Does something like DLSS have a future in rendering?

[Xntric]

It's incredible how these AIs can upscale and extract details!

Personally i have tried to use topaz gigapixel to upres my renders, but it doesn't work very well.
Their AI is probably trained more for real images and certain objects, cause i have tried to upscale textures too, and it doesn't really add any detail either.

But i imagine if you trained an AI to upres CG images, and maybe gave it access to some of the renderdata, like normals and such, that it could be great.
Perhaps Nvidia could even provide their DLSS tech for rendering.

This is all just speculation from my side, but what do you guys think?
I think it's some very exciting technology.

[dougWilkinson]

yea this tech has been in Vray in the form of the NVIDIA AI denoiser. I'm not sure of the differences between DLSS and the Nvidia AI Optix denoiser, but they are essentially the same thing conceptually. DLSS is def not perfect, even the 2.0 version struggles with patterns and other details it hasnt been trained on. I wouldn't WANT to use this for final renderings, but the way its implemented in apps like vray and blender is smart; its the default for interactive IPR rendering but off for final frame rendering. the chaos denoiser is better for final renders, IMHO.

The only render engine with custom tuned AI upresing is octane (AFAIK). its pretty good, for what it is. i wouldn't WANT to use it because like all neural networks its never as good as the real thing, but its a great concept and when its trained specifically by the render engine devs it seems to be alot more useful.

So my thought is that DLSS is is a solution to a different problem than "vray". I think we'll see this tech evolve quickly tho.

if topaz made a version of gigapixel trained on CG renders that would be very helpful!

[Xntric]

Although both are based on AI trying to guess what should be in the image, i still think AI denoising and AI upscaling is quite different, but what do i know.

DLSS seems to be working great on games, but i guess it has been specifically trained for those games.
Could be awesome for faster previews or for animations. When there is a lot of motion in an animation, you aren't gonna notice the small mistakes the AI does anyway.

I think it could be very exciting

[vlado]

There are a few tools for upscaling already out there and some of them do a pretty good job.

Best regards,
Vlado