Merry Christmas 2023 everyone!

[D-503]

Now this is a real nice one!



Hm, doesn´t seem like a standard uniform, though...

[wolf329]

That's because AI image generators don't know anything about Outsider, and thus can't make anything that actually resembles the setting, characters, or outfits.

[D-503]

AI generated? Really? Not so bad!

[QuakeIV]

Is it? It seems real to me

[Gudo]

These generative processes are really good at generating human forms, but I'm always amazed at how bad the symmetry and details are, in comparison to the rest of the work. They look great at first glance (I'm actually really impressed with the muscles in the legs) but then you start to notice things like the tops of the boots, the bracelets, the different colored sleeves... They're just really bad at making clothing and accessories symmetrical. Then there's issues like the planet on the screen having a big lump.

But the real issue is what Wolf pointed out.

AI image generators don't know anything about Outsider, and thus can't make anything that actually resembles the setting, characters, or outfits.

[Overkill Engine]

It did do a pretty good job with the legs especially, and it didn't give each hand 11 fingers, several of them floating. It did flub the left foot/boot though. The sleeves could be written off as artistic license on the part of the tailor. Otherwise nailed the "Space Opera skimpy minidress uniform" aesthetic.

It would be interesting to see what one could do if an artist were to directly train one with their own artwork samples.


Edit: Dear gawd, I just had a horrific idea - a bot trained on a diet of nothing but Liefeld. We'd *deserve* what the machines eventually do to us.

[Arioch]

Generative AI trained on 2D images doesn't understand the context of a 3D world, which is why it's especially bad at symmetry and objects which pass in front of or behind each other. And it has no persistence of memory, so if you try to create a sequence of images (for animation or sequential art) each frame exists in its own universe, with only a tangential relationship to each other frame.



There are workarounds, but it quickly reaches the point where you spend as much time fiddling with it as you would have spent just drawing it the conventional way.

All that said, it's very impressive how it handles light and shadow within the context of a scene. That's something that I would intuitively expect it to have problems with.

[wolf329]

Well yeah, it's blindly regurgitating artist's works and artists for the last few thousand years have worked out chiaroscuro pretty well. It's not surprising it copies art sort of accurately.

There are workarounds, but it quickly reaches the point where spend as much time fiddling with it as you would have spent just drawing it the conventional way.

Trying to do things the "clever, easy, faster" way always ends up being the harder, longer, worse way unless you're satisfied with inferior work and only care about pushing out the content.

Otherwise nailed the "Space Opera skimpy minidress uniform" aesthetic.

Yeah if you don't know anything how scenes are put together or anatomy or perspective, and your understanding of art ends with GPU box design, it gets (popular artistic style that has the most visibility and sourcing in the dataset) "aesthetics" down fine.

It would be interesting to see what one could do if an artist were to directly train one with their own artwork samples.

I'd rather slit my own throat than spend the rest of my life typing prompts instead of actually making art, but apparently computer touchers seem to think typing a prompt is the same thing as making art for some unfathomable reason.

Edit: Dear gawd, I just had a horrific idea - a bot trained on a diet of nothing but Liefeld. We'd *deserve* what the machines eventually do to us.

Liefeld can draw which is more than I can say for anybody who tries to pretend AI images are "art".

[Overkill Engine]

Yeah one would almost be better off with an AI built specifically to generate initial assets in 3d format then handing those assets off to a human or another specialized AI for any sort of animation.

And I imagine the AI dataset included enough art and photos with "proper" lighting that it was able to imitate a passable understanding.

[Said2k]

Generative AI trained on 2D images doesn't understand the context of a 3D world, which is why it's especially bad at symmetry and objects which pass in front of or behind each other. And it has no persistence of memory, so if you try to create a sequence of images (for animation or sequential art) each frame exists in its own universe, with only a tangential relationship to each other frame.



There are workarounds, but it quickly reaches the point where you spend as much time fiddling with it as you would have spent just drawing it the conventional way.

All that said, it's very impressive how it handles light and shadow within the context of a scene. That's something that I would intuitively expect it to have problems with.

I wouldn't say it doesn't have no knowledge because image diffusion models are being used as priors for text to 3d. It handles light and shadow decently to enough but shadow alignment, vanishing points, and perspective errors are still good ways to detect if an image is AI-Generated for now.

teaser-min (1).jpg (293.15 KiB) Viewed 239855 times

[D-503]

Whow, guys (and gals?), i just learned more about AI-pics than until now combined!

Is there some sort of "AI-pics for dummies" somewhere on the net?
Some quick links for old f*rts like me? (Entry level or easily understandable would be fine...)

[SVlad]

There is lots of free ai generator there.
Just open website, type some keywords, look at result, rinse and repeat:
Example:
https://getimg.ai/img/img-1EzNE9b80DanTlvHloHjj

[Clockwork Ninja]

Generative AI trained on 2D images doesn't understand the context of a 3D world, which is why it's especially bad at symmetry and objects which pass in front of or behind each other. And it has no persistence of memory, so if you try to create a sequence of images (for animation or sequential art) each frame exists in its own universe, with only a tangential relationship to each other frame.

Are there any AI models out there that train on and generate 3d models? Even if the user only took 2d images of the result.

[Demarquis]

No tutorials, then?

[Said2k]

Generative AI trained on 2D images doesn't understand the context of a 3D world, which is why it's especially bad at symmetry and objects which pass in front of or behind each other. And it has no persistence of memory, so if you try to create a sequence of images (for animation or sequential art) each frame exists in its own universe, with only a tangential relationship to each other frame.

Are there any AI models out there that train on and generate 3d models? Even if the user only took 2d images of the result.

I only know of Stable Zero123 from StabilityAI, but it's also trained on 2D images and 3D models.

[Arioch]

Generative AI trained on 2D images doesn't understand the context of a 3D world, which is why it's especially bad at symmetry and objects which pass in front of or behind each other. And it has no persistence of memory, so if you try to create a sequence of images (for animation or sequential art) each frame exists in its own universe, with only a tangential relationship to each other frame.

Are there any AI models out there that train on and generate 3d models? Even if the user only took 2d images of the result.

There are some being worked on, though I think that's still in a rough state. I suspect it won't be long.

[Overkill Engine]

Financial incentives alone are too large for there not to be. Though it will be interesting to see what happens once modern creative rights laws actually get enforced properly.

That's my main issue with these tools, is not them existing, but them utilizing sources for training data without the explicit consent of their authors.

[wolf329]

...That's my main issue with these tools, is not them existing, but them utilizing sources for training data without the explicit consent of their authors.

Since they literally can't and never will function without doing that, it actually comes down to, "Do you care about the artists/actors/voice performers/musicians who made everything that these tools use without any consent whatsoever, or not?" Because that's the real question.

And anyone who uses them very clearly doesn't.

[QuakeIV]

To be honest, I'm fairly sure it was always okay to look at artwork, visually assimilate that information, and for that information to potentially influence your own later work.

It seems like a mis-application of the law to try to apply copyright here. This artwork was in general publicly distributed onto the internet and has been obtained by those means by various companies producing AI models. You can retain a lawfully obtained copy of something for your own use as long as you don't redistribute it. If the AI can be proven to somehow be cutting and pasting bits and pieces of the source art together and redistributing that, that would be one thing, but it really doesn't look like it works that way.

[Overkill Engine]

Thing is, a human being given permission to look at and thus potentially learn from art could be argued to be legally distinct from giving an automated tool permission to do so.

Basically usage permissions need to be updated and clarified as technology advances since it is not reasonable to expect an artist to have perfect foreknowledge of such.