Been watching some videos of the leaked version. And it's honestly not bad. In fairly recent games at least it looks pretty good. Horizon, GTA 5, Forza, TLOU 2, Cyberpunk all look good with it. Didn't really see much of a difference in a cartoonish Cel-shaded game like Borderlands 4 but old games like GTA San Andreas look very cursed with the low poly character models and dated animations. probably best not to use it with those.

It seems like it's best used in newer games that are already going for a realistic look.

This is on a 5090.... and it has jotter and lag.... you notice it with the electric poles on the side of the road as he drives by.

Supposedly dlss5 cost more than half the frame rates, you normally have, to run it.
If a 5090 cannot run Forza Horizon, at high frame rates with this on, yeah, Its of no use to me.

IF you go to 1:53 (again at 3:24) in the video and watch him turn "intensity" of dlss5 to 0% and back up to 100% (I assume that is on/off)...
the visual difference really is not that big.  It changes the lighting of the scene and that is about it.

The best use and worse of it too, is still faces of characters imo.
As you can see towards the end of the video.
Not that that stuff matters much of anything in a racing game.

The changes to characters arn't always for the better though.... and the demand on the gpu, is not worth it imo.

Much better video, than the Forza one.
The one you linked (i'll embed for you):




" Imagine this kind of graphics or better for Switch 3 🤤"

DLSS5 is pretty demanding.... The "5090" can use ~575watts of power.
The Switch 2 current maxes out at like ~28watts when docked.

DLSS5 can make a 5090 "slow" at high resolutions in games.
Just like how Path-tracing is stupidly demanding.... this neurorendering is demanding on the gpu.

Imagine a lesser version of this on a Switch 2, but instead of games running at 30fps, they instead run at 10fps.
Handhelds are not were this tech will shine.

Yeah right... don't dream too much tho. I'm not even sure I could run 60fps games on my €2500 PC with DLSS 5 so don't get your hope too high.

Thanks mate for sharing the video. But the truth is this is like RT technology when it was first launched. Another problem is that current leaked version is using unoptimized FP8 neural network. The official one should use FP4 one, which should double performance in some cases. Still it is a baby step towards photorealism. In the future like in 2033 they should have a more optimized lighter model that could be run on low power hardware like Switch 3. If the performance hit is like 12%, then upscaling will even out the performance loss.

You're right I am overly excited about the future. But the performance penalty won't be this high on 50 series when this launches. The current leaked version uses FP8 instead of FP4 which a Blackwell GPU can compute or process at lower precision, but at a much more optimized speed and accuracy. So the performance penalty should be up to 30% based on the performance of modded version.

DLSS also uses AI-generated frames.  It inserts those frames between genuinely rendered frames.  That is called DLSS AI Frame Generation and has been around since DLSS 3.

Further, the AI in DLSS Super Resolution (the process you described) independently fills in “missing details” as part of its up scaling process.

Both of these censor the artist’s work in real-time by imposing changes they never intended.  It also means you are playing a simulated version of the game instead of the actual game if the AI generates enough frames and alters them in the upscaling process.

DLSS was made to be the AI upscaling tech, but Nvidia kept adding new techniques under the name. But does upscaling really change the developers intent or it is doing something that is not truly correct on the objective level? So basically, you're disliking AI frame gen and neural rendering. Prove to me that AI upscaling is changing the intent with DLSS 4.5 version set on quality or DLAA mode.

Are you demanding that I prove to you that AI fills in the empty spaces as part of its upscaling process?

Or are you demanding that I prove such a process introduces deviations from the artists’ intent?

I am asking for a proof that DLSS upscaling changes the intent of the artist or developer. It is either causing artifacts or missing details. It very rarely adds things or modifies the original image on artistic level. The intent is to upscale the image and make it look like native, but with anti aliasing. If there are hallucinations and artifacts, those are imperfections and shortcomings of the models not the change in artistic intent. That's how I would describe it.

Now you claim that rarely adds things that modifies the original image on an artistic level, but it is a fact that an AI fills in all the “missing spaces” that were created by the upscaling process.  This is, by its very nature, adding things that modifies the original image.  The developers certainly did not fill those “missing spaces”, an AI did.  

Do you accept this much to accurately describe what happens during DLSS upscaling?