Thread 720521959

Nvidia was right. Get ready for a future where 8 GB VRAM would be enough for gaming.

https://arxiv.org/pdf/2407.01866

Woah that's CRAZY! The same two pictures held side by side with words around them!

None of this shit ever delivers in practice

>>720522064
read nigga. one is almost 30 times smaller in file size than the other despite looking almost the same.

>bro just apply an extremely compression and sharpen it with a clarity shader
nvidia's out there doing optimization for (((game devs)))
i knew women were incapable of doing anything but this really takes the cake

>>720522141
You have no idea. Neural texture compression for assets can be done using Nvidia's tools right now. Developers just aren't making use of it.

>>720521959 (OP)
Depends on the entropy of the texture. You can see in the reference that they used that just from your intuition it should compress pretty well because of the background and parts. It's low entropy.

>>720522289
>le compression bad
retarded take. bet you think DLSS is shit as well.

>>720521959 (OP)
>posts low detail blurry images so losing detail means nothing

>>720522536
They have several references in the paper. It works just as well for most of them.

Developers are hacks if 12 MB ram on a Voodoo 2 isn't enough. You should never have to upgrade your hardware. No, this isn't because I'm poor I swear, I'm taking a principled stance.

>>720522467
It could work but regular tools should already be able to compress it very well, BCn is already in use today. Actually it could be compressed much better but that would require changing workflow because base color + noise is usually used to generate these PBR style materials and since the data is generated procedurally and then artist sculped to what they want to resemble. You could perhaps store the steps of the seed/PRNG with each mask applied directly in shaders, which is done probably in some niche applications already but it's always a trade-off between caching things so memory and compute having to do the work of decompressing within the target framerate.

>>720522559
Yes.

>>720522814
>BCn
BCn is said to be not even close to being as efficient as the intel / nvidia's new neural compression tech.

>>720522176
how long does it take to compress and decompress?

>>720522176
No, the file size is 547 KB whether I'm looking at the top or bottom picture. It says so right above them both

>>720523180
30 seconds to a minute.

saar look at all the free ram, now we can ship 24k resolution .bmp files as texture

>>720523180
Just look at the pass times in these screens
>>720522467
>>720523056

>>720523273
so op wants 0.03 fps gaming?

>>720523226
yeah but if you cropped out the top half of that pic it would only be 24 KB.

>>720523056
Where did I say it was? Also I wrote a neural net compression test myself and could compress it decently and after reading about it years ago I already foresaw this being a gamechanger but it always boils down to inference speed. Besides this ongoing research has been in the works for years now. https://arxiv.org/pdf/2006.09661 It's only recently that you're starting to see NVIDIA and others implementing parts of it.

why don't they just use multiple smaller textures of varying sizes depending on how detailed part of a model has to be instead of assfucking everything with 30 separate 8192x8192 auto-generated textures of skin pores and nose hairs

>>720523397
>3x slower than png
yeah it's not gonna work for games aiming at least 30fps

>>720523406
No see, the bottom is actually the bigger side. Nvidia is full of shit

>>720523541
They do it's called UV unwrapping with texel density and mipmaps.

>>720522467
And they won't, just like we never got the 500%+ performance increase we were promised with mesh shaders and VRS and all that other shit. Even on the off chance developers do start using it, it'll just be another crutch like MFG that enables them to get even fucking lazier and end up with the same or worse relative performance.

>>720521959 (OP)
it will never be enough

This looks like little more than a proof of concept that doesn't yet actually have the stats to mean a damn thing. You'd quite literally be better off just making lower resolution textures that have enough detail on them to get what's needed across, than to use an algorithm to do it for you that adds extra processing pass time to the mix to.. fake a lower texture quality for better gains elsewhere.
If someone is stuck on an 8GB of VRAM card in the first place, odds are their RAM or CPU taking those extra hits may not be great.

>>720523881
https://www.youtube.com/watch?v=OrTyD7rjBpw

>>720523402
That's like 3 fps using multimodal frame generation.

>>720523580
Actually this paper was from Intel and AMD.

>>720522910
You literally can't play in 4k 60fps without DLSS.

>>720523945
Why would the RAM or CPU be involved? Presumably, the decompression model would live entirely on the GPU

>>720522467
>2466.2 fps v 2088.3 fps
Nope. Nearly 400 fps drops are not acceptable.

Precompute a path of playable space then train an AI or store that data to generate what the density of each texture asset will be at each position and generate a long strip of textures at the max possible resolution needed for interpolation.

(c) if you steal this idea and employ it in your game or game engine you must use the +NIGGER license.

>>720525218
That's only with a single asset... imagine if you had to uncompress every texture in an entire scene. This shit is going to be useless. It would be cheaper just to add more VRAM than make a GPU more powerful to get the same amount of FPS

>>720525254
congratulations, you just made the game look like dogshit.

A proprietary solution is not a solution, just another trick to lock down the PC environment.

>>720522141
Because by the time another extraterrestrial rover gets exploded to some other solar body, the hardware and software used will be hopefully better. If this boondoggle does work, at least even a little bit, it would be a pretty good thing for smoothing out interplanetary data transmission.

>>720530026
It's not proprietary.