Alternative AA methods and their comparison with traditional MSAA*

Depending on how much developers will spend on ever higher definition graphics you may get away with supersampling on newer cards. Wouldn't that be simpler? It looks better to me even at low resolutions, but maybe I'm biased by seeing CGI cutscenes in old games.

Yes, I would approve 16xSSAA too. :)
 
Depending on how much developers will spend on ever higher definition graphics you may get away with supersampling on newer cards. Wouldn't that be simpler? It looks better to me even at low resolutions, but maybe I'm biased by seeing CGI cutscenes in old games.
I'm quite sure that only selective supersampling can be fast enough for new games and even then the shader alternatives would be better in terms of speed/quality.
Yes, I would approve 16xSSAA too. :)
If the game is old enough you can force up to 32xSSAA on nvidia cards and get 'playable' framerate.. ;)
 
Yeah even 4x MSAA looks significantly better to my eyes still... but I'd definitely like to see it on a real 3D scene.

4x MSAA isnt better, i checked it, maybe when You use really low angle on whole fence field, but still on some edges is much worse. 16xQCSAA is better on subpixel features on low angles, but its worse upclose and with shading option on. 8x CSAA is comparable except for one cases, a link of two fences links, there is some aliasing that its not solved by SMAA.

I have quite low res monitor [1280x1024], i'm interested how it's solved on 1080p or higher res. It would be nice to also have supersample option like in FXAA 4.
 
I'm looking at ability to integrate over regions, not edge gradients (which are easy, and uninteresting) as I've mentioned about 100 times in this thread. Tilt the grid and jiggle it back and forth and it's very clear that even 4x MSAA is superior temporally, which is what matters.

The goal of these things anti-aliasing algorithms is not primarily to make static images look good, but rather to limit the amount of temporal flickering, shimmering and flashing that you get. To that end I'm glad they've posted a demo rather than static pictures, but I would love to see some real scenes in there to see how much the theoretical grid temporal stability results impact the results in practice. I'm also glad that people are starting to (re)incorporate non-uniform sub-sampling patterns into these reconstruction filters, but I'd really like to see results that are strictly better than MSAA (i.e. *worst case* they revert to the MSAA results at similar sample counts).
 
Last edited by a moderator:
well at any rate t2x looks way better than smaa 1x, while having similar cost in frame time.

4x and s2x look significantly better, but take twice as long at least on my 4890. i think it honestly looks better than 4x msaa for aliasing on diagonals, but it has much worse moire if you zoom all the way out.
 
Yeah I'm not sure how much to trust the performance on a contrived scene like that. All of them seem far too slow really so I imagine it's just bottle-necked on rendering ridiculous geometry mostly. Furthermore without considering a realistic cost of shading, you can't really make a comparison at all, so I'd take any performance numbers in the demo with a huge grain of salt.

Also the "temporal AA" checkbox doesn't seem to do anything. in the cases that I've tried... I like to disable any temporal AA since it just confuses the comparison. Show me the version with it pre-integrated, and if it's too slow then we can fall back on temporal super-sampling. Fundamentally it really is just a somewhat less robust way to do super-sampling, and thus uninteresting when comparing raw quality.
 
Last edited by a moderator:
Please take into account that what we are showing in our demo is a worst case scenario. A very high contrast grid which can be moved very very fast.

The performance numbers obviously cannot be taken into account, for that we already measured in real scenes and gave numbers in the paper ;-)
 
Turn the shading on and it becomes a real pixel shimmering hell. :D

On a side note, I'm surprised the CSAA 16x mode is giving worse output than 4xMSAA, especially at the fine details. I thought the extra gobs of coverage samples should provide more balanced weights for the pixel resolving. The same could be concluded for CSAA 16xQ vs. MSAA 8xQ to some extent.
 
I think it would be nice if there's an option to load a background image in the demo for more fair IQ evaluation -- or just a uniform colour picker. All black is indeed too extreme.
 
Please take into account that what we are showing in our demo is a worst case scenario. A very high contrast grid which can be moved very very fast.
Oh I know, but that's precisely how we should be comparing these techniques of course :) In cases where they all look similar, it's obviously harder to spot where the real differences lie. That said, I would love to see it in a more representative scene as well.
 
say you built a cheap brazos system with a video capture card, would that be able to run fxaa 4 or smaa tx2 at a reasonable framerate, just processing a video feed from a dedicated gaming computer ofc.

maybe in the future amd or nvidia, intel, maybe even lucid would implement it as an extension of virtu. offloading the post aa to a separate videocard, or integrated gpu.

just seems like a good feature to add, if theyre already routing the video through the integrated graphics. might as well spruce it up a bit.
 
Hello everyone,
i would like to inform You that we decided to implement SMAA 2.7 into our titles
(alongside already existing FXAA 3.11)

atm. it's available in public betas of

ARMA 2: Operation Arrowhead
Take On Helicopters

and it is expected to be released in next full updates of the games (both titles are DX9)

ofcourse we are looking into improved SMAA for our DX11 title ARMA 3

readme with legal and credits given and thanks are included with the betas and will be part of patch

next to thanks to authors and theirs partners
i must thanks everyone in this thread for information, insight and feedback cause that's what helped me in discussion within our company about them
 
Unofficial TXAA Info
Technically I should be able to say something about this now. There will be more official information at some point on the NVIDIA blog.

Prior I had posted some images of the early R&D work towards the next version of FXAA, then went quite as I shifted my focus to finishing up TXAA for Kepler, something I personally feel is a giant leap forward. The core idea of TXAA is to do correct hardware anti-aliasing as close as possible to what is done in the CG film industry and to target aliasing in motion directly.

TXAA is based on hardware multi-sampling, a high quality sample to pixel filter, and temporal super-sampling (which is optional but provides a 2x quality improvement). TXAA supports forward and deferred rendering pipelines.

TXAA is designed for engines which want to extend physically correct linear HDR lighting throughout their engine pipeline. With TXAA, hardware anti-aliasing has perceptually smooth gradients in the cases where traditional MSAA resolve does not. With TXAA, hardware anti-aliasing still works with high-dynamic range input and with the correct color bleeding from the over-exposed areas just like one gets when taking a real photo using a camera.

For many of you, the first response will be to judge TXAA compared to other AA or no-AA filters using still images, and then remark how TXAA does look less aliased, but also looks less sharp. This is the correct response too, because it is physically impossible to remove aliasing, especially temporal aliasing, without resulting in a perceptual reduction of sharpness. Motion however is where the real battle for AA is fought, and where TXAA really starts to shine compared to all prior methods.

For those who have a problem with lack of sharpness compared to no-AA GPU rendering, and judge AA techniques by still image screen shots, I'd suggest pressing the pause button on your Blue-Ray player and take a look at one of the CG film shots in your favorite movie, and then comment on sharpness. If enough of you light up the forums I might be able to do a side project at NVIDIA where we ship a video filter which makes movies look like traditional GPU rendering in games, something so sharp that your eyes will bleed.


http://www.timothylottes.blogspot.pt/2012/03/unofficial-txaa-info.html
 
From neogaf, dunno where guy got it.

RPsNO.jpg


hTKHx.gif


1332370311kfvezRMQlH_1_17_l.gif


There is some info on hardcocp
 
Back
Top