FXAA On PS3....Why use this over MLAA?

I was thinking more stencil buffer/MRT and a custom attribute.

 

Yes, I was wondering if we can use FXAA on MLAA selectively, only for certain subpixel features.

These post processing AA... They can address shader aliasing right ?

 

Graphics dev from 343 got his hands on FXAA3 too.

Meanwhile Timothy posted update how can you get sharper results on console,but with less sub pixel results.

http://timothylottes.blogspot.com/2011/06/sharper-fxaa3-console.html

 

Nice!! I was really hoping that 343i would get their hands on FXAA.

 

hope this is will be all the PS3 and 360 SDK so everyone can take a look.

 

It basically doesn't blur well done textures at all, since textures are already mipmapped (antialiased). If it touches a texture surface there must be a big luminance difference, and the most likely reason for that is specular aliasing.

This works. But stencil masking (or using depth to bounds for NV hardware) provides better performance (less pixels to process).

Yes, post AA filters address shader aliasing and hard edged (clipped) transparencies. So if you have POM/QDM/etc on every surface and lots of (clipped) foliage and leaves, post AA is a pretty good choice. Post AA filters also help a little bit with specular aliasing, but techniques such as lean mapping are solving that issue better, since specular aliasing is caused by really fine subpixel details. Of course nothing beats real supersampling in quality, but post AA comes pretty close in many cases. MSAA has it's fail scenarios as well (transparencies and shader aliasing).

 

It should be easy enough for the devs to do. The Edge and 343 folks have already done some work on it. I'm sure it will be in the SDK nonetheless.

Assuming the quality is good enough, I wonder if they can spare 1ms to get rid of shader aliasing in GT5. ^_^

EDIT: Sebbbi, what threshold values are you using ? The fidelity of the final image will depend on the thresholds right ? Do you have to tweak them based on different scenes at the moment ?

 

shouldn't fxaa be cheaper than 2XMSAA or 4X in GT5. Since Polyphony Digital still constantly working on the game, I hope they will do it. Might even get more frame rate boost.

 

I was wondering if the MLAA you tried was the Jimenez MLAA or something else. It was quoted to run at 1.3ms on the 360, though I'm not sure if it's a lower quality than the MLAA you tried out.

Here's what I think is the tweet (pulled from the Gears 3 thread:

I assume you're talking about MLAA here, right? Thanks for the explanation, always love to read the details on things I'm hazy on, like this.

You think it would be a good idea just to ditch the MSAA entirely and go with FXAA as it's replacement? Not sure if GT5 is still have some slight frame rate issues or not, but if so then this could maybe help.

I've always wondered what the cost of 2x/4xMSAA is on consoles as well.

 

kinda goes along with patsu, some assets which they know are going to be subpixel like these powerlines could they just do what they did with nfs:hp my god i've never seen such magnificent powerlines

http://www.eurogamer.net/articles/digitalfoundry-nfs-hot-pursuit-demo-showdown

 

^Speaking of which...I remember La Noire having just as smooth looking powerlines.

 

The NFS: HP powerlines are analytically anti-aliased thanks to some awesome rendering code Criterion did

 

So would it be possible (in futur, or may be in present?) to have selective AA method in function of needs in the same frame ? Not to complex to manage? To costly in RAM? Or the same Post AA method can be tweak to apply different parameters depend of the needs for AA (ex:Criterion AAA for powerlines, FXAA for sub-pixel)?

And if responses are to "Hot" and can damage your rep, don't answer now… Make a game with the answers.

 

Yes, and I've long expounded as much. Power lines can be drawn as perfectly antialiased 2D curves and composited, ignoring completely the clumsy attempt at rendering them with triangles. A shift to deferred renderers makes compositing different image elements easier, and I hope, with a mix of programmable hardware, developers use more varied rendering strategies, picking a method that's better for a particular job rather than trying to fit everything into the one representation.

 

I wonder if you could get a good result with standard quads alpha blended but using ddx / ddy to control the filter width.

 

I fused MLAA+DLAA and now MLAA+FXAA3 with good results. MLAA (because of the depth-information) takes care of low-contrast high-depth situations, FXAA or DLAA of everything else. DLAA is rather costy and blurry, FXAA is quick, and the fusion is very quick as well. I took the standard color-edge detector out of MLAA because that's doing FXAA then.

 

Do you have some results and comparison pics to show the final result? Or maybe even a movie were we can see this AA in action? Would be super interesting...

 

Oh finally I'm really interesting to see this combination at work, great someone has tried!!!

 

Let me try doing some screens, I can switch the effects at runtime so the screen will be pretty static. I'm not sure the given setup (with my Oblivion) is fluid enough to record videos.

 

This is an excerpt from the first scene of three which you can find here (50MB).


I tried to select a region with no movement, though the fire-socket gets lost sometimes (auto-lod-culling ).