AI Image Generation for Game Art, PETSCII Robots

[RevEng]

I've been wanting to play with AI image generation for a while now., but there were a couple things stopping me. 

 

1. I've been busy with work around the 7800 Attack of the PETSCII Robots game release.
2. I have a potato GPU in my PC.

 

While the PETSCII Robots work was wrapping up, I happened across the Hugging Face Stable Diffusion site *, which allows you to enter a prompt for Stable Diffusion and spits out the resulting pictures, all for free. The site is missing more powerful features you'd get if you ran Stable Fusion locally - I'll touch on those features a bit later - but it's a great way to get your feet wet, since it's quick and the price is right! (Note: The download version of Stable Diffusion is also free, but I hear that capable GPUs aren't free, for some stupid reason.)

 

So I figured now was as good a time as any to start talking to the AI about the game, to see what it could make.

 

 

Specifically, I wanted to see what it could make in terms of game art - not really box or cart art, but rather the sort of thing you might use to flesh out a game's backstory, and fire up imaginations - so I thought about different scenarios in PETSCII Robots. When you first start off in the game, you're usually positioned somewhere near your spaceship.

 

 

So I wanted to see what kinds of related images I could create the AI could create. After a little bit of work, I had these...

 

\

 

...I think that's a pretty good result! Mind you, getting there took a bit of prompt crafting, and there were also a whole lot of bad or weird images generated that I rejected; In addition to sometimes giving you an odd interpretation of your prompt, sometimes the AI will get confused about human body parts. I don't judge; sometimes I find human body parts confusing too. Anyway, you can always just generate more images, or if there's something really good you can save it and touch it up.

 

I thought it might be cool to see some of the events leading up to the game. Here the colonists leave home and travel to their ultimate destination...

 

...the spaceships don't match, of course. Consistency of people or objects between attempts is something you need to put more effort into, by using the more advanced features in a local version of Stable Diffusion. For this exploration, I just decided to roll with it. Maybe these are "historical artist recreations" from some distant post-robo-revolution era, and the exact details have been lost to time.

 

The PETSCII Robots game also features contrasting game locations. Sometimes you're hunting bots in the woods, and other times you're navigating futuristic colony buildings...

 

 

I thought it would be cool to first look at these settings, pre-robot revolution.

 

 

Then we can get a look at those settings in the aftermath of the revolution and resistance...

 

I've always had an interest in propaganda art, so at this point I wanted to see what the robot propaganda art might look like...

 

I think maybe the AI has a sense of humour. To me the last one looks awfully close to Rosie the Robot, after she's decided that she's dusted her last bunny...

 

While this whole exercise was just a bit of playing around, the mood caught me to add some game related text to a couple...

 

So there you have it - AI created images for a game, from someone who is a complete AI image noob. I think these tools could be very useful in our hobby, in terms of support material, and will only get more useful as they become more powerful. In the same way that there's a niche for music, pixel art, cartridge and box art, a prompt jockey could create some nice backstory imagery.

 

While I've only touched on a simplified introduction to the topic here, If you have a semi-capable GPU on your PC, you can run Stable Diffusion locally and it opens up some interesting features:

• you can input images along with your prompt, to give the AI a starting point closer to what you want.
• you have access to "negative prompting" to remove certain types of results.
• you can run other publicly available pre-trained models, such as...
  • Robo Diffusion, which has been trained specifically to produce futuristic robot images. 
  • Double-Exposure, which was trained to produce artistic composite images. A lot of the original Atari box art uses a similar composite style.
  • F222 was trained on... adult images... and even when producing non-adult content, it has a superior ability to generate realistic anatomy.
• in-painting, where you have the AI replace some area within an existing image.
• out-painting, where you have the AI extend an image beyond it's current borders.
• training your own custom model sets. You can train on custom images of a particular person, place, or thing. Using it's custom keyword in your prompts, you can get original images that incorporate that thing in a consistent and recognizable manner.

 

I'm sure there are some features I'm not yet aware of yet, and even more soon to be released. This space is moving pretty fast.

 

That's all I have. If you guys want to chat below about Stable Diffusion, the place for AI image generation in our hobby, or your love/hate for AI tech - the AI says it's cool with all of that - for now, anyway. 

 

____________________________________

* = note that I've linked to the online portal for Stable Diffusion 1, instead of the latest release (2.1 at the time of this writing). Stable Diffusion 1 seems to be better than 2 for beginners in a few different ways. The quality from v1 tends to be better than v2 if you're not using more advanced techniques, like negative prompting. Also, the v1 portal runs significantly faster than the v2 portal, which makes trying out different prompt experiments a lot less painful.

[Thomas Jentzsch]

You got some really cool results. 👍 

 

I have experimented with Stable Diffusion in the past too, but I didn't had the patience for creating the input for such great images.

[RevEng]

Thanks! For sure there's a bit of art to prompt crafting, but I did find some of the examples on this site a useful starting point.

[+Karl G]

Cool stuff for sure! I'm not going to be the one to start a discussion about the potential ethical issues with AI art. All I'll say is that the discussions I've seen contain a lot more heat than nuance. At any rate, I've enjoyed seeing what you have gotten the tool to produce here.

 

Oh, and this isn't AI-generated, but I can't resist adding this image to the topic, since it is at least thematically relevant.  

 

[CapitanClassic]

5 minutes ago, Karl G said:

I'm not going to be the one to start a discussion about the potential ethical issues with AI art

There aren’t any. “good artists borrow, great artists steal” — possibly stolen by Pablo Picasso

 

You cannot copyright an art style. Whether a human or an AI copies your art style to make new original art, the new creator of that art is the copyright holder.

 

AI art will make the creation of “good enough” art cheap for small developers of board games, video games, other publishing, it will also make artists more productive at pumping out their own new art when the AI is trained on the artists own database of work.
 

Custom made art will likely always have a marketer, but the market will be much smaller when the AI art gets closer and closer to good enough.

[+Karl G]

16 minutes ago, CapitanClassic said:

There aren’t any. “good artists borrow, great artists steal” — possibly stolen by Pablo Picasso

 

You cannot copyright an art style. Whether a human or an AI copies your art style to make new original art, the new creator of that art is the copyright holder.

 

AI art will make the creation of “good enough” art cheap for small developers of board games, video games, other publishing, it will also make artists more productive at pumping out their own new art when the AI is trained on the artists own database of work.
 

Custom made art will likely always have a marketer, but the market will be much smaller when the AI art gets closer and closer to good enough.

I do agree. I have seen a lot of claims about AI art being "stolen art", but without a lot of evidence. The nuance is that not all AI art is created equal, so there may well be some less sophisticated apps that may contain recognizable portions of artwork harvested from the internet. If an AI creates art in the style of human artists by analyzing existing art though, then I would argue that that is pretty much how human artists learn to create art as well.

[RevEng]

14 hours ago, Karl G said:

Cool stuff for sure! I'm not going to be the one to start a discussion about the potential ethical issues with AI art. All I'll say is that the discussions I've seen contain a lot more heat than nuance. At any rate, I've enjoyed seeing what you have gotten the tool to produce here.

Yeah, I'm completely aware that people are divided on this topic, and it's entirely possible that I'll be accused of being an awful person for bringing the discussion here. Still, I think it's one of those conversations that should happen, ugly or not, if some subset of people in the community are truly opposed to AI art being used for homebrew purposes.

 

I agree with both you and CaptainClassic that style can't be copyrightable. But I do think there's some nuance here too. Some researchers found a very small percentage of generated images were effectively the same as an image in the training set - the main cause seems to be when your prompt references someone or something that only has one image representing it in the training data. So you need to be careful around these sorts of edge cases.

 

For anybody keen on the legal stuff, there's a pretty interesting legal analysis of a Stable Diffusion lawsuit on the Corridor Crew channel...

I disagree with his assertion that the AI generated images are "derivative works" in the normal legal sense - to me it still seems no more derivative than an artist pulling from their own experience of art - but he makes a good case for Fair Use in the end.

 

[CapitanClassic]

That lawyer is an idiot, "the suit claims that the nearly 6 billion training images which have been scrapped from web pages and across the internet were obtained without the consent of the image owners or the website operators. The datasets is so large (tangent) ... the class action lawsuit asserts that this is a violation of the exclusive rights of the copyright holders in those images, and they're right. Stability doesnt deny that huge swaths of the training images are copyrighted, which in the most basic terms is copyright infringement on a mindbogglingly massive scale."

 

The website operators granted consent by making the images publically available.  The copyright images owners likely did as well (unless they are claiming someone else uploaded the images to the website). The AI developers aren't even distributing the images. The entire internet doesnt work if somehow they are claiming if my browser caches an image that the copyright holder put up on a public website, that I have infringed copyright. 

 

He does say that this isnt their strongest argument,  but it isnt an argument at all. Then he goes on to say, the argument comes down to if the training data is fair use or not. Which isnt the argument at all. What only matters is are the output images derivative to tne training images.

[Thomas Jentzsch]

I think part of the problem is, that when the copyright owners made their images available, the did not foresee this new use case. Probably we will see more and more limited image releases in the future which e.g. only allow a certain number of use cases.

[RevEng]

5 hours ago, Thomas Jentzsch said:

I think part of the problem is, that when the copyright owners made their images available, the did not foresee this new use case. Probably we will see more and more limited image releases in the future which e.g. only allow a certain number of use cases.

I think that's probably quite true, but if the usage doesn't involve distributing copies, then it doesn't actually matter what their consent is. To limit usage in other ways they'd need to hide the content behind a wall, and have some kind of click-through agreement. (i.e. a contract with consideration)

 

Of course, I'm saying the usage here isn't copying, and it's impossible to say how the courts will weigh in.

 

7 hours ago, CapitanClassic said:

That lawyer is an idiot [...] What only matters is are the output images derivative to tne training images.

Agreed entirely. The fair use is a second much-weaker leg, which I'd rather not stand on. But the history of copyright is rife with bad cases and corporate shenanigans, so who knows where this will end up.

 

[Sohl]

On 2/9/2023 at 2:09 PM, RevEng said:

While this whole exercise was just a bit of playing around, the mood caught me to add some game related text to a couple...

I love this poster, @RevEng!

[RevEng]

1 hour ago, Sohl said:

I love this poster, @RevEng!

Thanks! Now we know the reason why plasma guns and emp weapons are stored willy nilly.