.). This layer has to be creative, otherwise it will not be protected by copyright. If creative, then this new layer is protected by copyright. Anyone can use the public domain layer without the need to ask permission by the right holder (that’s the concept of public domain) however no one can use your layer, the one you created.

The same structure applies even if your derivative work is based on a previous work that is still protected. The lower layer is the original author’s layer, then you add on top an additional layer of yours. In this case, not being there any public domain situation to help, whoever wants to use your work shall ask permission to both you (the top layer) and the original author (the lower layer). If you want to use your work (the top layer) by yourself, you shall anyway ask the original author’s permission (the lower layer).

This is why a work based on an original work that is in the public domain can be anyway protected by copyright.

 

[whydoibother]

Here is the site. Create discord account and click the option "join beta". You can generate your images with going on "newbie" rooms (ie: newbie#47) and use the "/image" prompt without quote marks. It has 20 uses for every discord account.

Its /imagine, not image.

 

[whydoibother]

Prompt: medieval town guard grizzled veteran sour stern facial scar photorealistic night torch
Result:

I like the fourth, lets get a few variations:

Not great, not terrible. Needs less scars and more torch.
These are done in 10-15 seconds, by the way, only input is the text prompt above. Everything else is quickly generated.

Edit: Here's your default protagonist avatar

 

[Morgoth]

DALL-E when I type "I am an artist."

Spoiler