-
Ever wanted an RSS feed of all your favorite gaming news sites? Go check out our new Gaming Headlines feed! Read more about it here.
-
We have made minor adjustments to how the search bar works on ResetEra. You can read about the changes here.
AI Neural Networks being used to generate HQ textures for older games (You can do it yourself!)
- Thread starter vestan
- Start date
Hopefully one day AI can just make the entire game. Then all the game devs would be freed from working those 100 hour per week jobs and could find more normal lines of work. /s
Hot damn. This is very exciting stuff for the future of re-releases -- I'm the type of person who is totally fine with legacy games simply re-released with better image quality (rather than demanding full re-masters), so this is really neat to me.
Socom II... *prays*
Edit: This won't be a popular opinion because I know folks get a little sanctimonious about pixel art but I've never liked it much and that pic is so much more interesting looking... maybe because it's weird looking? Idk, but I prefer it.
Edit: This won't be a popular opinion because I know folks get a little sanctimonious about pixel art but I've never liked it much and that pic is so much more interesting looking... maybe because it's weird looking? Idk, but I prefer it.
Last edited:
OP
OP
I'd play this tbh
ok this is magic..
basically publisher can just use this to do like 70% of the bulk work for remaster while the remaining 30% will just be use to touch up by the devs. Way cheaper and efficient to do remaster now!!!
I wonder is this a Nvidia exclusive feature?
OP
OP
Exactly what I was thinking. Can't depend on it entirely but it can give you a base to work with. Humans do the rest.
For now.
Though it is clearly not perfect, the results are very impressive.
I don't expect this to change game development much, except possibly for remastering older games. These neural networks typically produce significantly better results when the source image has been created from a higher resolution asset that has been downsampled/compressed or pre-rendered. If I am not mistaken, I believe the Doom sprites were originally photographs rather than hand-drawn.
I don't expect this to change game development much, except possibly for remastering older games. These neural networks typically produce significantly better results when the source image has been created from a higher resolution asset that has been downsampled/compressed or pre-rendered. If I am not mistaken, I believe the Doom sprites were originally photographs rather than hand-drawn.
you're still going to be stuck with the same number of polys. It's for textures/bgs.
OP
OP
For their textures yeah
No idea if I'm allowed to post textures ripped straight from games on here but you can try it for yourself by following the guide in the OP.
this improve texture quality and resolution..it doesn't add more polygons..
also, if the texture is just single shaded/gourad shaded originally (think the character graphics in FF7), i reckon it can't really do much
Where does this go, though? Can A.I. be trained to include results in texture upscaling into extrapolating more polys on low res models?
I realize I'm about to create an internal monologue for a neural network.
"I was trained to see this in a 2D image, and things that look like what I saw take 3D shapes kinda like this, and that 2D image is applied over a 3d topography, so I can compare the two and then I can guess based on what I know about both spaces and the interpretations I made to then..."
Yeah I was thinking something like a Gabe Logan where he's super low poly but has a big gross texture for a torso. lol
I don't know if you would ever want that. The computation involved means that'll most likely be strictly worse than just generating higher-res textures beforehand.
Most likely it can be done, but I imagine the limited use case (old videogames) makes it less tempting to spend time developing on as opposed to enhancing imagery.
it will be madness if that is even possible but wildest things had happened (like Google Camera's Nightsight) so who knows at this point.
Some of these have levels of sharpness that are rather unpleasant looking, and some are even quite straining to my eyes. Some extra touch-up might be preferable for a couple of them, and others seem unsuitable. Still, results look pretty impressive. Makes you wonder where this sort of thing will go in the future.
As an Nvidia investor, i hope this can bring in some $dough$ which in turn increase the $share price$ :P
i agree with you that in terms of using Neural network AI training to enhance gaming, Nvidia is definitely at the forefront of everyone in the gaming industry. I mean their latest DLSS for turing cards is essentially using AI to output a more efficient way for anti-aliasing.
Nvidia built an entire 'renderer' from a neural network trained on on images from cars driving in traffic:
https://www.hardocp.com/news/2018/12/03/nvidia_uses_neural_network_to_render_3d_environments
While very resource intensive and poor looking, it's still a really cool and impressive result. ML will factor into games increasingly heavily as time goes on.
Yes guide is in the op but if you want to do a whole game it will take a while especially on your card.
Thanks. I will try it but I was thinking of upgrading to a desktop PC next year so maybe I will wait until I have a new machine.
OP
OP
Follow the instructions on this page.
https://github.com/xinntao/ESRGAN
Naturally it'll be way slower for you if you have AMD as your only option is using your CPU.
I'm not quite sure how to run it despite instructions.
It says you need Python?
Now I´m wonder when we will get neural networks that do things like smart increase of polycounts on 3D models (that's not just subdividing the whole thing indiscriminately) and interpolate between frames to generate smoother animation frames that look more lifelike.
For audio there's also technology out that can increase fidelity of low sample rates by guessing which overtones are missing.
For audio there's also technology out that can increase fidelity of low sample rates by guessing which overtones are missing.
OP
OP
Yes. You need Python 3 and PyTorch
https://pytorch.org/get-started/locally/
https://www.python.org/downloads/release/python-366/
It'd look like shit because they're already large in dimensions. This kind of thing is better for smaller stuff. Take a look at the Morrowind rock image I posted on the last page.
Is there a particular reason to go with Python 366 and not 371?
Do I need to install Python to run, or is there a portable option?
Also just to emphasis, I have an AMD GPU, but the instructions still tell me I still need CUDA? Wha?
Edit: Also it tells me to go into test.py to change settings? How do I do that?
Last edited:
