The leading 3d gen techniques are latent diffusion models that work off voxelized intermediate representations of objects. When you convert these into meshes, very often they will have poor topology. Topology is important if you need to modify these meshes or animate them. Bad topology and excess verts is also bad if you're building assets for a game.
There's still more refinement of existing techniques that need to be done before AI generated 3d models match what humans produce.
My experience here has been so bad with the topology that I usually use the created mesh as an example and rebuild the whole thing from scratch to reduce the verts. Not even worth the time.
A good auto retopologizer (AI or not) would be one of the greatest possible innovations in 3D. Add to that better unwrapping and rigging, and you’ve eliminated most of the mechanical grunt work in 3D, leaving the truly important sculpting and animating to the artists.
I'm sorry to say this, but the voxelization diffusion model I mentioned is effectively AI sculpting, and AI animation is already quite good (for humanoids at least)
Virtually all the dominos have fallen except good topology at the moment, and unfortunately that's the biggest domino there is.
AI videogen and image gen will hollow out a lot of 3d work that isn't explicitly gaming oriented IMO, and eventually even that work may go away once better World-Action-Models come online.
The most interesting/surprising development for me has been that LLMs are getting better and better at synthesizing procedural content.
Fable 5 can write procedural texture shaders, CSG, surface meshing, l-systems, even rigging + realtime IK, etc. and do it all in a visual loop within an environment. The model understands the topology, and can tune things with mathematical precision where necessary. Which is much closer to using 3D tools like an artist would.
> CGTrader, an online marketplace for 3D assets, found that one in six models uploaded to the site these days is AI generated, but that AI generated assets account for only $1 out of $90 in revenue on the site. These numbers show that AI generated assets are quickly flooding the marketplace, but that most people are not interested in paying for them.
> ...
> AI generated content is creating a discoverability problem across the internet. On Instagram, porn sites, YouTube, and music streamers, human creators are being drowned out by a flood of AI generated slop.
The obvious solution here is to to increase the upload rates to speed AI adoption. Maybe direct integration with the generation tools? If we can flip the ration from 1:6 AI generated to 1:10 or even 1:100 human generated, we can probably drive up the AI generated revenue proportion.
Why is that a goal? More likely the site that can surface the stuff people actually want to pay for (which seems to be human generated) will get more revenue in total.
We need to increase slop tolerance to speed AI adoption, and the best path towards that is to increase the availability of slop while simultaneously gaslighting people about how good it is. Eventually people will accept slop and buy it or ideally generate it for themselves (which uses more tokens in aggregate).
If we can keep that up for only a little while, we can frustrate artists economically, causing people to de-skill, eventually leaving slop as the only (and therefore best) option. Then people's revealed choices will prove that slop is really all they want.
At first I honestly didn't know if you were being ironic or not. Now I'm still unsure, but I'm pending towards believing you are.
Now, even if you are just joking around, this makes me wonder about how fucked up we are, if we have to admit there are real people out there who might actually really think like that.
There is an army of slop-merchants, ai-evangelists, scam artists, and people who hate having to pay for artistic talent because they have none that think exactly this way.
is to force more low quality AI generated slop down people's throats? sorry, have i misunderstood what you're proposing? (the ratio part of your comment was super confusing to read).
> (the ratio part of your comment was super confusing to read)
Basically, we need to get the the point where people rarely encounter anything but slop. Right now on that site 1 out of every 6 models is slop, it would be better if 5 out of 6 were slop. Getting to 99 out of 100 would be even better.
Having purchased a number of 3d models both professionally and personally - it’s just like any other asset you might need. They can be useful in games, as props in larger modeled scenes, and so on.
This article is mostly talking about assets used outside of printing or manufacturing, so it’s probably going to be closer to formats like OBJ or FBX.
You can almost always tell when a model was created with 3D AI tools. The topology is usually an absolute mess, and when you combine that with bad UV maps for textures, it practically becomes nightmare fuel.
"Absolute mess" may even be giving them too much credit, since most seem to have shot past "steaming pile" and are offroading in hitherto unexplored regions of sucky-space.
yes. people who are making content professionally need to make sure they actually have a license for the content they use. model, sound effect, whatever.
the best way to do that is either create it yourself or buy it from place like cg trader that gives you a license.
does everyone do that? no im
sure not. but for any reasonably sized company the risk vs cost
is just not worth it.
I bought an STL for a very specific use-case because of how niche it was and how few purchases the creator had... and because I couldn't find anything similar enough elsewhere, but still I did buy one.
Usually FBX, USD, or GLTF files, if you’re in the gaming industry. AI models rarely do well there because deliberate topology matters for optimization and animation.
I have, but only a few times. For one-offs, it's been when I've come across a paid model that included a particular technique or mechanism involved that I really liked and hadn't seen before. In those cases, it's worth the cost just to get a working example of it I can pick apart, rather than flail around trying to replicate it from scratch.
The other time was paying for a subscription to this[1] Thangs profile, which was more or less the same. I love making stuff using these techniques, and having access to all of his paid models really helped me get familiar with it using working examples. I mainly make my own designs now instead, but still pay for the subscription out of pure appreciation.
(reads:) "AI Generated Content Floods Market, But Almost No One Is Buying"
So if that is a Thesis, in Terms of: heck the 'AI' has perfectioned it's own 'due to circumstances'-Copyright ??
I may add (reading further:) "AI" content in terms of "voxelized intermediate representations of objects" for example...
...converted to meshes, they will have poor topology. But topology is important if you want to modify the meshes or animate them -so
"bad 'ass'-hats (assets) AI technology exchange models may sound like a solveable prob, and lesser Ideological than hu politics which was added to those LLMs before, not?
Maybe pick a size first (limited) 320x480 if thats the old PAL, dunno...now let 'AI' slove the mystery you all seem to have...
And for motivation (btw), if there are some pixel-games falling off from, may be they are liked, too...
But as always, just my 2 Cents... (somebody listed US$20,- that may count a lot more...)
Isn't this basic economics? Supply and demand. If slop floods the zone then human work becomes even more scarce and valuable.
What is my incentive to pay you? A person who spent no time working on something and who probably won't spend any time working with me if I need support. You're literally a meat puppet and a useless middleman when I could pay $20 and do it myself.
Unfortunately we've experienced the basic economics of spam. It costs virtually nothing to spam emails and so the spammers exploited that. Now it costs virtually nothing to spam markets with AI slop. Platforms need anti-slop filters and the easiest one is to charge x dollars per model to publish there.
yep basic longtail effect. back in the day at a PRO (performing rights organization -- music copyright stuff) we would analyse the longtail to death because the ~10% artists would get ~90% of play time (=> apportioned funds).
i guess in this case and probably some other cases that AI slop is just bulking up the longtail even more than before. 90% of people don't listen to AI generated music, it's only weirdos on HN that i've seen claiming to want access to it for personal listening.
> Our goal isn't to increase the volume of AI-made assets on the marketplace. It's to give designers better tools to work faster and focus more of their time on creative work.
>
> Source: https://www.404media.co/ai-generated-3d-models-flood-market-but-almost-no-one-is-buying-them
This is not art. I do not and never will believe "AI" supports the creativity mentioned. More to that, I believe only not artists would say that nonsense. Or, sorry, a dear idiot even, who could not care less about art, human, the future at all.
Do you even realize how ineffably much an artist may imagine, consider, realize, choose while actually creating? It may be weeks, months, years... when artists developers, creates, enchants their unique miraculous creation... making it invaluable... an actual art...
These human experienced choices are immeasurably important where no "AI" should ever exist...
The effortful lovely choices, where every single choice polishes an artwork... reifying it into a perfect or an imperfectly perfect miracle...
AI Creators and NFT Bros are cut from the same cloth. Just polluters who think because they can type on a keyboard it makes them special despite having spent most of their lives as talent-less hacks.
I won’t be surprised if 3D modeling and rendering software will soon add some sort of mechanism or watermark or a hashfile to prove that it was modeled with this software rather than some slop.
Yeah obviously, it’s better tho to filter the noise in advance, but if AI slop flooded a market place, real designers might stop uploading and sharing their work.
The buyers are saying it's because the human-created models are better. This might be true, but it could also be some other things. Maybe they're doing the AI dream and using AI to create the models themselves. The article cites revenue maybe AI model are selling, just cheaper. Maybe human creators are buying from humans out of solidarity.
The leading 3d gen techniques are latent diffusion models that work off voxelized intermediate representations of objects. When you convert these into meshes, very often they will have poor topology. Topology is important if you need to modify these meshes or animate them. Bad topology and excess verts is also bad if you're building assets for a game.
There's still more refinement of existing techniques that need to be done before AI generated 3d models match what humans produce.
My experience here has been so bad with the topology that I usually use the created mesh as an example and rebuild the whole thing from scratch to reduce the verts. Not even worth the time.
The real algorithm that's needed is an intelligent AI mesh retopologizer.
Honestly more of a reinforcement learning AI problem than a generative AI problem IMO
A good auto retopologizer (AI or not) would be one of the greatest possible innovations in 3D. Add to that better unwrapping and rigging, and you’ve eliminated most of the mechanical grunt work in 3D, leaving the truly important sculpting and animating to the artists.
I'm sorry to say this, but the voxelization diffusion model I mentioned is effectively AI sculpting, and AI animation is already quite good (for humanoids at least)
Virtually all the dominos have fallen except good topology at the moment, and unfortunately that's the biggest domino there is.
AI videogen and image gen will hollow out a lot of 3d work that isn't explicitly gaming oriented IMO, and eventually even that work may go away once better World-Action-Models come online.
The most interesting/surprising development for me has been that LLMs are getting better and better at synthesizing procedural content.
Fable 5 can write procedural texture shaders, CSG, surface meshing, l-systems, even rigging + realtime IK, etc. and do it all in a visual loop within an environment. The model understands the topology, and can tune things with mathematical precision where necessary. Which is much closer to using 3D tools like an artist would.
> CGTrader, an online marketplace for 3D assets, found that one in six models uploaded to the site these days is AI generated, but that AI generated assets account for only $1 out of $90 in revenue on the site. These numbers show that AI generated assets are quickly flooding the marketplace, but that most people are not interested in paying for them.
> ...
> AI generated content is creating a discoverability problem across the internet. On Instagram, porn sites, YouTube, and music streamers, human creators are being drowned out by a flood of AI generated slop.
The obvious solution here is to to increase the upload rates to speed AI adoption. Maybe direct integration with the generation tools? If we can flip the ration from 1:6 AI generated to 1:10 or even 1:100 human generated, we can probably drive up the AI generated revenue proportion.
Why is that a goal? More likely the site that can surface the stuff people actually want to pay for (which seems to be human generated) will get more revenue in total.
We need to increase slop tolerance to speed AI adoption, and the best path towards that is to increase the availability of slop while simultaneously gaslighting people about how good it is. Eventually people will accept slop and buy it or ideally generate it for themselves (which uses more tokens in aggregate).
If we can keep that up for only a little while, we can frustrate artists economically, causing people to de-skill, eventually leaving slop as the only (and therefore best) option. Then people's revealed choices will prove that slop is really all they want.
I'd disagree if it weren't so likely!
So you are betting on things getting better by… letting things get worse.
At first I honestly didn't know if you were being ironic or not. Now I'm still unsure, but I'm pending towards believing you are.
Now, even if you are just joking around, this makes me wonder about how fucked up we are, if we have to admit there are real people out there who might actually really think like that.
Good job on that one.
There is an army of slop-merchants, ai-evangelists, scam artists, and people who hate having to pay for artistic talent because they have none that think exactly this way.
> The obvious solution here ...
is to force more low quality AI generated slop down people's throats? sorry, have i misunderstood what you're proposing? (the ratio part of your comment was super confusing to read).
> (the ratio part of your comment was super confusing to read)
Basically, we need to get the the point where people rarely encounter anything but slop. Right now on that site 1 out of every 6 models is slop, it would be better if 5 out of 6 were slop. Getting to 99 out of 100 would be even better.
3D model generation is for end-users to generate bespoke without waiting for an artist. It is not for artists to slop onto a store.
You don't go to a restaurant to order 1 minute hot pockets. That's why 3D artists are not being replaced by AI.
I feel like this article doesn't understand how people use AI in the industry, or how products integrate it.
Has anyone really ever bought a STL (I assume) model? It's like WinRAR, does anyone ever buy a license?
Having purchased a number of 3d models both professionally and personally - it’s just like any other asset you might need. They can be useful in games, as props in larger modeled scenes, and so on.
This article is mostly talking about assets used outside of printing or manufacturing, so it’s probably going to be closer to formats like OBJ or FBX.
You can almost always tell when a model was created with 3D AI tools. The topology is usually an absolute mess, and when you combine that with bad UV maps for textures, it practically becomes nightmare fuel.
https://www.cgtrader.com/3d-models/electronics/other/vintage...
> The topology is usually an absolute mess
This.
"Absolute mess" may even be giving them too much credit, since most seem to have shot past "steaming pile" and are offroading in hitherto unexplored regions of sucky-space.
yes. people who are making content professionally need to make sure they actually have a license for the content they use. model, sound effect, whatever.
the best way to do that is either create it yourself or buy it from place like cg trader that gives you a license.
does everyone do that? no im sure not. but for any reasonably sized company the risk vs cost is just not worth it.
I bought an STL for a very specific use-case because of how niche it was and how few purchases the creator had... and because I couldn't find anything similar enough elsewhere, but still I did buy one.
Usually FBX, USD, or GLTF files, if you’re in the gaming industry. AI models rarely do well there because deliberate topology matters for optimization and animation.
I have, but only a few times. For one-offs, it's been when I've come across a paid model that included a particular technique or mechanism involved that I really liked and hadn't seen before. In those cases, it's worth the cost just to get a working example of it I can pick apart, rather than flail around trying to replicate it from scratch.
The other time was paying for a subscription to this[1] Thangs profile, which was more or less the same. I love making stuff using these techniques, and having access to all of his paid models really helped me get familiar with it using working examples. I mainly make my own designs now instead, but still pay for the subscription out of pure appreciation.
[1] https://thangs.com/designer/kaizen3dprints
[dead]
I have a little thought to add...
(reads:) "AI Generated Content Floods Market, But Almost No One Is Buying"
So if that is a Thesis, in Terms of: heck the 'AI' has perfectioned it's own 'due to circumstances'-Copyright ??
I may add (reading further:) "AI" content in terms of "voxelized intermediate representations of objects" for example...
...converted to meshes, they will have poor topology. But topology is important if you want to modify the meshes or animate them -so
"bad 'ass'-hats (assets) AI technology exchange models may sound like a solveable prob, and lesser Ideological than hu politics which was added to those LLMs before, not?
Maybe pick a size first (limited) 320x480 if thats the old PAL, dunno...now let 'AI' slove the mystery you all seem to have...
And for motivation (btw), if there are some pixel-games falling off from, may be they are liked, too...
But as always, just my 2 Cents... (somebody listed US$20,- that may count a lot more...)
regards... (-:
(Edited typo^^)
Isn't this basic economics? Supply and demand. If slop floods the zone then human work becomes even more scarce and valuable.
What is my incentive to pay you? A person who spent no time working on something and who probably won't spend any time working with me if I need support. You're literally a meat puppet and a useless middleman when I could pay $20 and do it myself.
Unfortunately we've experienced the basic economics of spam. It costs virtually nothing to spam emails and so the spammers exploited that. Now it costs virtually nothing to spam markets with AI slop. Platforms need anti-slop filters and the easiest one is to charge x dollars per model to publish there.
yep basic longtail effect. back in the day at a PRO (performing rights organization -- music copyright stuff) we would analyse the longtail to death because the ~10% artists would get ~90% of play time (=> apportioned funds).
i guess in this case and probably some other cases that AI slop is just bulking up the longtail even more than before. 90% of people don't listen to AI generated music, it's only weirdos on HN that i've seen claiming to want access to it for personal listening.
Do you even realize how ineffably much an artist may imagine, consider, realize, choose while actually creating? It may be weeks, months, years... when artists developers, creates, enchants their unique miraculous creation... making it invaluable... an actual art...
These human experienced choices are immeasurably important where no "AI" should ever exist... The effortful lovely choices, where every single choice polishes an artwork... reifying it into a perfect or an imperfectly perfect miracle...
AI Creators and NFT Bros are cut from the same cloth. Just polluters who think because they can type on a keyboard it makes them special despite having spent most of their lives as talent-less hacks.
I won’t be surprised if 3D modeling and rendering software will soon add some sort of mechanism or watermark or a hashfile to prove that it was modeled with this software rather than some slop.
Or you can just look at the geometry. 3d slop is orders of magnitude easier to spot than text slop.
Yeah obviously, it’s better tho to filter the noise in advance, but if AI slop flooded a market place, real designers might stop uploading and sharing their work.
The buyers are saying it's because the human-created models are better. This might be true, but it could also be some other things. Maybe they're doing the AI dream and using AI to create the models themselves. The article cites revenue maybe AI model are selling, just cheaper. Maybe human creators are buying from humans out of solidarity.