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

[Norfleet]

If you can't already do art, you shouldn't be doing art. That is to say, I can understand the argument that AI is potentially valuable if used by artists, but art that's produced by relying on the AI's art-making capability alone is not going to be art. You have to have artistic sensibility there already, in order to sift the wheat from the chaff in what the AI produces.

The ability to perceive quality in art is not the same as the ability to actually create art, though. It's just like how I can taste bread, but I have no real clue how exactly that bread came to be. Something involving wheat and ground-up bones. If I find an instant bread machine where I can throw in these things and get bread out, I can thus make and eat bread without any real understanding of how to actually make a bread. Because ultimately, I have no aspirations of being a breadsmith. It's just that people demand bread to go with my product.

 

[Raghar]

Social issues require social remedies, not halting our technological progress in the name of "speaking to each other more" or whatever. The key reason behind atomization of society is the ongoing urbanization of the world's population, not the fact people have smartphones. People never had tightly-knit communities in cities (unless you count street gangs, lol), and you are idealizing the shit out of "contributing to something greater". No, a vegetable seller really didn't feel like he was contributing to "something greater" regardless of the period.

Czech dissidents and bookworms were quite tight societies during socialistic era. Then capitalism came, earning money become glorified, they could earn money without giving a shit about others, and they were free to have tourism around the world.

Book market collapsed, and everyone started to shit on everyone just to get better salary, frankly western like society isn't helping to create some groups. People who are using cellphones 24/7 might claim they are connected as a part of a group, but that's not connection. That's like when you went to doctor, and your neighbor threw a brick into your window to give you something to improve relationships. Or when some government officials had scribbles "kurwa" on theirs front doors. There were plenty of contacts in old socialistic urban areas.

Actually when I think about it, the increased differences between highest income and irrelevant strata of society might be another reasons why modern European societies sucks balls. (As Monkeypox proven, German's LGBT groups do that literally.)

 

[Norfleet]

The main reason why AI generated art is a problem is nobody made a law that puts AI generated images into free use.

Technically, the opposite is true: There isn't currently a law that enables you to copyright AI art, as it is not the product of a person, so you can't necessarily stop anyone from ripping off the art because you didn't actually make it, it's just the vomitings of a computer algorithm.

 

[Raghar]

If a person uses AI image generation to make similar art, who cares the person spend an effort. And when games look too similar, the second arrival could be pressured to alter the game to look different to avoid confusion between his game and the original. The game as a whole is copyrighted, and there are some protections prevents degenerate behavior.

For example, Xenonauts evaded problems by having theirs own art, but if Xenonauts and XCOM would use THE SAME AI generated images. Then both games would be too similar, and XCOM is PRIOR ART.

 

[Skorpion]

 

[Poseidon00]

I have become mildly obsessed with AI since first being shocked with how great the new models were. Some of the people in the field who I follow are talking about building good text-to-music and text-to-video generators using the current model, to say nothing about more powerful ones.

The next 5 years are going to be absolutely wild bros

 

[gurugeorge]

People never had tightly-knit communities in cities.

This is not true. Tightly-knit communities in cities faded out with the car, with urban sprawl, etc., before that, cities did have tightly-knit communities of various kinds, which grew around walking distance and walkability.

I think we've said our respective pieces on this, so I'll stop here. It is possible that I might be misunderestimating how human ingenuity will deal with all this. But I doubt it - as I said, human life is about more than individual action per se, there has to be a meta level of agreement-in-action, otherwise the individual rationality of the Prisoner's Dilemma runs wild, and everyone gets the worst possible outcome.

 

[gurugeorge]

If you can't already do art, you shouldn't be doing art. That is to say, I can understand the argument that AI is potentially valuable if used by artists, but art that's produced by relying on the AI's art-making capability alone is not going to be art. You have to have artistic sensibility there already, in order to sift the wheat from the chaff in what the AI produces.

The ability to perceive quality in art is not the same as the ability to actually create art, though. It's just like how I can taste bread, but I have no real clue how exactly that bread came to be. Something involving wheat and ground-up bones. If I find an instant bread machine where I can throw in these things and get bread out, I can thus make and eat bread without any real understanding of how to actually make a bread. Because ultimately, I have no aspirations of being a breadsmith. It's just that people demand bread to go with my product.

Yeah, but wouldn't you be better off finding a living, breathing breadsmith to collaborate with, who can select better bread from the plethora of options the bread machine produces for you?

 

[Raghar]

I have become mildly obsessed with AI since first being shocked with how great the new models were. Some of the people in the field who I follow are talking about building good text-to-music and text-to-video generators using the current model, to say nothing about more powerful ones.

The next 5 years are going to be absolutely wild bros

Artists are screeching already:
"The AI used images of artists to learn, thus they are copying our work."

Do you remember about these SF books about societies before collapse? AI, homosexuality a society norm, rich having too much money... And the rest of decalogy books was about live after collapse.

Current leadership is doing its best to start WW3. Or perhaps we would see just changes of laws to turn the society into old cliché called dystopia. As they say... We had AI and we wuz Kungs.

 

[Skorpion]

People never had tightly-knit communities in cities.

This is not true. Tightly-knit communities in cities faded out with the car, with urban sprawl, etc., before that, cities did have tightly-knit communities of various kinds, which grew around walking distance and walkability.

I think we've said our respective pieces on this, so I'll stop here. It is possible that I might be misunderestimating how human ingenuity will deal with all this. But I doubt it - as I said, human life is about more than individual action per se, there has to be a meta level of agreement-in-action, otherwise the individual rationality of the Prisoner's Dilemma runs wild, and everyone gets the worst possible outcome.

continues to argue
bro you use individual comments like other people use commas,

 

[Nutmeg]

How many years before AI, under the control of an untrained person (i.e. not someone with digital art training of any kind), can produce 240p pixel art animated sprite work at the quality level of Metal Slug?

Estimates please.

My own estimate is maybe another 20-25 years.

In terms of AI being employed by digital artists as part of their process and this being an undeniable productivity boost (as opposed to some idiosyncrasy of the artist) to produce *quality* game art work, I would estimate another 10 years and considerable developer effort to build the tooling.

 

[The_Sloth_Sleeps]

How many years before AI, under the control of an untrained person (i.e. not someone with digital art training of any kind), can produce 240p pixel art animated sprite work at the quality level of Metal Slug?

Estimates please.

My own estimate is maybe another 20-25 years.

In terms of AI being employed by digital artists as part of their process and this being an undeniable productivity boost (as opposed to some idiosyncrasy of the artist) to produce *quality* game art work, I would estimate another 10 years and considerable developer effort to build the tooling.

A bit difficult to estimate as we have never had an "AI" tool for any application that could produce consistent quality results. Using the current approaches, probably never is my guess.

These tools seem to rely heavily on large amounts of data and regression, undoubtedly there are some CG techniques on top of that(feature extraction, edge detection etc). It does not appear to have any grasp of form, pose or motion, etc. Although I suspect there is some hard coding around humanoid generation (it knows a frontal face has two eyes a mouth etc).

This does not scale up too well, as we can see. Ask it to draw a cube, rotate it 25 degrees left and 15 up, and redraw. Trivial for a human with understanding of form. But this is a very weak approach "AI" cannot even manage the basics. Instead it stupidly falls back to make a composition based off a regression of white cubes images.

"White 3d cube on a black background, lighting top left corner. Cube rotated 15 degrees on the up axis. Top side is red colour."

Its worth noting that its been trained(?) or algorithmically modelled such that it can make a guess at where "top" is. I believe this to be a few tricks employed to make it seem smarter than it is. They have obviously had a stab at trying to do "holding" as well.

But using the same prompt but extending...

''White 3d cube on a black background, lighting top left corner. Cube rotated 15 degrees on the up axis. Top side is red colour. There is a monkey in a tracksuit holding a watermelon sitting on top."

Can form really be taught from images and tags alone? I would imagine you would need skin to sense through touch, as well as a muscle system to be engaged for handling weight. Underpinning that is a sense of space, and objects in space. I suspect it may not be possible to encode this in a regression model using only images and tags.

 

[Nutmeg]

Can form really be taught from images and tags alone?

No but it might be possible from 3D scene data paired with descriptions and renders. There was actually a gofai system that performed well in the kind of tasks you gave examples of.

I would imagine you would need skin to sense through touch, as well as a muscle system to be engaged for handling weight.

I agree that ultimately for AI to be human like it will need to be embodied. But human like is not necessarily ideal for all tasks, or even for use by humans. IMO it's still low hanging fruit days for audio visual sensor (no motor, no chemical (smell, taste) or pressure (touch) sensor) input based AIs.

 

[mkultra]

How many years before AI, under the control of an untrained person (i.e. not someone with digital art training of any kind), can produce 240p pixel art animated sprite work at the quality level of Metal Slug?

Estimates please.

My own estimate is maybe another 20-25 years.

In terms of AI being employed by digital artists as part of their process and this being an undeniable productivity boost (as opposed to some idiosyncrasy of the artist) to produce *quality* game art work, I would estimate another 10 years and considerable developer effort to build the tooling.

Maybe a year, it's going tremendously fast now with many new additions, extensions every week.
What we have now is going to be absolutely laughable in just a year, just like Disco Diffusion was incredible 1 year ago, today we don't even use it, and it seems useless if we look at it today. Actually i think the development has speeded up, so i think what we have currently will be seen as even worse.

For textures (3D textures) i could easily use it today, and it could compete with AAA-games where they've worked for years with those textures. I have models for flat plane 3D grass/plants and models for making 3D textures. I have around 6000 textures on my HDD right now, improving every month (because my model improves, and my knowledge of how to use ML, what mix of textures works best etc etc)

 

[gurugeorge]

People never had tightly-knit communities in cities.

This is not true. Tightly-knit communities in cities faded out with the car, with urban sprawl, etc., before that, cities did have tightly-knit communities of various kinds, which grew around walking distance and walkability.

I think we've said our respective pieces on this, so I'll stop here. It is possible that I might be misunderestimating how human ingenuity will deal with all this. But I doubt it - as I said, human life is about more than individual action per se, there has to be a meta level of agreement-in-action, otherwise the individual rationality of the Prisoner's Dilemma runs wild, and everyone gets the worst possible outcome.

continues to argue
bro you use individual comments like other people use commas,

Who rattled your cage? And have you never heard of a Parthian shot?

 

[v1c70r14]

How many years before AI, under the control of an untrained person (i.e. not someone with digital art training of any kind), can produce 240p pixel art animated sprite work at the quality level of Metal Slug?

Estimates please.

How long would it take you personally until you can do it? Poland is supposedly a game development rising star of a nation, how long until someone there can produce pixel artwork at the quality level of Metal Slug? How long until continental Eurasia and the Americas reach the level of pixel art quality the Japanese had during the 90's and in the early 00's? Estimates until any African nation achieves it, hardware with the same processing and graphical power as in the 90's should be affordable enough, surely, plus labor costs there are cheap, you could probably personally hire a crack team Congolese pixel artists, right? Unlike the far-off AI models they would get right at it and we'll see a Metal Slug spiritual successor any day now.

I'm not defending the capabilities, or future capabilities of AI, so much as I'm putting your highly specific demands into context. If AI gets to Metal Slug levels of pixel art, something there isn't much demand for by the entities developing it with proper funding, within a decade or a couple then that's leaps and bounds over what humans can achieve and a miracle.

Maybe it will never get there, although never is a long time and so that is unlikely unless its development cease for some reason, but it'd be hard since they are thriving on big datasets and there are very few game assets it could be trained on to produce something of that quality. Will humans ever get there again though? Where is this contemporary Nazca/SNK that you are comparing AI output to and that you want it to do for free and for whatever ridiculous reason want to be doable by a single unskilled and untalented guy (presumably yourself)?

 

[fizzelopeguss]

But is doing art boring work? Is writing boring work? Office type stuff, sure, that's boring enough and we could happily see that being done by machines. But when it's reaching into things that are more personal and human like art and writing, then I think we're getting into dangerous territory.

The better question is, can the machine do something I CAN'T do? If so, then it is immediately useful. I have zero skills of an artist, so any machine that can generate art, even art of inferior quality, gives me the capacity to produce an art.

Would that be desirable or good though? If you can't write worth a shit you could get an AI to draft a story for you. Even if its dogshit, it gives you the capacity to write. But if you rely on machines to do the heavy lifting, you're not acting as a designer, at best you're a commissioner deciding if the outsourced work you're given is up to snuff and worthy of putting on the market.

The entirety of the younger population will be utterly reliant on these programs to write their CV, dissertations, business correspondence, wedding speeches. Anything that even required the slightest bit of thought. It's only a matter of time before some tranny tech creature starts putting a blacklist on certain words, phrases or language patterns. If you think the Americanisation of the universe isn't bad enough already... We'll memory hole and double good plus ourselves within a decade.

 

[The_Sloth_Sleeps]

But is doing art boring work? Is writing boring work? Office type stuff, sure, that's boring enough and we could happily see that being done by machines. But when it's reaching into things that are more personal and human like art and writing, then I think we're getting into dangerous territory.

The better question is, can the machine do something I CAN'T do? If so, then it is immediately useful. I have zero skills of an artist, so any machine that can generate art, even art of inferior quality, gives me the capacity to produce an art.

Would that be desirable or good though? If you can't write worth a shit you could get an AI to draft a story for you. Even if its dogshit, it gives you the capacity to write. But if you rely on machines to do the heavy lifting, you're not acting as a designer, at best you're a commissioner deciding if the outsourced work you're given is up to snuff and worthy of putting on the market.

The entirety of the younger population will be utterly reliant on these programs to write their CV, dissertations, business correspondence, wedding speeches. Anything that even required the slightest bit of thought. It's only a matter of time before some tranny tech creature starts putting a blacklist on certain words, phrases or language patterns. If you think the Americanisation of the universe isn't bad enough already... We'll memory hole and double good plus ourselves within a decade.

That is definitely a concern. However its not really good enough that you can switch off your brain and create content that most people will want to absorb. Thankfully, as low as people standards are its just not that good to be able to do that.

I think how it will play out is that A) the majority of people do not read much anyway so will not be affected. This is the type you really have to fight for their limited attention B) the discerning reader types will of course hate these programs, which leaves C) a smaller yet loud group of people, probably redditors, who will be the ones reading this stuff.

 

[mkultra]

Some of my models, here's a Sci-Fi themed one for doors, walls, hulls, crates etc. I currently have 6000 textures from this particular model, so this is just mostly random images of those.. will obviously take me hours or likely days to go through them all, find the best ones, but it's pretty consistent and doesn't really output anything that couldn't be used in some way.

Spoiler

This is just a tweaked model of above but more geared towards sci-fi horror, these were some 2048x2048 generations i did

This is my flora model:

Some of the flora models in-game:

These are constantly improving though because i still consider myself a newbie of ML.. There are just so many variables and hard to know what gives the best results. What i've improved the most is learning what kinds of images to mix to get a good and varied result. The training images are all my images, though the latest models also contains images it has generated from the previous model (so its like a feedback-loop almost).

 

[V17]

How many years before AI, under the control of an untrained person (i.e. not someone with digital art training of any kind), can produce 240p pixel art animated sprite work at the quality level of Metal Slug?

Estimates please.

My own estimate is maybe another 20-25 years.

In terms of AI being employed by digital artists as part of their process and this being an undeniable productivity boost (as opposed to some idiosyncrasy of the artist) to produce *quality* game art work, I would estimate another 10 years and considerable developer effort to build the tooling.

All the way to something so specific and imo iconic as Metal Slug is really hard to say. The last bit of quality is the most difficult even for a talented and experienced human.
If we make it easier to "approximately 90% of the way to Metal Slug quality", or "good quality Metal Slug tribute without any obvious AI mistakes" I'd say 2 years at most.

How many years before AI, under the control of an untrained person (i.e. not someone with digital art training of any kind), can produce 240p pixel art animated sprite work at the quality level of Metal Slug?

Estimates please.

My own estimate is maybe another 20-25 years.

In terms of AI being employed by digital artists as part of their process and this being an undeniable productivity boost (as opposed to some idiosyncrasy of the artist) to produce *quality* game art work, I would estimate another 10 years and considerable developer effort to build the tooling.

A bit difficult to estimate as we have never had an "AI" tool for any application that could produce consistent quality results. Using the current approaches, probably never is my guess.

These tools seem to rely heavily on large amounts of data and regression, undoubtedly there are some CG techniques on top of that(feature extraction, edge detection etc). It does not appear to have any grasp of form, pose or motion, etc. Although I suspect there is some hard coding around humanoid generation (it knows a frontal face has two eyes a mouth etc).

This does not scale up too well, as we can see. Ask it to draw a cube, rotate it 25 degrees left and 15 up, and redraw. Trivial for a human with understanding of form. But this is a very weak approach "AI" cannot even manage the basics. Instead it stupidly falls back to make a composition based off a regression of white cubes images.

"White 3d cube on a black background, lighting top left corner. Cube rotated 15 degrees on the up axis. Top side is red colour."

Its worth noting that its been trained(?) or algorithmically modelled such that it can make a guess at where "top" is. I believe this to be a few tricks employed to make it seem smarter than it is. They have obviously had a stab at trying to do "holding" as well.

But using the same prompt but extending...

''White 3d cube on a black background, lighting top left corner. Cube rotated 15 degrees on the up axis. Top side is red colour. There is a monkey in a tracksuit holding a watermelon sitting on top."

Can form really be taught from images and tags alone? I would imagine you would need skin to sense through touch, as well as a muscle system to be engaged for handling weight. Underpinning that is a sense of space, and objects in space. I suspect it may not be possible to encode this in a regression model using only images and tags.

The issue here is you don't discern between different architectures of image generators, each of which creates and "understands" images in a different way. Current diffusion models are notoriously bad at understanding descriptions like yours, where you need specific composition of two or more objects with different qualities. There are even some standard tests prompts that go something like "a photo of a small red cube on top of a large blue cube".

There are other image generation models which do not suffer from this and also can for example do text pretty well, something which Stable Diffusion hilariously struggles with. However afaik none is available outside of academic use yet and I believe they need beefier hardware, so they won't be as prevalent as Stable Diffusion. So we don't know what their weaknesses are or how fast they're going to improve.

What you're doing here is showing a task that current models have been equally bad at from the beginning because it's currently one of their inherent weaknesses, which is valid, but it doesn't show that the technological improvement in other tasks has been immensely fast and that this tempo of development may realistically be repeated when something like DeepFloyd drops (or it may not).

 

[thesecret1]

DeepFloyd

Fentanyl-powered AI model?

 

[Nutmeg]

All the way to something so specific and imo iconic as Metal Slug is really hard to say

Fine, lets say the Fire Emblem 7 attack animations.

See my point? The highly skilled human touch will be necessary for another decade at least yet.

 

[mkultra]

I think synthesized art rather soon (years) will be comparable to synthesized music (which is the norm for most games). Some of the bigger studios instead hires an entire live orchestra and everything else required around that, this to really have great quality, to be taken seriously but more importantly its something you can actually promote - "we use a real orchestra" vs "we used a computer to make our soundtrack" (no one can use that..). For synthesized art it will likely become similar. "We use real artists" vs "we generate our art in a program" (again, this can't be promoted and it will never be seen as the same quality/effort, no matter how great it looks.)
Though it doesn't mean there won't be tons of great games using it (and tons of shit too, like always).

But in the long run we also might get reputable ai artists, then its something which will be taken a bit more seriously if attached to a game. I do think that's pretty far off into the future though.

 

[V17]

All the way to something so specific and imo iconic as Metal Slug is really hard to say

Fine, lets say the Fire Emblem 7 attack animations.

See my point? The highly skilled human touch will be necessary for another decade at least yet.

I'm pretty sure that is going to be real within 2 years with two conditions: firstly this style (or whatever other style we pick for this) has to be popular at the time and secondly it's still going to require a person who's possibly unable to make any art themselves but able to recognize what's good.

I am about 95% sure that the technology is going to be there, but for it to work with a specific style the style will either have to be popular enough for a corporation to invest money in recreating it, or just liked enough in the community for someone to train a free tool to do it. So what may happen is that we will have the technology but it will only be used to generate asset store-like crap spritesheets instead of actually good stuff.

Feel free to mark this conversation in your calendar 2 years in advance, might be fun to argue about who was more accurate.

 

[Norfleet]

The thing is that vidya game art doesn't have to be GOOD, it just has to be GOOD ENOUGH. Once you hit that threshold, bored and now unemployed artistes will line up to make it better for you for free in the new HD MEGAMOD!!!111.

 

[Üstad]

It would get better if only I bothered to learn to use comfyui or controlnet.

nsfw

One of them is not Stable Diffussion though.