Actually, I wouldn't mind if they would 'test' it in Ps1 & Ps2 back compability first..
Last edited by a moderator:
Alternative AA methods and their comparison with traditional MSAA*
- Thread starter mitran
- Start date
brain_stew
Regular
My experience with temporalAA solutions in the past has always led me to believe its something that only really works if you can guarantee a consistant framerate. How then does Crytek's temporalAA solution differ considering its targeting a platform where you can't guarantee a consistant refresh (the PC) and everything we've seen of the engine on consoles so far has suggested it will regularly drop frames even there?
http://advances.realtimerendering.c... 2010 Advanced RealTime Rendering Course).pdf
They talk about it at the ned of that PDF. Basically, edge blurring for foreground objects and temporal AA for background objects which seems like a pretty smart solution to me assuming it works as advertised. There's a video of it in practice here as well:
http://advances.realtimerendering.c... 2010 Advanced RealTime Rendering Course).avi
http://advances.realtimerendering.c... 2010 Advanced RealTime Rendering Course).pdf
They talk about it at the ned of that PDF. Basically, edge blurring for foreground objects and temporal AA for background objects which seems like a pretty smart solution to me assuming it works as advertised. There's a video of it in practice here as well:
http://advances.realtimerendering.c... 2010 Advanced RealTime Rendering Course).avi
ATI's "temporal AA" from R3xx days is different; it changes the multisample position per frame, so that's where a high framerate is necessary to perceive double the MSAA. The change is swift enough to fool the eye.
Crytek's temporal AA stores the previous frame & depth to do comparisons and etc as described in the presentation. I wouldn't be surprised if Halo Reach's temporal AA is similar considering what I observed. i.e. samples with little depth delta accumulate, giving the AA effect. So for far away objects, you get the AA most of the time. Of course, objects closer to the screen change in position (or rather depth) much "faster" (i.e. compared to the previous frame), so there's no accumulation. Hence, Crytek's hybrid approach with the additional post-process AA for objects nearer the camera.
Crytek's temporal AA stores the previous frame & depth to do comparisons and etc as described in the presentation. I wouldn't be surprised if Halo Reach's temporal AA is similar considering what I observed. i.e. samples with little depth delta accumulate, giving the AA effect. So for far away objects, you get the AA most of the time. Of course, objects closer to the screen change in position (or rather depth) much "faster" (i.e. compared to the previous frame), so there's no accumulation. Hence, Crytek's hybrid approach with the additional post-process AA for objects nearer the camera.
brain_stew
Regular
Thanks for the explanation, I was quite impressed with Halo's implementation (especially for distant scenery and more static scenes) so I look forward to seeing how well this works in practice.
I'm curious, since its distant (sub pixel) detail that MLAA has its biggest problems with would a hybrid MLAA+temporalAA not be a potentially decent solution for other titles? I know Crytek discounted it due to performance reasons but it'd be interesting to see someone else attempt it. From my very basic understanding, it would seem this approach would cover most aliasing sources and deliver potentially fantastic image quality.
Seems like Under Siege for PSN will be using MLAA in 720p mode.
http://www.neogaf.com/forum/showpost.php?p=22661479&postcount=29
http://www.neogaf.com/forum/showpost.php?p=22661479&postcount=29
user542745831
Veteran
Came across this post:
http://www.crymod.com/thread.php?postid=735373#post735373
Seems like as if they posted some more about it somewhere else:
http://www.gamedev.net/community/forums/topic.asp?topic_id=580517
http://www.crymod.com/thread.php?postid=735373#post735373
Seems like as if they posted some more about it somewhere else:
http://www.gamedev.net/community/forums/topic.asp?topic_id=580517
Last edited by a moderator:
Is there any way we can download this and try it ourselves?
Since we recently got our new console tech running at stable vsynch locked 60 fps, I tested the old school perspective matrix (screen space +0.5, +0.5 subsample jitter) temporal AA. It looked surprisingly good. Basically it provides almost 2x supersampled quality (if also reduces pixel shader aliasing and alpha clip aliasing) completely free of cost. And it works perfectly with any rendering technique (we use deferred shading). Naturally without perfectly vsynched 60 fps it looks pretty bad, since you will notice the jittering. But for 60 fps vsynch locked games it looks like a promising alternative for games that don't have any extra performance to spend on AA. Too bad this kind of TAA isn't visible in screenshots
Since we recently got our new console tech running at stable vsynch locked 60 fps, I tested the old school perspective matrix (screen space +0.5, +0.5 subsample jitter) temporal AA. It looked surprisingly good. Basically it provides almost 2x supersampled quality (if also reduces pixel shader aliasing and alpha clip aliasing) completely free of cost. And it works perfectly with any rendering technique (we use deferred shading). Naturally without perfectly vsynched 60 fps it looks pretty bad, since you will notice the jittering. But for 60 fps vsynch locked games it looks like a promising alternative for games that don't have any extra performance to spend on AA. Too bad this kind of TAA isn't visible in screenshots
Wow, that's fantastic sebbbi, keep up the good work.
Of stationary polys, sure, but if an edge has moved a couple of pixels between frames, it'll be fully aliased. It's certainly a good idea to incorporate in any 60fps title, and should be a good fit for some types of games, but in others I'll be surprised if it's visibly comparable to conventional 2xMSAA, at least in the majority of contexts. I suppose any into-the-screen type 1st/3rd person game (FPS or racer) would receive equivalent to 2xSS on distant geometry that isn't changing much, although swinging the camera around will break it. And sadly so few devs target 60fps these days to be able to get it for free!
Yes, this kind of TAA looks best for the distant geometry. But in the fast moving near part of the scene the aliasing is harder to notice. For example Halo Reach disables their TAA solution completely for the fast moving near objects, and Cryengine 3 switches their TAA for edge blurring. Fortunately we have around 50x higher view distance in our new game, so there's lots of distant geometry to draw. And free things are always tempting... If the cheap TAA works on real HDTVs as well as it does on computer monitors (no visible jittering or motion compensation getting crazy), I am pretty sure we'll include it in the final product.
I agree with you that we would take fast hardware 2xMSAA any time over the free 2xTAA, as would any developer. Unfortunately hardware MSAA does not mix that well with deferred shading (and many other image space techniques). As we are vsynch locked to 60 fps there's no way we could resolve the MSAA scene in 2x resolution like 30 fps games such as LittleBigPlanet and KillZone 2 did and brute force the lighting calculation in double resolution. Post process antialiasing filters seem tempting, but even the cheapest ones take around 2ms (that's already 12.5% of the whole frame at 60 fps) and have drawbacks.
MLAA could also give you 1x, the point is what the average is, not the peaks nor lows alone. 4-8xMSAA gives a far more balanced average and includes all geometry, even subpixel sized.
Still, pictures ? In an actual game, the developers will choose the tech that best showcase their games overall. And they will cover up flaws via various techniques.
Games are X number of screens in motion.
And what the hardware is cappable of else all games would be super sampled for best IQ.
But obviously flaws still exists and the average still lags behind quite a bit.
Calm down now, you are getting over-excited!
The average of edges getting AA in a scene/game. You would want as many edges as possible getting AA and also transparencies getting AA. So as said you want as few as possible edges to have crawling jaggies that disturb. For example I can play BFBC2 with 8xMSAA but game looks far worse (as in IQ cleanliness) than SC:Convition/MW2 with 4xMSAA just becouse BFBC2 has lots of edges missing out AA. It's a mish mash of smooth edges and jaggy edges and in motion it sure is one helluva deal breaker.
The average of edges getting AA in a scene/game. You would want as many edges as possible getting AA and also transparencies getting AA. So as said you want as few as possible edges to have crawling jaggies that disturb. For example I can play BFBC2 with 8xMSAA but game looks far worse (as in IQ cleanliness) than SC:Convition/MW2 with 4xMSAA just becouse BFBC2 has lots of edges missing out AA. It's a mish mash of smooth edges and jaggy edges and in motion it sure is one helluva deal breaker.
You'll have to check out PC screenshots at different sites or articles/benches with MSAA/SSAA.
I can however fast link to these screenshots as I remember the page from screenshot thread (my own). They are from HL2 CM10.4 and 8xMSAA/TSAA. MSAA might break on some edges with high contrast and HDR and this game is purely DX9 based.
http://forum.beyond3d.com/showpost.php?p=1428928&postcount=1367
http://forum.beyond3d.com/showpost.php?p=1428929&postcount=1368
http://forum.beyond3d.com/showpost.php?p=1428931&postcount=1369
I can however fast link to these screenshots as I remember the page from screenshot thread (my own). They are from HL2 CM10.4 and 8xMSAA/TSAA. MSAA might break on some edges with high contrast and HDR and this game is purely DX9 based.
http://forum.beyond3d.com/showpost.php?p=1428928&postcount=1367
http://forum.beyond3d.com/showpost.php?p=1428929&postcount=1368
http://forum.beyond3d.com/showpost.php?p=1428931&postcount=1369
Similar threads
