Alternative AA methods and their comparison with traditional MSAA*

What makes me wonder: is it smart to apply MLAA recursively?

I mean: use it first the standard way, i.e. apply it once.
But then, in theory you could apply the same algorithm again on the de-aliased image?!
Would this make IQ better of worse - I don't know?

If you already have detected much of the edges in the first appplication, you only find a few edges left in the second application - so the computational overhead may not be so large compared to the first application, as you only need few averaging?

If a repeated application of MLAA further improves the IQ, one could apply it until one has the perfect IQ, i.e. no edges are detected anymore!
And each application should be cheaper then the one before...resulting in the perfect IQ!!
 
The Saboteur uses luminance to detect edges...but how does it smooth them? It's not blurring. It's not MLAA. What is it?

I thought both of them use some sort of color blending to get rid of AA ?

Anyway, here's a short and simple MLAA article:
http://www.realtimerendering.com/blog/morphological-antialiasing/

From where else are you taking the 16XAA quote from? I understand that MLAA must offer a similar or better technique that the one used on saboteur when it comes to remove aliasing, but I have not seen a comparison to normal MSAA as in Saboteur.

I found an Intel MLAA paper here: http://visual-computing.intel-research.net/publications/mlaa.pdf

They may be referring to this section ?

4 COMPARISON WITH SUPERSAMPLING

...

We compute pixel coverage analytically (black line on the top chart) and compare it with MLAA (blue) and two supersampling approaches, using 4 and 16 samples per pixel. All techniques estimate area with 15% accuracy, while MLAA provides a linear estimation, without any unwarranted frequencies.

...

On Figure 8, the half-plane silhouette is almost horizontal, straddling 24 pixels per Z-shape. For this view, 2x2 sampling results in significant errors in area estimation. For MLAA, actually the opposite is true since a longer Z-shape allows for smaller variability in the true image. Accordingly, to match MLAA quality, 8x8 sampling is required.

...
 
i think one case where 4xmsaa will destroy mlaa is for far objects especially fences and wires (any very thin items) because the sampling will help from things from just simply disappearing into the distance but you can sorta make art assets in a way that favors mlaa like those fences in that pic you just posted.

Yes but to contrary 4xmsaa with HDR or high contrast not work anywhere and when we works remains well simple 4xmsaa :LOL: however I think 2xmsaa with these limited edges could be the better solutions.
 
What makes me wonder: is it smart to apply MLAA recursively?

I mean: use it first the standard way, i.e. apply it once.
But then, in theory you could apply the same algorithm again on the de-aliased image?!
Would this make IQ better of worse - I don't know?
One could go into a technical discussion about why this wouldn't work, but it's a very easy question to answer with a more low-brow argument. Look at the God of War images. Tell me honestly, does it look like it needs better antialiasing? ;)
 
One could go into a technical discussion about why this wouldn't work
please,it would be great to learn more about this!
We are here in a tech forum after all ;-)

but it's a very easy question to answer with a more low-brow argument. Look at the God of War images. Tell me honestly, does it look like it needs better antialiasing? ;)
But we don't know how other games compare, right? It is not obvious if one MLAA appplication always results in such a perfect IQ in every game...you cannot translate the GOWIII results to every game!
 
Last edited by a moderator:
The MSAA sample comparisons are really not particularly useful, because sample count determines two distinct features of MSAA: subpixel resolution and the number of samples used to construct a pixel, which gives you the number of "steps" in a line (the smoothness, so to speak).
If you compare that to GoWAA (I like that. Sounds like GWAR.), we have no subpixel information (so 1x), but we have more then enough resolution for any slope possible in the image. In a 720p image, that would be 921600xMSAA.
So, please, forget about samples for methods that rely on sampling. Thanks :)

There are objective ways to compare AA methods by using error metrics, but that's not really practical.
 
But we don't know how other games compare, right? It is not obvious if one MLAA appplication always results in such a perfect IQ in every game...you cannot translate the GOWIII results to every game!

Yeah, use of DOF for background aswell as color contrast does a lot to impact how easily jaggies are percieved.
 
But we don't know how other games compare, right? It is not obvious if one MLAA appplication always results in such a perfect IQ in every game...you cannot translate the GOWIII results to every game!
Well, broadly speaking, yes, you can, just as you can say 4xMSAA will produce a given, predictable degree of alias reduction. The edges that are identified and interpolated with GWAA's routines will have this same high quality to them, whether in a racing game, platformer, or anything else. We can look at LBP2 screenshots for another implementation, and they're looking just as clean.
Yeah, use of DOF for background aswell as color contrast does a lot to impact how easily jaggies are percieved.
DOF in background would more likely highlight jaggies, lowering the frequencies behind the jaggied edge. And again with the colour contrast, look at example images with (near-) white edges contrasted against darker backgrounds. GOW3's lack of aliasing isn't an optical illusion that breaks in more demanding situations, a lack of stepping until a metallic highlight reveals the pixelated nature of the framebuffer.
 
The MSAA sample comparisons are really not particularly useful, because sample count determines two distinct features of MSAA: subpixel resolution and the number of samples used to construct a pixel, which gives you the number of "steps" in a line (the smoothness, so to speak).
If you compare that to GoWAA (I like that. Sounds like GWAR.), we have no subpixel information (so 1x), but we have more then enough resolution for any slope possible in the image. In a 720p image, that would be 921600xMSAA.
So, please, forget about samples for methods that rely on sampling. Thanks :)

There are objective ways to compare AA methods by using error metrics, but that's not really practical.

Ah ha, I was hoping the caged animals will venture outside their habitats again. I noticed that the MLAA paper says their implementation is massively parallelizable but yours is not. Are you using a more "accurate" version of edge detection ?

The "SUMMARY AND FUTURE WORK" section hints at possible future development. I suspect your framework is more general than their implementation (because SPU cores is more flexible). What other areas are you looking to reapply the framework, if you can reveal the info ?
 
please,it would be great to learn more about this!
We are here in a tech forum after all ;-)
MLAA first searches the image for a edges which are clearly defined and have visible aliasing and then fits antialiased lines on them.

If you would run MLAA several times you wouldn't really improve the image quality as the first MLAA pass already did get rid of any visible edges.
If secound time finds any they are most likely false edges and just grow additional artifacts.
 
I thought both of them use some sort of color blending to get rid of AA ?

Christer Ericson said it was just edge blur. But the folks here said it was inconsistent with the results blur usually produces. I don't think it was ever discussed further..
 
MLAA first searches the image for a edges which are clearly defined and have visible aliasing and then fits antialiased lines on them.

If you would run MLAA several times you wouldn't really improve the image quality as the first MLAA pass already did get rid of any visible edges.
If secound time finds any they are most likely false edges and just grow additional artifacts.

Single Processor case:
I understand this, if the image is pure black and white!
But if one considers a colored pic, I am not so sure if the definition of a edge is unique!!

Multi Processor case:
if one considers parallel version, I think that the edge detection alters due to the domain decomposition. In this case, the edge detection could depend on the domain decomposition...thus you could use MLAA recursively on different domain decompositions to improve IQ...
 
Well, broadly speaking, yes, you can, just as you can say 4xMSAA will produce a given, predictable degree of alias reduction. The edges that are identified and interpolated with GWAA's routines will have this same high quality to them, whether in a racing game, platformer, or anything else. We can look at LBP2 screenshots for another implementation, and they're looking just as clean.
DOF in background would more likely highlight jaggies, lowering the frequencies behind the jaggied edge. And again with the colour contrast, look at example images with (near-) white edges contrasted against darker backgrounds. GOW3's lack of aliasing isn't an optical illusion that breaks in more demanding situations, a lack of stepping until a metallic highlight reveals the pixelated nature of the framebuffer.

But when looking at two different games with e.g. 2xMSAA...the amount of jaggies you percieve drastically depends on the game?! Heck, sometimes even on the actual setting you are currently playing! Compare for instance Mirrors edge (Xbox360) and the GOW3 demo: both had 2xMSAA iirc...
 
That's because MSAA isn't particularly good. ;) MLAA produces antialiasing akin to massive PR shot supersampling. Antialising akin to quality vector line-drawing routines. Thus every AA'd edge is smoothed to a 'perfect' progression of intensities, which you cannot improve upon with our discrete display technologies (and given the resolving limits of our eyes, do not need to imrpove upon). MLAA, or rather GWAA specifically as there'll be different implementations of MLAA, will no more break down into jaggies on edges than digital photography will.

Whereas MSAA will have a limited AA effect when it is applied, very noticeable on very shallow edges, which varies in AA ability given triangle alignments because of the sampling patterns, and can get broken with HDR, resulting in cases of no intermediary steps at all and discrete stepping (2xMSAA, 4xMSAA will have at least some intermediaries).

Variations from game to game will happen not because of framebuffer content, but because of implementation, either missing some edges (and worst-case edges whould be the ones most detected, so visibility of non-AA'd edges will be minimal), or lots of tiny-object noise, or somesuch. Santa Monica described their first implentation only as good as 2xMSAA, which shows MLAA can be created in different flavours. With a common library of code though, games adopting GWAA specifically willl have the same IQ as GOW3, whether they are racers, platformers, shooters, horror-games, puzzlers, or any other genre of high or low contrast scenes.
 
And what about fast moving scenes? How does MLAA compare to MSAA?
In GOW3, you don't have so much fast moving cameras (in LBP as well)...compare this to a FPS/TPS game, where you can just stand still and fast turn around, i.e. fast moving cameras!
So, you cannot translate GOWIII results automatically to fast moving camera games (for the record: I *don't* say that GOWIII has fixed cameras ;)).
I remember repi posting about tests with BF:BC2+MLAA, he mentioned that you get some artefacts when moving around...
 
And what about fast moving scenes? How does MLAA compare to MSAA?
In GOW3, you don't have so much fast moving cameras (in LBP as well)...compare this to a FPS/TPS game, where you can just stand still and fast turn around, i.e. fast moving cameras!
So, you cannot translate GOWIII results automatically to fast moving camera games (for the record: I *don't* say that GOWIII has fixed cameras ;)).
I remember repi posting about tests with BF:BC2+MLAA, he mentioned that you get some artefacts when moving around...
Artifacts are easier to see when the camera moves around slowly, not fast.

Basically you lose sub-pixel accuracy on the edges and that causes some edges appearing and dissapearing.
It also looks worse than MSAA on every small or thin edge, this is very apparent while moving.
 
And what about fast moving scenes? How does MLAA compare to MSAA?
In GOW3, you don't have so much fast moving cameras (in LBP as well)...compare this to a FPS/TPS game, where you can just stand still and fast turn around, i.e. fast moving cameras!
So, you cannot translate GOWIII results automatically to fast moving camera games (for the record: I *don't* say that GOWIII has fixed cameras ;)).
I remember repi posting about tests with BF:BC2+MLAA, he mentioned that you get some artefacts when moving around...
Are you kidding? Try to grab an enemy with the circle button and see the change of the camera that zoom to you (in and out) that move left to right etc. The MLAA implementation of GoW 3 is simply stunning, even if we consider that is implemented in a game that have a framerate that fluctuate between 30 and 60fps.
 
Are you kidding? Try to grab an enemy with the circle button and see the change of the camera that zoom to you (in and out) that move left to right etc. The MLAA implementation of GoW 3 is simply stunning, even if we consider that is implemented in a game that have a framerate that fluctuate between 30 and 60fps.

Basically, I am always kidding;)

When you grap an enemy, the camera zooms in fast, right!
But at the same time lot's of cool post processing kicks in, focusing on Kratos and the foe, making jaggies hard to detect!

For the record: I don't (I would never!) downplay GOWIII MLAA implementation in any form (I hail T.B and the rest of the MLAA ninja gang!!)- after all, it is my favorite game (not only graphics, but gameplay too!)

I just want to find out, if MLAA can be applied to every other game out there! We can either wait, and see if all games on PS3 in future have MLAA or try now to come up with situations, where MLAA may not be so good...
 
Artifacts are easier to see when the camera moves around slowly, not fast.

Basically you lose sub-pixel accuracy on the edges and that causes some edges appearing and dissapearing.
It also looks worse than MSAA on every small or thin edge, this is very apparent while moving.

Thanks for explaining!
Nebula mentioned in another thread that you should not combine SSAA+MSAA due to a massive performance drop!
What about combination of SSAA+MLAA? If MLAA is cheaper compared to MSAA, you could combine this with SSAA to fight small or thin edge aliasing?
 
Back
Top