What are your thoughts on AI on hard surface modeling?

DrBombs

What are your thoughts on AI on hard surface modeling?

Do you guys think in the next few years it will be able to make really clean meshes?

poopipe

There's already generative ML CAD tools. 
For meshes though, I don't think we'll see anything for a few years - at least not in the form of a commercial product. 

An ML driven retopology tool is entirely possible but it's unlikely that you'd be able to generate a decent return on investment quickly enough to satisfy the sort of people that are throwing money around at the moment. 
It'll all chill out eventually, costs will drop to the point where it becomes a viable commercial option and we'll start to see it all trickle down into our tools I'm sure. 

I wouldn't want to estimate a date but I'm pretty sure there'll be a lot more fast money to be made investing in other industries for at least another 5 years 

Shrike

These are like estimations from last year

At current pace and recent breakthroughs 1 year max that you get 80-90% quality photoscan style assets which are usable
AI image generation evolved from "kinda usable" to "1 word photorealism pixar quality" in a couple months
I mean just look at this
https://styledrop.github.io
All these 3d icons were made based on only the left side reference image as input

Getting this into some topology thats usable is surely not far away (does exist already but someone's gotta make a good version), depends on what your standards are
But its not like people care about scanned asset topology
Id say 3 to 12 months how things are progressing right now to get high quality 3d model generation.
- You can make photoreal references with solid continuity from many angles
- You can imagine a 3d space just from a few images
- You can make textures
- You can generate point clouds
Someone has to simply combine all thats there in a good version, theres not too much missing here
If machines can learn something then topology, there are very clear patterns and best practices in topology with highly consistent logic.

pior

Yeah, I think the mistake that many people make is to assume that for 3D generative AI to work, it would need massive amount of 3D models to train on. I think this is incredibly short-sighted, because generative 2D can already create convicing pictures that look like perfect renders (see above). Hence extracting turnarounds and then 3D information from that seems really quite doable.

IMHO automatically generated 3D stylized sculpts will be there before AI-driven retopo.

dimwalker

I will be the sceptic one here to balance it out then.
AI will generate nice models for sure, but I expect it to be similar to what's going on with image generation atm - it's amazing if you don't really care what you get and just want a pretty picture. If you have very specific goal it can be a lot of frustration.
 Maybe it's different for other NNs, can only speak for MJ since used it mostly. Some things are impossible to generate reliably if at all, it's all about luck and  result would probably still require fixing in photoshop. Atm it's not a tool for production, though is quite nice for inspiration, concepts, quick mockups.

I expect 3D to be similar for a while.
No support loop here, infinite density triangulated spot in the middle of flat quad surface, weird diagonal cut going through whole mesh. Other stuff like that and a lot more that I can't even think of that would make rendered model look like a kaleidoscope pattern.

Eric Chadwick

The auto retopo modifier in latest 3ds Max is already solving this problem, imho. It makes really nice looped topology, with very little user guidance needed. 

The current problem I see with AI models is the baked in lighting, no PBR textures, blurry transitions between material types, no bottoms, poor handling of high frequency details, etc.

iam717

can we compile all these a.i. threads into one, its a bit annoying how it keep popping up a new thread every week almost like hammering until we "agree to it?"... whatever i can also not come here anymore which is what it will do eventually.  So if that is the goal you are 100% on the way there, i also know no one cares but i still want to say it.

poopipe

This is a bit of a different topic.

Will we see a robot generate useful topology?  isn't the usual 'robots are taking my job' post
 
It's not a question of whether it could or not, the answer is yes. When it comes to when I still don't think anyone's going to invest the money required into a commercial solution for a long time

Where it would likely improve over things like Max's auto retopo is in intelligently detecting features. Any conventional algorithm is really just looking at curvature and markup such as hard edges - any concept of shapes and forms tend to be quite localised .

Pior is 100% right about it being relatively straightforward to generate a model
We have done most of the work already for 2d generation and it'll get better quickly.