Could AI ever get into 3dcg and animation? Posted on December 28, 2022 by Anonymous Could AI ever get into 3dcg and animation?
photogrammetry is about as close as it can get to something usable
The fuck are you babbling about?
It is still working like utter ass and taking too long. There's a whole universe left for improvements.
UV, Animation, Texturing, Modeling, and Retopo all have AI assisted software now. Only thing I can't think of is rigging and weight painting, but I wouldn't doubt there isn't AI for this too, or being worked on.
No. Best auto option is still traditional algo.
Might become viable options in a few years
> Animation, Texturing, Modeling
Very primitive. Decade away from production readiness.
Texturing is the only thing I can think of that is good with AI so far. I have the stable diffusion texture addon in blender and it works very well for simple textures. extremely convenient for things like wood, walls, grass, etc.
For me the quality isn't good enough and not worth the hassle yet.
There's no shortage of high quality pbr maps for most surfaces.
I don't do NPR stuff though so it might be decent for that.
using ai to rig basic bipedal two armed models shouldn't be too difficult considering we can assign bones and determine their position for vrchat using nothing other than a standard webcam and Driver4VR, if some dude in the EU can do it from his basement I'm sure people with more field knowledge can do better. https://www.youtube.com/watch?v=D5x_aqM0rAw
Certainly, it already exists in rudimentary form.
Works by first generating 2d images, converting that into neural radiance field and then into a mesh.
The outputs suck (low res, bad topology and glitches) and inference is slow af.
Suitable for background scenery if you really want to use it productively.
More sophisticated approaches are possible where the model is trained on a dataset of 3d meshes, but meshes aren't as easy to come by as images so scraping them isn't really feasible.
I suppose you could extract them from a large library of video games but this isn't something a commercial entity would do because it would be a legal minefield.
Buy the games, extract the datas.
Nothing of the OG remains in the model so it'd be fine.
>Buy the games, extract the datas.
>Nothing of the OG remains in the model so it'd be fine.
Hmm, you might be right.
I was thinking about the various end user license agreements that would supposedly prohibit that.
But then you don't exactly need to adhere to it when you're not playing the game. Not that this would stop a game publisher's legal department to try something. These are the kind of entities that try to extract cash from torrent users.
Anyway the whole legal stuff is a mind fuck, I have to stop thinking about it.
Back to a more technical concern, game models usually are stored in triangle topology but for source models you want quads. So an implementation detail might be to recover the original quad topology.
>But then you don't exactly need to adhere to it when you're not playing the game.
yes you do
They can refuse updates, lock your account, etc... essentially refuse to do further business with you.
All of which you don't care about if you just want to look at the data contained in the game files.
AFIK there's no case of litigation for a TOS violation. This goes along similar lines as the right to repair issue.
If you purchase something, you can do with it as you damn please.
>If you purchase something, you can do with it as you damn please.
No you can't, it's lllegal, you're making up retarded shit
Nothing will happen to you of course, but it's not legal
cite some sources so we can have a real debate.
I'll start: https://www.findlaw.com/legalblogs/ninth-circuit/violation-of-a-websites-terms-of-service-is-not-criminal/
simply violating tos is not a crime, if you do anything with the models you extract from the game though, you're breaching copyright and they could sue you if they found out and they wanted to
>if you do anything with the models you extract from the game though, you're breaching copyright
That's not how copyright works, as long as you don't re-publish the original data in parts or it's entirety you are not violating copyright.
It's the same as with ripping your own music or movies.
That is unless you agree with the "artist community" sponsored by media companies that are trying lobby to _change_ copyright law to _make_it_ illegal.
If you used it to train an AI and then you published what it generated, you would be republishing the work
Although that would be pretty hard to prove
>If you used it to train an AI and then you published what it generated, you would be republishing the work
According to this argument all publicly available AI models would be subject to that.
This is the argument brought forth by the people trying to outlaw AI generators.
Spoiler: Current copyright law sees content generated using AI as "inspired work" which is legal.
This is evident by the fact that the images generated by SD & co don't look like copies even if you formulate the prompt in such a way that it should produce a certain image by a certain artist.
If you prompt "The Mona Lisa by Leonardo Da Vinci" you won't get The Mona Lisa but an image that looks like someone looked at it and painted it from scratch, which is also legal.
AI generated work isn't "inspired work", it's just that you can't really pinpoint one discrete copyright infringement on a generated work so who is going to sue the creator
I disagree, but neither of us are lawyers. Guess we will have to wait and see what comes out of the inevitable lawsuits next year.
that's not my opinion, that's the law
if there was an AI trained on the works of just one person or copyright holder, they could absolutely sue the person who did it
the issue is they train it on lots of stuff
>thinks reading a blog means understanding a nuanced legal issue.
No you don't.
I have read no blogs on the topic
I work in a field that deals with this issue
>I work in a field that deals with this issue
If you did that you would have explained it by now.
Oh yeah? Well I also work in a field that deals with this issue. You're wrong.
I know it's coming, I just hope it isn't cheap.
t. 3d modeler
No not possible, go ahead and study for 3 years, become a master.
Is there even software that can create 3d from rudimentary nets? Obviously if you know the exact net you don't need an algorithm for it, but, like, is there something that can create 3d from side and top views?
>is there something that can create 3d from side and top views?
you don't need AI for that
You can get away with any number of errors in 2d but you can't always get away with horrible topology in 3d. No doubt ai will be used to produce background objects within a couple years but I don't see it being capable of making anything from start to finish that is both animated and shown right in front of the camera for a long time.
topologizing in 3D is not difficult for a machine to do and can be done with procedural techniques without even having to resort to AI
yes macgay, a.i. will eventually be able to do everything
Could you ever get into an unemployment office?
You've never held a job, have you OP
yes but there is a very important difference - namely that 3d models need to be optimized for performance and texturing and animation, and they need to match concept art
>yes but there is a very important difference - namely that 3d models need to be optimized for performance and texturing and animation, and they need to match concept art
Performance, texturing and animation optimizations all are very technical and are something that an algorithm is inherently very well suited for.
You could go straight from concept art to a uv-mapped, textured and rigged model with matching topology.
You could use prompts to define the idiosyncrasies of the model.
For example, "add bones for the eyebows and eyelashes, add look at target to the eyeballs, teeth and eyes should be a separate mesh" or "use a brick wall texture with normal mapping, add a tiled roof and a stone masonry chimney"
Either by SD style prompts or better yet by the time we have it a ChatGPT style interface.
Protip: Optimization can be done with AI, and it is a specially good target for that.
Think of it like this, deep learning is all about maximizing a function, in other word, optimize.
The only difficult part is gather a good dataset of unoptimized vs hand optimized models. Once you have that, a model can be trained easily.
Fuck I really really hope so.