• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

Could AI in theory reduce file size of textures, making games much smaller storage wise?

Knightime_X

Member
Shower thought:
With games blasting well over 100+ gigs, a lot of that data is from big ass textures.
I wonder if ai can upscale smaller textures into much larger ones while retaining image quality but greatly reducing storage requirements.
My fantasy is games revert back to 10-60gb+ in file size. Maybe even less?
 

IAmRei

Member
cloud will not benefit player who cannot online much. and the size it self is actually depend on the developer to reduce without much loss in quality. and also there is some compression which could help the dev to ease the size. dev nowadays are mostly work with tight deadline, it may result of they doesnt have much time for optimization or such. it's actually in the dev side for this. cloud is not the only option i guess. cloud or not, it's depend on the asset treatments.
 

deriks

4-Time GIF/Meme God
Yes, since upscale is a thing, a AI upscale on the go could get a small texture and improve on it

How good can it go? Have no idea. Still, like boobs, native always wins

But I can say that Nintendo has a patent on this AI upscale. No joke
 

Alexios

Cores, shaders and BIOS oh my!
There was a thing called procedural textures before AI too but I guess it didn't see that much use outside crazy demoscene stuff like kkrieger (an FPS demo in < 100 KB) which is kinda nuts since without such extreme limits they'd probably have pretty good results & have improved the tech by now...

Of course certain techniques like these, AI or otherwise, may actually require other system resources to enable on the fly real time enhancement (including decompression) of stuff like this making it less beneficial, especially for weaker platforms than the norm like, say, the coming Switch successor...
 
Last edited:

Shifty1897

Member
Yes, this is exactly why the Switch 2 is leveraging DLSS. Those Switch cartridges only go up to 32 GB (the rumor is that Switch 2 cartridges will use the same or similar format) and using AI to upscale the image (including textures) saves on storage.

You probably won't see it leveraged as much on consoles where disk space is less of a concern.
 
Last edited:

yurinka

Member
Shower thought:
With games blasting well over 100+ gigs, a lot of that data is from big ass textures.
I wonder if ai can upscale smaller textures into much larger ones while retaining image quality but greatly reducing storage requirements.
My fantasy is games revert back to 10-60gb+ in file size. Maybe even less?
As far as I know from friends I have working there, minimum two major AAA studios already have been experimenting/working with AI upscaling techniques in real time to use in future games source textures smaller than the current ones we have in games to upscale and enhace them during the real time loading/streaming time, getting as result bigger and more detailed textures than the ones we currently have in games.

The thing is that it requires some dedicated horsepower and they either need to do it in loading screens or in what today are high end GPUs only. Unless they further optimize it, in consoles it would be pretty likely next gen only stuff.

Yes, this is exactly why the Switch 2 is leveraging DLSS. Those Switch cartridges only go up to 32 GB (the rumor is that Switch 2 cartridges will use the same or similar format) and using AI to upscale the image (including textures) saves on storage.

You probably won't see it leveraged as much on consoles where disk space is less of a concern.
No, this is a separate thing that pretty likely Switch 2 won't be able to implement due to too weak GPU and VRAM.

Switch 2 is supposed to have DLSS which is upscaling for the flat fullscreen 2D image after it has been rendered in a lower resolution. And unrelated to AI, a chip to decompress in real time the game data downloaded or in the cartridges in a similar way to how PS5 does it. These are two things.

What the OP mentions is a 3rd one to individually upscale & enhace textures that later are used to compose the 3D gameplay scene, which is complementary to the other two things. The idea behind this is that nowadays AAA games use up to 4K textures (and will be bigger in the future), which take a lot of space in both video ram and in the disc drive.

With these AI upscaling techniques, instead of having a let's say 4K texture source image, they could have an up to 4x, 8x or 16x times smaller one in the SSD, and in the process of loading it from there to the VRAM they'd apply an AI upscaler to turn it into a big ass one that would be stored in the memory.

Looking at the rumored specs Switch 2 won't have enough RAM to store big ass stuff, or horsepower to do such real time textutre AI upscaling & enhacing (different than/on top of DLSS).
 
Last edited:

Hudo

Gold Member
Yes, neural texture compression is a thing that has been discussed in published papers over the last couple of years. And it seems that Nvidia's SDK/APIs support that functionality as well.
 

hlm666

Member
If your interested in AI advancements for gaming I think you may aswell start paying attention to nvidia, they are basically the ones everyone else is following at this point.

 

poppabk

Cheeks Spread for Digital Only Future
How many game sales are lost due to file size? Probably not enough to make reducing file size a priority.
 

ReyBrujo

Member
File size is the least of gamers worries with ever increasing broandband speeds and cheaper storage upgrades.
Unless you want the full Flight Simulator 5Pb textures.

And yes, AI can do a lot of things, you can train the model to generate the game without having textures, dialogues or logic even. Differences aside, think of generative AI as the closest we are to Star Trek's Holodeck.
 

Knightime_X

Member
As far as I know from friends I have working there, minimum two major AAA studios already have been experimenting/working with AI upscaling techniques in real time to use in future games source textures smaller than the current ones we have in games to upscale and enhace them during the real time loading/streaming time, getting as result bigger and more detailed textures than the ones we currently have in games.

The thing is that it requires some dedicated horsepower and they either need to do it in loading screens or in what today are high end GPUs only. Unless they further optimize it, in consoles it would be pretty likely next gen only stuff.


No, this is a separate thing that pretty likely Switch 2 won't be able to implement due to too weak GPU and VRAM.

Switch 2 is supposed to have DLSS which is upscaling for the flat fullscreen 2D image after it has been rendered in a lower resolution. And unrelated to AI, a chip to decompress in real time the game data downloaded or in the cartridges in a similar way to how PS5 does it. These are two things.

What the OP mentions is a 3rd one to individually upscale & enhace textures that later are used to compose the 3D gameplay scene, which is complementary to the other two things. The idea behind this is that nowadays AAA games use up to 4K textures (and will be bigger in the future), which take a lot of space in both video ram and in the disc drive.

With these AI upscaling techniques, instead of having a let's say 4K texture source image, they could have an up to 4x, 8x or 16x times smaller one in the SSD, and in the process of loading it from there to the VRAM they'd apply an AI upscaler to turn it into a big ass one that would be stored in the memory.

Looking at the rumored specs Switch 2 won't have enough RAM to store big ass stuff, or horsepower to do such real time textutre AI upscaling & enhacing (different than/on top of DLSS).
Yes! This exactly!!
 

AzekZero

Member

yurinka

Member
The rumor is Nintendo is doing just that.
Nah, the patent is Switch 2's DLSS implementation, which includes a lightweight variant that gets activated once the game doesn't achieve certain fps, in addition to local or server side upscaling for Nintendo's cloud gaming service.

Alex is wrong when assuming that in the following image "output images at 4K", "images produced by the video game engine" listed here are textures when it isn't the case, because textures are input images not produced by the engine. The output images produced by the engine are instead the rendered frames that DLSS upscales.

image.png


What Nintendo is saying in this image is that since with DLSS you upscale engine's 1080p native resolution output frames/images to 4K, they don't need to use as big/detailed textures/models/etc, which reduces the total disc/cartridge size needed.

Meaning, let's say a PC game that has a setting to choose texture sizes, one in Ultra Textures is 4K and other one in Low Textures is 1K.

Let's say for that game the PS5 port uses the Ultra textures and will render in native 4K (or close), and that Switch 2 renders at 1080p and upscales from there to 4K via DLSS so it's ok for Switch 2 to use the Low. So -compression apart- the game file would be smaller in Switch 2 even if it's outputting at 4K (after rescaling via DLSS), because if the native resolution will be 1080p (or 540p/720p/etc) won't make sense to put there 4K textures, which in any case wouldn't fit in Switch 2's memory.

Switch 2 (not including here the cloud gaming variant) just uses DLSS to upscale the engine output from a native resolution to other one. The OP instead means to upscale textures when loaded from the SSD/cartridge, not the engine output.
 

Gambit2483

Member
Not only is it possible but I'm pretty sure this is one of the major reasons Nintendo partnered with Nvidia implementing DLSS into Switch 2.

They desperately want to keep files sizes down for these ginormous next gen games that will need to fit on Switch 2 carts/SD cards
 

Gambit2483

Member
Nah, the patent is Switch 2's DLSS implementation, which includes a lightweight variant that gets activated once the game doesn't achieve certain fps, in addition to local or server side upscaling for Nintendo's cloud gaming service.

Alex is wrong when assuming that in the following image "output images at 4K", "images produced by the video game engine" listed here are textures when it isn't the case, because textures are input images not produced by the engine. The output images produced by the engine are instead the rendered frames that DLSS upscales.

image.png


What Nintendo is saying in this image is that since with DLSS you upscale engine's 1080p native resolution output frames/images to 4K, they don't need to use as big/detailed textures/models/etc, which reduces the total disc/cartridge size needed.

Meaning, let's say a PC game that has a setting to choose texture sizes, one in Ultra Textures is 4K and other one in Low Textures is 1K.

Let's say for that game the PS5 port uses the Ultra textures and will render in native 4K (or close), and that Switch 2 renders at 1080p and upscales from there to 4K via DLSS so it's ok for Switch 2 to use the Low. So -compression apart- the game file would be smaller in Switch 2 even if it's outputting at 4K (after rescaling via DLSS), because if the native resolution will be 1080p (or 540p/720p/etc) won't make sense to put there 4K textures, which in any case wouldn't fit in Switch 2's memory.

Switch 2 (not including here the cloud gaming variant) just uses DLSS to upscale the engine output from a native resolution to other one. The OP instead means to upscale textures when loaded from the SSD/cartridge, not the engine output.

This is just false, based on the recently released patent.



Perhaps the most interesting piece of this, after a lengthy read, is that one example use case given is explicitly to reduce overall game sizes, to fit a modern game onto "smaller capacity physical media", e.g. Switch carts, which get exponentially more expensive for larger cart capacities
 

yurinka

Member
This is just false, based on the recently released patent.
You're wrong. I made a TLDR of the patent/DF video and particularly the part that DF wrongly says talks about textures, explaining why he is wrong, including the related screencap of the patent they did put in the video to talk about this, and also explained why they save game size by using DLSS (since they are going to render the game in a way smaller native resolution they don't need big textures).

When the patent talks about "output image of a game engine" isn't talking about game textures that the engine uses to build a 3d scene, it's talking about the result image that the engine rendered at a given native resolution, the one that scaling solutions like DLSS use to upscale.

If you think that a game textures are "output images of a game engine" you have no fucking idea of how gamedev, game engines or DLSS work.
 
Last edited:
Top Bottom