Using GANs to Create Fantastical Creatures
ai.googleblog.com
ai.googleblog.com
Here are the color names most disproportionately popular among men:
* Penis
* Gay
* WTF
* Dunno
* Baige
A lot of things can be done procedurally, but going from 2d concept to actual working 3D Character in video game is a long jump.
On top of that every game will have very different asset and gameplay requirements.
On the other side of it, from the triple-A point of view, automating even one of the steps you mention will be a huge time saver. Artists tend to worry about machines replacing them, but I think it will only unlock more, better art. You'll be able to do so much. So there is a big financial incentive to get this right; the first one to capture the market will be in a strong position.
But isn't that exactly why market places like the Unity Asset Store, TurboSquid, and UE Marketplace exist?
The asset stores for Unity and UE in particular offer assets that are ready for use with these frameworks and can also be easily modified if need be.
I rather think the problem is the expectation to get high quality assets for free, which just isn't a realistic assumption, unless you're willing to team up with an artist who is willing to volunteer their time.
It's also funny you mentioned Notch, as Minecraft and its aesthetic are a direct result of lack of artistic talent, proving that you don't need great looking 3d models at all to create the best selling video game in history...
I agree, though that ML/AI tooling for asset generation would be a major improvement for the gaming (and film!) industry.
Chimera Painter demo itself is unbelievably fun. When AI Dungeon and GPT-3 launched the observation was made that we are all "prompt engineers" now. It's the quality of the input that differentiates final results ;)
GAN-generated art, name and flavor text, a perhaps more structured rules engine for the card type and behavior, and AlphaZero to play the card in a deck against itself a few thousand times to tune the balance.
It's choosing a suitable net architecture, loss function, and training methodology all of which require in-depth knowledge and lots of experimentation.
Even after that you're only half way there, because now the fun and exploration is over and the pain begins: finding (and worse even - annotating!) and validating training data.
Once this herculean task has been finished, it's time to get your credit card ready and shell out hundreds of dollars/pounds/euros for days of GPU time to train the basic model.
With that out of the way, it's time for a little fun to return and finetune your model(s), which usually can be done using even a mediocre desktop PC or laptop.
There's a reason you can make serious money with even a comparatively simple and straight forward idea that uses ML and AI to get quality results. It's much, much more involved than just "learning some tensorflow".
Yes, things have progressed since my time in undergrad studying NNs and such. Presumably the work done on MtG transfers.
> Even after that you're only half way there, because now the fun and exploration is over and the pain begins: finding (and worse even - annotating!) and validating training data.
Well, for A:NR there's already multiple databases and a thousand cards. That's your training set. Throw in the online league cards if you require more.
> It's much, much more involved "learning some tensorflow".
I already work in SRE for a huge AI you've heard of, but thanks for the lecture.
A few years ago there was a community project training RNNs to generate M:tG cards [1], that if I'm not mistaken became Roborosewater [2]. M:tG already had around 12-15k cards by the time and still the majority of generated cards did not make sense. In the mtgsalvation thread in [1], most of the time people crack out with the nonsense that's generated by the RNNs people train [3]. There are also plenty of cards that make sense and are interesting and even usable, but they had to be hand-picked out of buckets of nonsense. A lot of curation and probably editing generated cards by hand would be needed, which somewhat defeats the purpose of the whole endeavour.
To generate new cards it would make much more sense to use good old procedural generation starting with a hand-crafted grammar of the rules text. This is certainly doable for a game of the size of M:tG, let alone Netrunner. For instance, M:tG Arena, the online p2p version of the game, runs on a game engine with a hand-crafted parser, the Game Rules Parser, that essentially resolves spells like an interpreter executing a script [4][5].
I mean to say, sometimes jumping feet-first into learning how to use a new set of complex tools is not necessary. Simpler tools that should already be in every programmer's toolbox can do the job fine, sometimes, even when it seems easier to just get a bunch of data and put it through a machine learning meat grinder. "Easier" might turn out to mean you need to do a lot more work before and after you can use the "easier" method (e.g. labelling, training, curation, etc) and not even get very good results in the end.
___________
[1] https://www.mtgsalvation.com/forums/magic-fundamentals/custo...
[2] https://twitter.com/RoboRosewater
[3] The project was started by the user Talcos, then more people started training their own nets using Talcos' code.
[4] https://www.reddit.com/r/magicTCG/comments/74hw1z/magic_aren...
[5] To tout my own horn a bit, I did the same thing in my degree as a final year project, but I didn't have enough time to get full coverage of the entire card corpus at the time - still, I was just one undergrad student and I did manage to get a big chunk of the game working with a rules parser. The parser was written in Prolog meaning I could "run it backwards" as a generator so the project had an M:tG ability text generator that spat out grammatically correct, if not always particularly useful, text. I'm not linking to the project because it was 9 years ago and it's painfully embarrassing looking at it now.
Discussion: https://news.ycombinator.com/item?id=25152764
https://i.paste.pics/8bed0cd17629f0e9c852a24162bf381e.png
Otherwise my clumsy, misshapen caricature turned out surprisingly nice. (I mean, relative to how nonsensical its anatomy is.) The shapes are followed very precisely, so yes, blobby input begets blobby output.
Huamns can produce novel artworks that a computer cannot, usually with more control over the intagibles like creativity and the fundamentals like drawing and perceiving the world.
The same could be said of humans. A lot of humans would say a random piece of quality GAN-generated art is a creative work. They may not say so if they first looked at every piece of training data that went into it. But then, they might not do the same if they looked at every piece of art produced and viewed by a given human artist previously.
As a thought experiment, if the line in the sand is whether a computer can generate a piece of art that people think is creative, even with full knowledge of every piece of training data- I think that's not at all difficult to overcome long term. Possibly even with current models.
When that happens people will start training GANs with GAN data.
It will become a self feeding loop with humans only curating data.
Of course even the curating of the data is ripe for automation via ML.
Example context: "Fear and Loathing in Las Vegas run through Google's Deep Dream neural network is pure nightmare fuel" https://www.independent.co.uk/arts-entertainment/films/news/...
I'm also envisioning a Ready Player One type of online escape world where content is generated from social media images/films.
In that one you couldn't get much control like drawing letter clouds in the sky
Sounds like it would become pretty dull. Art and experiences are interesting because they are unexpected. Experiencing someone else’s unique and surprising artistic voice is what I love most about good games / art experiences.
Also, on a purely economic scale, automating some aspects of art asset creation for games would be a complete game changer. It would probably up the art capabilities of indie game developers quite a bit, although it's hard to predict these types of things (in the sense that it might benefit AAA developers even more - who knows).
I doubt that. Simplified and (semi-)automated instruments have been around for a long time since there is legitimate interest in such tools from musicians.
The difference here is that people have a tendency to regard machine generated output as more "correct" (whatever that means). The same psychological mechanisms that make Level 3 autonomy in cars and aeroplanes so dangerous can lead to a future where a closed loop automated content generation (e.g. train the next gen GAN on output from the previous gen GAN) kills all creativity and conditions us to prefer the generated aesthetic.
Don't worry, it's still dystopian -- games will become exceedingly good and compelling; you won't want to do anything else.