What do you make of this AI texturing tool?

Blond

As a working artist in the VFX/Game Industry; I was sharing with a friend a piece of work that had been produced by an AI from a Youtube video and he as a working modeler became concerned seeing huge part and swathes of his work being automated; specifically as an artist in smaller mobile/independant studio.
The video in question; he's done a lot of textures works on mobile games.

https://www.youtube.com/watch?v=1VHCIKMrWHk

However I did remark to him that the finished result that the AI Gave was akin to a Baked Out image from which it became very difficult to make minute detail and change since you had to deal with a finished image you could only paint-over on and not a file with sets of layers where you fine touch and tweak parts of it layer by layer, group by group. 

Edit; Decided to change thread's overal direction,

sprunghunt

It doesn't look faster than using some pre made smart materials in substance painter. And you're right - using something like smart materials is easier to adjust later.

Alex_J

i had read the original, just didnt have anything to add.

its an important point that tons of value is within the actual process of creating art layer by layer and at least right now you cant get that with AI - it just makes a "finished image" and that's that. Not very much flexibility at all.

I bet anybody who has worked in the industry understands that. Probably some people with more money than sense who get pulled in by hype might not. Eventually they'll learn, or eventually the AI can generate all the intermediary layers retroactively given an image.

I guess if you were freelancing and its becoming hard to make 2d art because of AI, you might moonlight as a prompt engineer. Kind of like those grown men who pretend to be teenage girls and send foot photos to make $7k a month, lol. Just rob fools who have too much money.

Eric Chadwick

It's also using camera mapping, which omits the sides and rear. I guess if you had a multi-camera neural generator it might resolve those angles? The neural solver would have to solve all the angles at once. But would it make a blurry mess where they overlap, ala lazy triplanar?

Eric Chadwick

Oh oops I had the video muted. OK they reproject multiple angles, and can adjust the blending some. Still, pretty ugly results with directional lighting baked in. I guess you could adjust the prompts to change the lighting. But yeah, seems pretty limited at this point. 

poopipe

It's not a massive leap to derive diffuse/specular/roughness values etc. - I saw an ML paper a couple of years ago that was able to derive pbr valid materials from photos etc. so it's just a case of refocusing that effort. 

I think this sort of thing is going to hit outsource providers and those working on shovelware first but it'll spread out as it gets more controllable and if you're not able to do more than the robot you're going to find yourself getting replaced. 

Alex_J said:

 Kind of like those grown men who pretend to be teenage girls and send foot photos to make $7k a month, lol. Just rob fools who have too much money.

Does this work? My friend asked

Alex_J

Lol, just something I read somewhere, who knows. Hard to know what is true or not on the internet, they let anybody on here. I've heard certain strange markets that you might find near the bottom of the internet can be lucrative to serve. The only downside is that you have to live down there long enough so that the mole people believe you are one of them. Not that there is anything wrong with being a mole person!

I keep telling myself, maybe I'll take a break from all this high-effort work for a couple months and see if I can't sell some absolute trash for a high price. With the high-effort work the big fear is failure, right? I spend a lot of time to make a game or w/e, and then nobody likes it. Ouch. With the other idea, I'm more afraid of success. What if they love me, and I become one of them?

sorry, I won't derail any further.

Blond

https://www.reddit.com/r/StableDiffusion/comments/13lcv91/bringing_360_worlds_to_life_with_a_sketch/

This one here's getting a lot of hype too but behind all those flashy generated assets and decors, there's a lock of control and fine tuning.

The biggest questio now is how will all those AI tools give back that kind of control/trackback and optimization in the future?

Alex_J

I think there is definitely a lot of value in able to being able to quickly and with almost no effort be able to see something representative of a final scene - i mean concept art is a thing but if you can get concept 3d art that's even better right?

but yeah this definitely isn't actual art assets you can directly use... probably to some degree we might "trace" over some of the stuff to build individual assets from just to save a little time.

some people might say that it will lead to more derivative artwork, but I could see it going the other way too. if you can easily generate a billion variations, it will be easier to quickly find the mean and then you can spend your efforts on finding ways to pull away from that more. Whereas if I have to slowly and carefully do a lot of work to get to a final result there is more pressure to stick closer to designs that have already been "vetted" by the market.

Personally I don't see myself bothering with this kind of thing very much. Usually I have a clear mental image what I want - I don't think having more options on menu necessarily is a good thing, there is already a real problem with digital work like, I could create anything, there are endless possibilities, so how do I narrow down and just do one?

poopipe

IMO etc.

There's a lot more money in providing tools for the masses than there is in providing tools for the creatives (there's a lot more masses than creatives) As a result, the current pile of hype-driven generative ML apps are aimed at providing a finished image rather than a useful tool and are utterly useless outside of generating some fun ideas for a concept artist to paint over. 

When the hype and rapid-return investment panic dies down we'll be left with a load of interesting tech that can be redirected towards something actually useful and we'll start to see tools that legitimately extend the capabilities of artists. 

ML is extremely powerful in terms of replicating and predicting the behaviour of systems - this allows for all sorts of very abstract possibilities but if you want a couple of simple examples that aren't threatening to the artist... 
a: its a perfect slot-in upgrade to the (largely noise based) procedural systems we're all using today
b: it can do simulation (cloth/fluid/soft-body dynamics) much, much faster than running the sim traditionally

zetheros

seems pretty slow and inaccurate. I could paint a lot faster and accurately with just a tablet, until we inevitably see improvements, which then it'll be a great tool for quickly blocking out large areas of colour. I'm interested in seeing what large companies are doing. If they're smart, greedy, purely focused on short-term benefits to the detriment of all else, which they are - they'll already be well on their way in investing into proprietary software to generate reliable AI art. I'm sure we'll be seeing mass layoffs sooner than later at ABK.

Personally I'll never pay for anything AI related. If they stole from us, it is our tech by right. Just like tech companies dealing in user data. If they take our data without our consent, we get to use their services for free. Lucky for us that we're the most suitable people to use it.

Alex_J said:

With the high-effort work the big fear is failure, right? I spend a lot of time to make a game or w/e, and then nobody likes it.

your game will be great man, you won't fail if you make the game you want to play.

sprunghunt

I saw this today and It seemed like a good example of an actual practical use of machine learning
https://youtu.be/h-cUB4U2VWc

DrunkShaman

Considering why AI came into play in every other community. I would stay away from it. I would rather do it myself then ask a robot to do the job. That's just my 2 cents.

chien

i don't use any of these whatever AI they are call, i model or recycle models, or texture in substance