Info Zone - gEngines, Ray Tracing, AI, DLSS, DLAA, DLDSR, TSR, FSR, XeSS, DirectSR and MODS etc.

Thanks Onna.

 

NP, finally i've snaged it and quickly uploaded into my Domain

Demo looking cool, so much in perspective, also take a look into Lumen Lightning Demo.
But do not move and click mouse to much (direction of the sun).
-> https://drive.google.com/file/d/1yWDoiLBHvCvyqbN9ruSURFV4RwX8f9vf/view

 

That's the review I want, I don't play with 600x zoom. Will try with gow

 

OP updated:
How to Use Nvidia's DLDSR to Improve Image Quality on Older Games

-> https://www.techspot.com/guides/2428-dldsr-vs-dlss/

DLDSR is a method of supersampling aiming to improve over the older DSR (Dynamic Super Resolution).
It renders supported games at a higher resolution than your monitor's native resolution, then shrinks the image to fit within the monitor's native output. This results in superior image quality with less flickering and better anti-aliasing.

 

Gof of War DLSS vs FSR2.0

For still images FSR2.0 is probably indistinguishable from DLSS, in motion DLSS is much cleaner since FSR2.0 suffer from some edge distortions (like the antenna in Deathloop)

 

GoW FSR2 & DLSS

-> https://imgsli.com/MTEwMzg5/2/1
-> https://imgsli.com/MTEwMzc0/5/1

 

The main issue is that FSR 2.0 completely breaks with transparent fog effects. If this will be true also in future games, it will be far less useful than initially thought. GoWs own TAAU doesn't have this issue and is faster...

 

Sackboy: A Big Adventure PC version details : GamingLeaksAndRumours (reddit.com)
June 2, 2022

 

Example?

 

Another:
https://forums.guru3d.com/threads/fsr-thread.440398/page-20#post-6022150

Look how with DLSS usually each blade of grass can be identified, but with FSR 2.0 it turns into clots behind fog. You could argue that FSR 2.0 looks more detailed in areas without such fog, but that makes the issue even more striking. And it's not just clots, it's also much noisier. You can already see that in the game's main menu.

I disabled DoF via .exe patcher, but I already verified that this doesn't affect such issues.

FSR 2.0 is badly flawed until this is fixed. Imagine fog appearing in RDR2 and AA to completely break with hypothetical FSR 2.0 implementation...
It btw. was exactly the same issue with smoke effects in Deathloop.

 

Wont be fixed until explicit hardware acceleration is added.

 

Thanks aufkrawall2. Nice catch!

It has nothing to do with hardware acceleration. Stuff like this has to do with how the algo is doing things like clamping, detection etc. And if the error doesn't happen with regular TAA, then it can definitely be fixed in FSR 2. Hardware or no hardware.

 

Lol no sorry. This wont be fixed until AMD incorporates ML hardware on their cards. Nice try tho... Think of the difference in processing power between MSAA and SSAA. Not very hard to deduce, easy to deny.

 

you're better removed from fsr/dlss discussion for the amount of misinformation you spread.

and it wasn't there in deathloop

 

The particle ghosting issue wasn't, but the the issue with transparent effects, such as smoke, was already there in Deathloop. This kind of effect was just barely used in that game.

 

This doesn't require ML to fix. It requires refinement of how it's processing the image. None of that requires ML, it requires further changes to how it's processing parts of the data. That may or may not require more power at all. In a process like DLSS you would try and focus the AI on this one element till it looks right, and then use the resulting pre-trained AI. With hand written code you would look at how it's dealing with the reconstruction around transparencies, like for example how is the vector or jittered data from the fog is being considered compared to the data from grass, vines etc, and refine the settings till it looks right.

 

DLSS2 uses a generic model. The only AI is the generated algorithm, and its becoming very apparent that "Tensor Cores" are required to run it efficiently.

 

didn't see you mention that as a disadvantage

anyway,I don't think this related to fp8/int8 acceleration,which Timmy is fixated on

or is it ? now that I think of it. not that it can't be done, but what about the performance hit ? dlss is already 5% faster than fsr. maybe they're just balancing the performance and IQ a little to stay close to it.

 

Well, here are some issues with FSR 2.0 vs. DLSS in Deathloop:






As you can see, it's mostly just small areas of the screen. Unlike in GoW, where often large parts of a rendered scene are affected. Hopefully these are bugs that will be fixed, and not inherent design flaws.

I wouldn't think it's related to shortcomings of the math format, machine learning part of DLSS should use even less precise formats.

 

Fixing issues with DLSS still requires training of that generic model. And you can help fix issues in that model by training on specific issues. And it's algorithm is what is typically called a pre-trained AI model.

But running a temporal up-scaler, doesn't need this method to work well. But the two released methods that don't, FSR 2 and TSR, are still early in there refinement.

It could be a performance optimization, or a maybe even a method to get more performance. Although I think such a method should be tweaked so that things are occluded more so we can't see it.