Nvidia DLSS antialiasing discussion *spawn*

why are you providing me with a link to dsr ?
dsr is downscaling aka rendering the game internally at 4k and displaying at a lower res pjbliverpool was talking about upscaling which is the opposite aka rendering the game internally at at a lower res and displaying it at a higher res (dlss 2.0)

 

With DLSS you select your output resolution (say 4k), then you activate DLSS and set the quality level.

So if you select "Quality" for example it renders internally at 66% on each axis(1440p in the 4k example), and if you select performance it'll renders at 50% (1080p at 4k). The new ultra performance mode renders at 33% per axis, so 720p at 4k.

However I've since read that ultra performance mode may be locked to 8k output resolution only. But that's still only 1440p internally.

Of course in order to select 8k as an output resolution you either need an 8k display, which basically no-one has, or you need to trick your system into thinking you have an 8k display via DSR. But at present that's only possible if you have a 4k display because DSR only supports up to a factor of 2 on each axis. My paltry 1080p display would need it to support a factor of 4 to enable 8k output resolutions on my system.

 

The resolution is impressive but look at the geometry around the wheel well. What's the point of 8K when the models are still low res?

 

Sure, making notable progress along one axis of photorealism will naturally expose the inadequacies of the other axes. But perhaps we should enjoy the moment for a short while, especially the fact that the stride was made without exhausting brute-force compute resources.

 

Well it did need a supercomputer

 

Even though I have a Turing card, I haven't played much with DLSS. Is there any hint of blurring in movement in 2.0 and 2.1 games, compared to native?

 

Well, AFAIU the ML component of DLSS 2.0 is doing exactly that - removes the blurring which is incurred by using temporal accumulation. So basically you could call DLSS 2.0 "DLTAAU".

 

No blurring, but breaking up of fine detail. DLSS is good at drawing complete lines when the camera is still, but in motion those lines will often break into pieces. A bright line with a dark background losing it's form can be quite noticeable - with less contrast obviously less noticeable.

 

That's news to me.

Edit: @Randomoneh - a good start would be to look at DF's Death Stranding review posted earlier in this thread.

 

Live and learn I guess.

This can be seen quite easily in Deliver us the Moon that has stuff like white antennae lit by sunlight against a dark space background in it. An antenna will be nice and full when still, but strafe the camera left and right and the delicate lines will get cracky. Not quite aliased but like I said, they lose their form in a noticeable manner.

 

Really not noticeable in any Deliver us the Moon reviews I've looked at. Another week will tell the "tale of the tape".

 

I wouldn't dismiss it. a) He sounds convincing. b) We need to give them all the crap that a certified floor or antenna inspector would give them. That's how they improve

 

This sounds passively aggressive and needlessly dismissive.

 

Then why bother asking your question initially? Too lazy to read DLSS 2.0 reviews or just looking to "stir the pot"?

 

DLSS 2.0 having a temporal component means that there is some breakup of reconstruction on camera movements and this is apparent on high contrast areas.
Control is a good example here - DLSS'ed image is almost perfect in static or with forward/back movements but the edges visibly break up on camera pans.
This however is similarly minor thing as some ghosting on objects with wrong motion vectors and pales in comparison to performance gains.
Also TAA is having similar issues and also ghosting which is just as visible in these cases.

 

That's related to motion blur being processed before DLSS (probably MB doesn't support viewport jittering), disable it and the issue with camera pans will disappear.
Ultimately, post-processing must come after DLSS. In a perfect world, VRS should be applied on top of full res post-processing to improve perf on top of DLSS with minimal quality losses on edges compared to full native res rendering.

 

ULMB users are looking for zero-smear image, TAA-lite ghosting won't make it. I'll test it out in fast movement scenarios.

 

Condescension is towards user dorf who shared his experience. I see no need for that.

 

I think this must be true. What's also important to note is that it's the faint detail where the breakup/loss of quality occurs (most visibly at least) when the camera moves. The detail TAA and NoAA couldn't even draw completely if at all.

The small red lights in these images are an example of the kind of detail I mean:


(DLSS is 50% resolution, NoAA and TAA are 100%)

DLSS manages to draw those red lights much more accurately from far smaller input resolution which is quite a feat - without AA some and with TAA most of them are missing.

DLSS can only do this because it very efficiently uses the multiple temporal samples given to it. Temporal accumulation is most effective when subsequent frames are alike which is the case in my example, the camera is still and the game is paused/stopped. Start moving the camera left and right and the frames become less alike and using them as effectively becomes impossible. As a result these red lights will start to break up/turn off here and there. Of this I have not and can not provide footage.

If the gallery doesnt work, add dot:
imgur com/a/4NvEjEh

 

It would be really interesting if someone did dlss tests with different framerates using same gpu+settings. In theory the quality should be better in higher framerates as there is less difference between frames.