I think ML is going to be huge no matter what anyway, be it for reconstruction tech and other uses in gaming, to non gaming uses where theres even a bigger market. Its the same with ray tracing, its a graphical feature thats going to matter going forward in gaming and other uses. It might be costly for NV (and AMD, Intel etc) but those are investments for IHV's future.
Nvidia DLSS 1 and 2 antialiasing discussion *spawn*
- Thread starter DavidGraham
- Start date
-
- Tags
- nvidia
Or it would indicate that devs of games that have been in development for years, didn't want to tweak their pipelines in the middle of a project.
DLSS was an unproven tech with little support and limit to a small fraction of newly released GPUs when it was first introduced. I think only now, devs are seeing its utility as RTX gpus are far more widespread and the heavy cost of RT is practically unavoidable.
DLSS inclusion at the start of a project is far more likely now. And we will probably see fewer games adding it as an afterthought.
Last edited:
"To implement DLSS2, a game designer will need to use Nvidia’s library in place of their native TAA. This library requires as input: the lower resolution rendered frame, the motion vectors, the depth buffer, and the jitter for each frame. It feeds these into the deep learning algorithm and returns a higher resolution image. The game engine will also need to change the jitter of the lower resolution render each frame and use high resolution textures. Finally, the game’s post processing effects, like depth of field and motion blur, will need to be scaled up to run on the higher resolution output from DLSS. These changes are relatively small, especially for a game already using TAA or dynamic resolution. However, they will require work from the developer and cannot be implemented by Nvidia. Furthermore, DLSS2 is an Nvidia specific blackbox and only works on their newest graphics cards, so that could be limit adoption."
D
Deleted member 2197
Guest
It's amazing that dev's can now do this in half a day!
I thought that was for enabling it in unreal?
Probably already had an early preview branch with the DLSS2 feature in our something.
D
Deleted member 2197
Guest
April 14, 2021 - GTC 2021 Video
Last edited by a moderator:
D
Deleted member 2197
Guest
D
Deleted member 2197
Guest
DavidGraham
Veteran
NVIDIA announced the plans to do so back in January.
https://www.nvidia.com/en-us/geforce/news/january-2021-rtx-dlss-game-update/
Less pixels = less rays. DLSS' internal resolution is less, and thus likely affecting lighting/quality, and causing it to lean more on DLSS/Denoiser to construct the image.
DLSS shouldn’t affect the number of rays per pixel. So fewer pixels yes but each pixel should receive the same light.
Most RT effects that I have seen in games use the pre-DLSS upscale res (internal res) to determine ray count.
To be honest, in this case, things seem to be the opposite.
There is RTXGI in this demo, and since it's a GI, it's additive to the lighting, i.e. it adds more lighting to the scene, so brighter image with DLSS might as well be caused by this.
In this case, decoupled probes should be updated at 2x speed with DLSS and per-probes rays budged is usully something like 144 rays per probe per frame.
This would obviously lead to faster lighting convergence with DLSS On and probably a little bit more energy with DLSS On.
Though, unlike RTXGI, which is decoupled from screen resolution and doesn't require screen space denoising, there is also the RTXDI, which is heavily reliant upon sampling in screen space and upon denoising.
Usually, modern denoisers filter out very bright signal to supress the shimmering "fireflies" graphics artifacts, which would be visible otherwise, hence image becomes darker due to the energy loss you mentioned (thought, the "bias" is more common term here since it addresses both the energy loss and other systematic errors in lighting).
Denoisers' blurring can also introduce some additional bias and this bias is typically higher in lower res, so reconstructing signal from lower res can add up to the energy loss.
Unfortunately, we don't have an unbiased path traicing reference images to compare against, so it's impossible to tell which image DLSS On or Off has more bias here, but I would not be worry about that anyway since there are literally thousands of other things which introduce discrepancies with the unbiased path tracing rendering.
Last edited:
https://imgsli.com/NTIxMzM
After a few seconds of waiting for lighting to converge, as you can see, no more difference in GI now (DLSS takes way less time to converge as expected due to 3x higher framerate), just a subtle difference in shadows.
After a few seconds of waiting for lighting to converge, as you can see, no more difference in GI now (DLSS takes way less time to converge as expected due to 3x higher framerate), just a subtle difference in shadows.
D
Deleted member 2197
Guest
April 19, 2021
It would be a huge surprise if ray traced effects didn't scale with the native resolution because I don't see how the performance gains would be anywhere near as convincing with DLSS enabled if games were tracing just as much rays on both the original and internal resolution for DLSS ...
It is true that we don't have reference path traced results but I would at least think that the "unfiltered result" (DLSS disabled) would be closer to the ground truth because filtering can introduce bias like you raised. I think the ethos here would be is that the less filters, the better ? In this case, it's the DLSS image that's noticeably dimmer compared to the original image in the results above since we're missing patches of high intensity lights on the couch and the and on the blanket if we take a closer look.
Rootax
Veteran
