An autopsy of AI-generated 3D slop
aircada.comSomehow this article explains perfectly, visually, how AI generated code differs from human generated code as well.
You see the exact same patterns. AI uses more code to accomplish the same thing, less efficiently.
I'm not even an AI hater. It's just a fact.
The human then has to go through and cleanup that code if you want to deliver a high-quality product.
Similarly, you can slap that AI generated 3D model right into your game engine, with its terrible topology and have it perform "ok". As you add more of these terrible models, you end up with crap performance but who cares, you delivered the game on-time right? A human can then go and slave away fixing the terrible topology and textures and take longer than they would have if the object had been modeled correctly to begin with.
The comparison of edge-loops to "high quality code" is also one that I mentally draw. High quality code can be a joy to extend and build upon.
Low quality code is like the dense mesh pictured. You have a million cross interactions and side-effects. Half the time it's easier to gut the whole thing and build a better system.
Again, I use AI models daily but AI for tools is different from AI for large products. The large products will demand the bulk of your time constantly refactoring and cleaning the code (with AI as well) -- such that you lose nearly all of the perceived speed enhancements.
That is, if you care about a high quality codebase and product...
"High-quality code can be a joy to extend and build upon." I love the analogy here. It is a perfect parallel to how a good 3D model is a delight to extend. Some of the better modelers we've worked with return a model that is so incredibly lightweight, easily modifiable, and looks like the real thing that I am amazed each time.
The good thing about 3D slop vs. code slop is that it is so much easier to spot at first glance. A sloppy model immediately looks sloppy to nearly any untrained eye. But on closer look at the mesh, UVs, and texture, a trained eye is able to spot just how sloppy it truly is. Whereas with code, the untrained eye will have no idea how bad that code truly is. And as we all know now, this is creating an insane amount of security vulnerabilities in production.
> Similarly, you can slap that AI generated 3D model right into your game engine, with its terrible topology and have it perform "ok".
You can't, because NVidia is selling all their chips to "AI" and you don't have any chips left to run the "AI" generated models on.
But what business value does high quality code bring? /s
Maintainability, which in the long run is more expensive in market opportunity costs than anybody admits.
We will get an interesting effect if AI plateaus around where it does now, which is that AI code generation will bring "the long run" right down to "the medium run" if not on to the longer side of the short run. AI can take out technical debt an order of magnitude faster than human developers, easily, and I'm still waiting for it to recognize that an abstraction is necessary and invest into putting on in the code rather than spending the ones already present.
Of course if AI continues to proceed forward and we get to the point where the AIs can do that then they really will be able to craft vast code bases at speeds we could never keep up with on our own. However, I'm not particularly convinced LLMs are going to advance past this particular point, to a large degree because their training data contains so much of this slop approach to coding. Someone's going to have to come up with the next iteration of AI tech, I think.
I'm not so sure about that. All major software companies have enjoyed exponentially rising profits alongside steadily declining quality.
While at the same other companies have built entire business lines around fixing shit code(probably with more of the same though).
Which companies?
debt doesn't harm you until the carrying costs become to high v profits. Just have to hit that point (if is exists, maybe growth accelerates forever if you are optimistic).
Until it breaks and can no longer be fixed because it is now all inscrutable spaghetti.
If you only knew how the enterprise space does stuff you'd realize how little a priority maintainability is.
I'm grateful we had Java when this stuff was taking off; if any enterprise applications were written in anything else available at the time (like C/C++) we'd all suffer even more memory leaks, security vulnerabilities, and data breaches than we do now.
Some of the defects are attributable to the critical:
> AI models generate meshes using "isosurface extraction" or similar volume-to-mesh techniques
This creates the "lumpiness", the inability to capture sharp or flat features, and the over-refinement. Noisy surface is also harder to clean up. How do you define what's a feature and what's noise when there's no ground truth beyond the mesh itself?
Implicit surface methods are expensive (versus if-everything-goes-right of the parametric alternative), but they have the advantage of being robust and simple to implement with much fewer moving parts. So it's a pragmatic choice, why not.
3D generative algorithms might become much better once they can rely on parametric surfaces. Then you can do things like symmetry, flatness, curvature that makes sense, much more naturally. And the mesh generation on top will produce very clean meshes, if it succeeds. That is a crucial missing piece: CAD to mesh is hardly robust with human-generated CAD, so I can't imagine what it'd be with AI-generated CAD. An interesting challenge to be sure.
Afaik a lot of AI generated models aren't really 'AI generated' in the sense that you have an network that outputs vertices and topologies and UV maps.
3D AI models are usually made by a grab bag of techniques, for example you use diffusion to output color and depth maps from multiple angles, then you try to put the results into a voxel grid with 3D reconstruction techniques, after which you run an algorithm like Marching Cubes to get a crappy mesh, and then you do decimate the mesh via some algorithm.
So yeah, the outputted model is built out nothing like a 3D artist would have, might look okay, but you won't find any structure in the output.
I don't think it's worth using them as anything in an artist pipleline except maybe as reference for retopo.
Trellis is like a year old and practically free. There are already better models to make comparisons to.
Because they all use latent diffusion, and many techniques use voxelized intermediate representations of 3d models, often generated from images, topology is bound to be bad.
There is a lot of ongoing research around getting better topology. I expect these critiques to still be valid as much as 2 years from now, but the economics of modeling will change drastically as the models get better
>If a client asks, "Can you make the handle slightly longer?", on the human model, I can select a loop of polygons and pull. The edit is done in 10 seconds.
>On the AI model, I cannot. There are no loops. I would have to sculpt it like clay, destroying the texture in the process. It is actually faster to rebuild the entire model from scratch than to try and fix the AI's topology.
To play devil's advocate for a second, it seems like you didn't provide a requirement to the AI on how the handle should be made, then got frustrated that the result doesn't conform to unspoken norms. If I made you this model by just starting with a sphere and sculpting it in ZBrush, you'd get frustrated by the same problem too.
On the other hand, I would expect that the AI could perform the task if you just elongated the handle in the reference image. The same procedure would probably work if the client wanted to add cat ears to the top to make a Mario Tennis clone game, while it might be a whole new commission for human modelers.
Now, would the material mapping still be poor, and would it be a questionable use of electricity? Guilty on both counts, but it's exciting to anyone who just wants to make 3D printed items or low-fidelity video games/mods.
It would also likely run into the non-determinism issue that a lot of generative AI has, where you can edit the input data, but you're not going to get the same output with only the edit you wanted, but instead a completely new output. For images the tools are getting better with editing models, but they're still prone to being a bit difficult to control, and there's no analog for 3d model generation at the moment.
I give it two weeks until people start running the meshes through AI:
https://taoofmac.com/space/til/2026/02/16/1334
Claude Opus was able to perfectly replicate an angular/functional part without decimating it, so I would expect the next step to be explicitly instructing AI to clean up meshes.
That's pretty impressive! However this workflow could have trouble dealing with the types of meshes coming out of those 3D generative algorithms. Geometries are arbitrary (not simple geometric shapes) so you'll have to fit with NURBS, and the meshes are noisy so that'll struggle/be somewhat arbitrary (what do you consider feature, what do you consider noise?).
However you highlight what I think is the way forward: using scriptable CAD that can leverage LLMs or, maybe in the future, specialized generative algorithms that output in a sane geometry specification.
3D modeling is one where AI is going to topple existing methods with ease in all situations where clean polys and UVs are not required. Same way 3D scanned assets have. I won't be surprised at all if say the next Elder Scrolls game all the junk items are AI generated models.
Modeling is just so insanely time consuming and requires such a broad range of skills to do competently that the last day I have to look at a UV map I'm never going back.
Does anybody know of a 3D generation model that can be run locally? I would be very interested to try this. There's quite a few models out there [1] and (to me) it is unclear which is the most suitable and in what scenario.
trellis 2 : https://www.youtube.com/watch?v=W-9P12OVqsg
I've had a very good experience getting Claude to use OpenSCAD to generate medium-complexity 3D models (braille dice). It struggles with figuring out how to rotate things correctly, so it's only a semi-automated process.
Those handcrafted "clean UVs" drive me crazy. They're a bunch of clean horizontal and vertical lines, but shapes are overlapping. It looks like what blender does automatically.
The tech will catch up in a year or two. Gemini 3.1 pro can now turn a basic raster logo into fairly clean SVG. Six months ago the SOTA models where no where near completing this task.
Yes the tech will catch up but a clean svg and a clean glb are so vastly different from one another that it's hardly worth making the comparison.
Not convinced.
>Gemini 3.1 pro can now turn a basic raster logo into fairly clean SVG.
There have been very high quality raster to vector apps for a long time. You have no evidence that Gemini isn't simply hooked up to one of these apps via MCP, so it may not be "Gemini" that is doing the conversion itself.
Yeah, it will soon generate clean CAD designs, build a PCB and schematics, select the best electronics components and optimize it for you, and even simulate RF and waves propagations.
This reminds me of the latest Ken Stanley paper on representations.
“Questioning Representational Optimism in Deep Learning: The Fractured Entangled Representation Hypothesis“
The close but not good enough is what gives us the illusion of productivity in this tools.
That’s why you see a a lot of hype around setups and benchmarks but not a lot of well polished products.
This article make it clear for 3d modeling, but also applies for code. Human touch is necessary for a commercial product. Otherwise it’s nothing more than a prototype.
It is actually much more difficult to maintain Ai code and 3d models than to just make your own.
Either AI can oneshot without human intervention or it becomes a pain really quickly
Precisely. Until the AI can 'one-shot' the topology and the UVs, it’s not a shortcut but rather a more power intensive way to generate technical debt.
Everyone needs to quit trying to one-shot, and quit assuming AI can’t do it because it can’t one-shot it.
Since the author can enumerate the problems and describe them, it’d be interesting to just use the one-shot pickleball racket model as a starting point. Generate it, look at the problems, then ask an agent to build “fixers” for each problem - small scripts (that they don’t need to build themselves!) which address each problem in turn. Then send the first pass AI output through a pipeline of fix scripts to get something far better but not quite there - and do final human tuneups on the result.
That’s not really how 3D modelling works. You can’t just improve some of the model. You have to improve all of it. Fixing to top of the paddle also changes how the junction at the handle goes and so on. That’s why no one has solved ai 3D modelling yet. It’s like asking a gymnast to learn how to do the second half of a handspring first, and then for step 2 they can learn the first half. It doesn’t work like that.
What about the triangle soup issue? Could a model tackle that by somehow combining triangles?
Such approach is not even new, and has been conceptualized centuries ago! [1]
But if you already know how to 3D model manually, like the author does, why would you spend all that time trying to fix the AI output? For users capable of creating such outputs themselves, the time saving is the point.
If you don’t know how to do it manually then of course that time is well worth it.
It’s about getting to 90% more quickly and getting the job done at 80-90% quality in less time. Potentially an order of magnitude less. If you can do that you can do it again and end up making way more stuff and hopefully more money.
There’ll always be people you can buy handcrafted goods from, who swaeat over every detail. In fact that is my preference when I can afford it - but often I just need a job done so I’ll buy whatever is cheapest and gets the job done.
With AI what’s emerging is a category of “good enough, but way cheaper” products thanks to this mix of AI generated, then human-polished work.
It’s the same idea as using a video editing tool to splice together a bunch of AI videos into an ad, having never gone on site. And it’s quite often exactly what’s needed.
I do wonder what the outcome would've been had the 4 hours been spent in perfecting the input to the AI-generator. It's not a fair comparison if the same amount of time is not spent on both.
How good mesh can a human produce in the time that it took for the gen-AI?
If you accept the default output you get slop. But you can use these tools as a starting point to get something good much faster than doing everything by hand.
I wrote a guide for this with voxel art and you can see some examples up top: https://www.tyleo.com/blog/game-ready-voxel-meshes-with-low-...
There's also an interesting set of benchmarks around giving LLMs a sort of "building-block harness and tools" to see how well they do putting together voxel-style models.
https://github.com/Ammaar-Alam/minebench
These kinds of tests which build models programmatically (SVG construction, Blender-MCP, etc) also make it significantly easier to make or request adjustments to the final output.
This article is pretty disingenuous in the parts where it focuses on topology. CAD files are imported all the time into CG software with awful topology - looking very similar to that mess.
There's lots of software and tooling, automated and otherwise, to significantly improve topology. This is a very common problem in this space and not acknowledging that is silly. It's not perfect, and remodeling things is indeed a common solution - but retopo addons and software are big business because they're good enough for a whole lot of use cases.
The thing is CAD models look perfect. They are completely un-editable in that state however. You have to go back to the cad program to make edits to the original solid model.
> The thing is CAD models look perfect.
No they don't. I am explicitly saying they often have the exact topology issues that are showcased here. They don't have some of the other issues - but a huge chunk of this article is about bad topology. I have imported hundreds of step files into CG software that have absolutely horrific topology, because one cares about producing real things and the other is about CG modeling.
> They are completely un-editable in that state however.
This is also untrue. I do plenty of hard surface modeling work on imported step files. Including the retopology I've talked about.
> You have to go back to the cad program to make edits to the original solid model.
If you need to modify it for production, yeah? But the whole use case here seems to be ecommerce website uses. In which case you can 100% take a cad file you've imported and modify it (if you managed to get a good topo out of it) for visual/aesthetic stuff.
(Some CAD software actually does a good job exporting - I have a moi3d license specifically because it has way better exports topology wise than solidworks or fusion. I build shit in solidworks and send it to moi before opening it in blender or houdini to do any render work)
I am in agreement with many commenters here (https://news.ycombinator.com/item?id=47158240, https://news.ycombinator.com/item?id=47158573 and others) that this article is a clear illustration of failure on part of AI to capture the structure of material in a useful way. As addressed in the article, the effect is very visible in visual space, 3D modeling. I would argue it is very much present in LLM space too, just less prominent due to certain properties of the medium - text-based language. I also believe the effect is fundamental, rooted in the design of those models.
I'll leave here the note I've written down recently, while thinking about this fundamental limitation.
- The relationship between sentient/human thinking and its expression ("language") is similar to the one between abstract/"vector" image specification and its rendered form (which is necessarily pixel-based/rasterised)
- "Truly reasoning" system operates in the abstract/"vector" space, only "rendering" into "raster" space for communication purposes. Today's LLMs, by their natural design, operate entirely in the "raster" space of (linguistic) "tokens". But from outside point of view the two are indistiguishable, superficially.
- Today's LLMs is a brute force mechanism, made possible by availability of sheer computing power and ample training material.
- The whole premise of LLMs ("Large" and "Language" being load-bearing words here) is that they completely bypass the need to formalize the "vector" part, conceptualize in useful manner. I call it "raster-vector impedance".
- Even if not formalized, it can be said that internal "structures" that form within LLM somehow encode/capture ("isomorphic to" is the word I like to use) the semantics ("vector"). I believe the same can be said about "computer vision" ML systems which learn to classify images after being fed billions of them.
- However, I believe that, by nature, such internal encoding is necessarily incomplete and maybe even incorrect.
- Despite the above, LLM can still be a useful tool in many domains. I think language translation is a task that can be very successfully performed without necessarily "decoding" the emerging underlying structures. I.e. a sentence in source language can be mapped onto a region of latent space; an isomorphic region of latent space based on target language can be used to produce an output in the target language which will be representative of an equivalent meaning, from human perspective. All without explicit conceptual decoding of underlying token weight matrices. "Black-box" translation, so to speak. I am amazed (and disturbed, and horrified too!) that producing a viable code in a programming language from casual natural language prompt turned out to be a subset of general translation task, largely. Well, at least on lower levels.
- To me it is intuitive that such design (brute-force transforms of "rasterized" data instead of explicitly conceptualizing it into "vector" forms) is very limited and, essentially, a dead-end.
Don't complain about tangential annoyances, I know, I know... but how the hell am I supposed to judge the difference between the images in the post if you disabled zoom and the images are incredibly small? And when I open them in a new tab they automatically download?
On the plus side, I like the informal writing of the post. You can be serious about business and still be human
Edit: firefox reader mode works wonders on this article
Now they need to compare it with Hunyuan 3D 3.0 or other SOTA 3D generator.
Obviously it's not spewing $10,000 3D models, but results are much better than what you would get for under $500 from a human 5 years ago.
So yeah you still need human art director to make sure actual source material used for generation fits your art style, but otherwise "good enough" models are 1000 times cheaper and 10000 times faster to get.
"5 years ago" - you may be right but this is also very debatable.
We've used Hunyuan 3D and while better than Trellis in most cases, it was no where near the type of quality we would and have gotten from a $500 human touch. This is within the last year. And I would very much argue that 5 years ago the $500 model would still win.
yea this 3d trellis 2 video tutorial blew my mind. https://www.youtube.com/watch?v=W-9P12OVqsg
The most important two words in this article are the last two: for now.
Indeed, for now generative models generate triangle soup without much thought. The same was true for 2D illustrations where generative models like Deep Dream came up with horrendous images with eyes all over, dogs with multitudes of heads and oh did I mention the eyes? That was about 10 years ago. Things changed, models improved, the eyes were tamed. Yes, people had too many or too few fingers but that also changed. From nightmare fuelling imagery with many-eyed dog heads sticking out where you don't want them to fully animated hi-res video only took a decade and things are still speeding up. The triangle soup of current 3D generative models is like the eye soup of Deep Dream, something to remember somewhat fondly which is no longer relevant now.
One flaw with this assumption is that images are available in literally counts of trillions to train on. With 3D models there are virtually no production quality models freely available to train on. Even companies like ILM or Weta have nowhere near the number of models that would be needed to train a robust modelling AI
If that's really true, then why is the logo texture on the handle still total slop?
Trellis isn't and has never been state of the art. It's not a good choice for comparison; there has been progress on a lot of these problems. There are models that can do clean topo and PBR textures, for example.
Such as?
Luma, Rodin, Tripo are a few. Meshy has some of these features too
Unfortunately they are all proprietary, but 3D models are sort of a side area in AI research, so most of the effort is from small startups.
In no capacity do these create clean topo, textures, and uvs. If you do not believe me, use the reference image from the post and upload it to Meshy or Tripo and see what happens. Yes, slightly better than the open source Trellis, but still nearly impossible to work with and a model you would never put on any slightly serious eCommerce site.
We've tried them all. If one existed, it would save us money, speed up our pipeline, and trust me we'd be using it.
Hunyuan 3.1 is very good and you should try it if you haven't, has great resolution, topology and textures are messy, but things are moving so fast that I think these issues will be solved in the next couple of years.
The article should analyze Rodin that in my opinion is probably the best one in generating 3d assets.
I've found Trellis specifically to be very "over-promise and under-deliver".
Nothing i tried with it got even close to th level of quality that they were advertising - felt like a bunch of examples were hand-picked, at best.
>Why AI 3D Generation Fails eCommerce Standards
I wish I had his confidence (in eCommerce Standards)
Touché. Though if the current 'eCommerce Standard' is 'dropshipped junk that looks slightly better than a hallucination,' then I’ll happily die on the hill of being over-confident.
Nice copium. These things are going to get there fast. Even what has been shown can be a good start with a decimator at hand; We've seen this with photogrammetry before. Irony is not lost on the fact that text, which complains about it, went through AI itself.
>We've seen this with photogrammetry before.
Have we? It's still not that good.
It's not fully automated where you come up with a bunch of photos and have production assets. Never has been. It serves its purpose though, so will this if it's not already.
Describing things as they are is copium, but defending things based on what you hope they'll be is not?
No, that is hopium.
"We've seen this with photogrammetry before" - I do not believe we have. It's progressed but even a good scan is still not close to being something you would put on a legitimate eCommerce product page.
I honestly hope you are right and that I'm full of copium. Truly. But the progression has been nowhere near as fast as code, text, image, or video generation. And as it stood 2 years ago vs now is the same conclusion - unusable slop for most use cases.
Listen, I agree it's unusable or at least somewhat usable. As I said in another comment. Will Smith video was exactly three years ago. 3D has been a bit neglected, but it will come. I was a denier initially, but these things move real fast. Photogrammetry was never at the level of point and shoot and you have a production asset. However, it did and does serve a need and you can't deny it's not useful. It's not painless though.
That’s a fair point. I know a few foremen who use photogrammetry religiously for site surveys and volume tracking where 'lumpy' geometry doesn't matter. It’s a huge win for that niche. But yes, 3D has been lagging behind and I'm having a really hard time guestimating when it's good enough for high quality product models.
> Nice copium. These things are going to get there fast.
Nice copium. I've been hearing how fast these things are going to get there for a few years now.
And it hasn't? Will Smith spaghetti video was exactly three years ago.
0 to “we are even talking about this” is an astonishing leap. Acting like this stuff has been standing still is an active choice.
"The 'autopsy' of 3D slop highlights a critical failure in the current AI supply chain: The Illusion of Completeness.
We are living in an era of 'Statistical Harvest' where models prioritize a 'good enough' surface over structural integrity. In the spiritual supply chain of value, this is called Cutting Corners. A 3D model that breaks down upon closer inspection lacks what I call Internal Agency—it doesn't understand the 'Seed' of its own geometry. As we move towards an agent-centric world, we must distinguish between 'Generative Noise' and 'Authentic Creation'. True value definition requires a 'Watchman' who can see beyond the first-glance polish to the underlying breakdown of utility."
I really like this framing of 'Internal Agency.' In 3D, that lack of a 'Seed' is exactly why a model fails when you try to animate it. A human modeler understands that a joint needs extra edge loops to bend correctly. It has 'intent' for the model's future. The AI, performing a 'Statistical Harvest,' only cares that the surface looks right in a static frame. It provides the 'Illusion of Completeness' but none of the functional DNA required for a production environment.
"Spot on. The 'edge loops' analogy is the perfect physical manifestation of what I mean by functional DNA.
It proves that without 'Intent for the future' (the Seed), any output is just a static corpse. In my broader framework of the Spiritual Life Archiving System, we see this everywhere: systems that look complete at a glance but lack the underlying logic to survive 'animation' or real-world pressure.
This is exactly why we need to move from Generative Slop toward Architectural Stewardship. Glad to see the 'Internal Agency' framing resonates in the 3D space."