they had a screen shots comparison for GOW3 demo and final game, the final game with MLAA doesn't looks any blur compare to the demo with 2XMSAA. Only way cleaner. The AMD MLAA varies on different games, some games looks a bit blurred but some don't. LBP2 also looks as sharp as always did on LBP1, if not sharper.
*merged/spinoff* for the Neverending Killzone Discussion on Graphics
- Thread starter KKRT
- Start date
Surely it's just voodoo?!?!?!
Lol, well I genuinely want to know! Take this image, ripped from an mlaa paper:
http://www.realtimerendering.com/blog/wp-content/uploads/2009/07/girlzoom_noaa.png
Everywhere black & white meet mlaa may try to smooth it. If those are polygon edges, or edges that overlap background then fine, smooth it. But what if that entire image was a texture, like a symbol on a dudes shield or whatever, in that case I want mlaa to leave it alone. So how do current mlaa implemtations know to leave those pixels alone?
MSAA is fine with HDR as long as it is implemented correctly. For example if you do it wrong and take 4 samples , 3 which land on a crate in shadow, but the 4th lands on the skybox whose brightness is 10 times that of the crate your pixel will now take on a level of brightness that is far higher than it should be. To do it right you must tonemap each subpixel individually and then do the resolve. Not a great explanation but look here for an example.
http://www.humus.name/index.php?page=3D&ID=77
Lucid_Dreamer
Veteran
We already know that Santa Monica Studios said their MLAA implementation goes "beyond" Intel's (?) original MLAA paper. We just don't know in what way(s), yet.
I wish I could.
(Yes, slightly dickish move, calling you out and then not explaining. Sorry about that. But my hands, as they say, are tied.)
Digital Foundry also compared LBP 1 [2x MSAA] vs LBP 2 [MLAA].
http://www.eurogamer.net/articles/digitalfoundry-lbp2-engine-face-off
Its really interesting, because LBP 1 actually looks blurry, i thought it used QAA first.
http://www.eurogamer.net/articles/digitalfoundry-lbp2-engine-face-off
Its really interesting, because LBP 1 actually looks blurry, i thought it used QAA first.
Of course it doesn't flicker like that. Everybody knows that medium bitrate video compression introduces edge flickering. It's very common. No, really. Very common.
So can a 3-pixel gaussian blur. Doesn't mean it's the preferrable method.
4xMSAA can also create a better image than MLAA under certain circumstances. Doesn't mean it's the preferable method too.
The point was:
According to the MLAA paper, MLAA can perform as well as (or better than) 16xMSAA under the right conditions.
Developers and artists can skew their work towards these better performing cases.
They should also continue to research in this area to minimize the MLAA breakdown cases.
As some one who games on a stupidly fast gaming PC with 3 ATI 5850 graphics cards even I'm impressed with those graphics
And I also have the ability to run every game with MLAA and I think it's horrible, Blurs the images way too much. Well on PC at least.
And on the previous page someone has posted shots of Dead Space claiming 4xMSAA, The game has NEVER supported any multi-sample anti-aliasing on PC, Forcing it through drivers doesn't help either.
But we talking here of MLAA ps3 solution or AMD? Because on the ps3 I haven't noticed the blur which you talking about
. Did you have play GOW 3 or KZ3? Because it's completely different in sharpness compared to the AMD solution.I searched for days looking for a AA fix for dead space for ATI and Nvidia hardware and came up with nothing.
Even forcing super sample aa doesn't work.
On topic the best solution is MSAA+MLAA as I've found from my own testing. There's certain situations were MLAA is useless and that's were running MSAA along side it comes in handy.
The IQ from running both types is amazing
Lol that's ok, at least you've confirmed that mlaa does not affect textures, unlike other implementations of it like Ati's which I wouldn't touch with a ten foot pole.
But now I'm left wondering what is making mlaa games look softer compared to games that use msaa or no aa? The last mlaa games in my head are LBP2 and KZ3. KZ3 isn't bad, but it still a touch soft compared to say Castlevania or Uncharted 2. LBP2 definitely has a noticeable softness to it that's easier to spot compared to KZ3. I don't remember Gow3 offhand, I'll have to go fire it up later.
Can you answer if you guys are using the depth buffer to determine edges? I guess you prolly can't, but it's worth a shot.
Yeah I'd like to know as well. Using depth buffer helps, but even that is not fool proof. Maybe they are writing out where polygon edges are in the stencil channel perhaps. In other words, imagine rendering the scene as wireframe where polygon edges are written out as 1 in the stencil buffer, and everywhere else in the polygon is written out as 0 in the stencil buffer. Maybe do it simultaneously during a depth prepass or some other prepass step. Once complete you would have a stencil buffer that is basically a map of edge locations. That buffer gets dma'd to spu along with the color buffer so that it now knows what to smooth. Doing something like that would leave true detail mostly alone, and only muck with edges.
Those images actually do show softening in the mlaa version, in a way errily close to what I was mentioning with the sample image I had linked. Look at Kratos's chest, left side (his left), above his nipple area, basically above the big splotch of red and more to where the fine red lines are. Those are sharp in the demo, and blurred in the final version. Also, look at the head that he's holding near the neck area. A lot of the pockmark detail is there in the demo version but gone in the final version. That's the types of details that I'm talking about which can get softened up by mlaa, any high contrast details. Intestingly a depth only solution would not work on those particular areas in terms of trying to determine if it's an edge or not because the Z values would all be too close together to be able to know.
Last edited by a moderator:
Similar threads
- Locked
- Locked