brain_stew said:
That does nothing to deal with sub pixel aliasing though, so thin lines and distant objects get pretty much no edge smoothing at all.
Well, it's a bit more complicated than that. They can be blended/smoothed, but the result won't be as accurate in these cases as with techniques that use subpixel info, and can lead to strange patterns. Though you could selectively choose to ignore those objects if the results were poor. See this comparison using Intel's algorithm:
http://i50.tinypic.com/2helvuw.png
Note just how thin your object needs to get before you see (notable) deviation from the 16-sample SSAA. (And perhaps, consider if that really wouldn't be worth the tradeoff vs no AA or 2xMSAA for example).
brain_stew said:
Its better than 2xmsaa but its hard to make a case for it being better than (properly applied) 4xmsaa in most cases.
Is it?
Intel did some technical tests with a number of edge slope types, and the results compared favourably to SSAA with very large numbers of samples (e.g. 16 or 24), not to even mention MSAA.
From the one game I've seen use it, it certainly looked well beyond anything I'd seen before on a console (i.e. 2x or 4x MSAA) in terms of edge quality. To me it looked like someone took a PS3 game, ran it on a PC emulator and cranked up the AA high. It was frankly a little weird to see it running on a console.
brain_stew said:
Even with 4xmsaa or MLAA, you're still going to get significant alpha and shader aliasing so you're still far from a aliasing free image.
MLAA will actually cover any edges on the image - not just polygonal edges, but those created by texturing (with transparency or no) or shader effects. It's one advantage over MSAA. MLAA doesn't care how an edge was produced, unlike MSAA.
A lot of people are saying strange things about MLAA that seem maybe half based on truth, and half based on total non-truths. I'd again recommend a read-through of Intel's paper on it (while bearing in mind that Sony's implementation has made further improvements over it, apparently wrt motion artefacts). I'd also recommend a realisation that neither MSAA or SSAA even are perfect either (Intel's paper, again, does a good job on comparing relative strengths and weaknesses of MLAA vs SSAA). People refer them almost as if they're reference/master quality, but they're not really...all of these techniques have deviations vs the ideal, perfect edge, with their own individual characteristics and contraints. It seems a little unfair to highlight all of MLAA's shortcomings (imagined or otherwise) while blissfully ignoring those of other techniques.
Whilst this is an interesting discussion, we risk veering waaay OT. But 'the talk' around MLAA seems rather confused.