Digital Foundry Article Technical Discussion Archive [2011]

Status
Not open for further replies.
Another top notch article from Grandmaster!:smile:

Digital Foundry: The Future of Anti-Aliasing

http://www.eurogamer.net/articles/digital-foundry-future-of-anti-aliasing

I got to add that I prefer FXAA console(sharp settings) rather than "filmic" one,it blurs image bit to much IMO.Also,less subpix AA but much sharper image.
fxaa39-c.png


subpix set to 0.

fxaa39-d.png


subpix at default.
 
its like using 4xMSAA without transparency AA turn on. This case, a cheap MLAA solution seems to be better if you want sharper iq.
 
its like using 4xMSAA without transparency AA turn on. This case, a cheap MLAA solution seems to be better if you want sharper iq.

Actually, looking back at the Digital Foundry article, in my opinion FXAA-sharp should pretty much match the quality of MLAA.
 
It'll really depend on the situation. Whomever is in charge of the final settings, I hope there's proper evaluation to see if it fits the art.
 
I dunno, all these post filters are still just a weak replacement for the real thing IMHO ;)

Sure, I understand, they're better than nothing, faster than MSAA, but still...
 
Actually, looking back at the Digital Foundry article, in my opinion FXAA-sharp should pretty much match the quality of MLAA.

the enslaved comparison on page 1 definitely show that it is not, even FXAA quality on PC I can spot the lower quality AA easily. Probably due to alot of long edge like what the author said. And the IQ sharpness take a huge hit. When games use QAA ppl hate it, same reason here for me that I prefer MLAA over FXAA. Even with castlevania, its noticeable that the AA quality is worse and the IQ also take a hit. If I have to loose the sub pixel AA to increase the sharpness for FXAA, I might as well use MLAA, since its suppose to cost around 1-2 ms with Jimenez MLAA on GPU, not much more expensive than FXAA3 right now. I wonder if games like R3 can benefit from FXAA, if its sub HD now with QAA, ditching QAA can free up more memory and FXAA blurriness won't affect they already have now if its sub hd. Potentially Higher res + fxaa should be better than low res + QAA right?
 
I dunno, all these post filters are still just a weak replacement for the real thing IMHO ;)

I wouldn't say they're so weak, considering how badly MSAA fails with a high dynamic range (unless you keep the subpixel info all the way through post processing, and that's just crazy).
 
My base of comparision is our offline stuff ;) We do significantly more complex shading and lighting and have a lot mroe geometry and texture detail so it wouldn't ever work with a post filter anyway. The general sharpness and clarity of the images we get is still completely out of reach for realtime rendering and these post filters always feel like a hopeless attempt to catch up, at least to me.
Of course I also realize that the performance loss for a similar image quality would be far above 50% and thus it'll always remain unrealistic - but in some ways it's actually a good thing. We'll still have a job for a while ;)
 
Well, keep in mind, we're talking about fixed console hardware from 5+ years back. In the PC arena, there's no denying the superiority of multiple samples.
 
Even that (MSAA) is only sampling geometry - but lighting and shading should be at least equally important, if not more.

Then we move to the realm of indirect lighting and area shadows - or as a replacement, ambient occlusion - and the various kinds of sampling errors and noise will also become an issue. Then there are glossy reflections, translucency, motion blur and DOF... all these can be described as spatial or temporal aliasing issues that can't be solved by a simple post filter. Cheats work in many cases, but the real deal is always to increase the number of samples. Especially when you start to move into the realm of raytracing, and although I've been against that some years ago, times are changing and it is now a driving force because of its simplicity and the easy trade-off between artist hours vs. rendering power (the later being cheaper most of the time, in offline CG)

But more samples will always be secondary when you have to fit into a performance envelope and trade-offs will have to be made to stay competitive. If everyone was content with 640x480 on the HD consoles than noone would research cheap post filter alternatives, but for now resolution is still the more important element of image quality to stick to, going under 1024*640 is considered to be too much, and righly so. For now, it is unacceptable to spend performance on more samples per pixel vs. more pixels.
 
I love the concept of FXAA 1.7-2 times higher resolution image and then downscaling it, it should eliminate almost every problem we have now with aliasing and almost dont touch the sharpness of the image.

What a pity that Richard didnt post 1080p videos from FXAA Quality test on RE 5 benchmark, it would be nice to see it in action.
 
I love the concept of FXAA 1.7-2 times higher resolution image and then downscaling it, it should eliminate almost every problem we have now with aliasing and almost dont touch the sharpness of the image.

What a pity that Richard didnt post 1080p videos from FXAA Quality test on RE 5 benchmark, it would be nice to see it in action.

Why not just display the higher resolution image? Would be an improvement on what we have today.
 
fixed resolution displays. Render slightly above 720p. 1080p is basically unattainable this gen.

Who has a 720p display these days? In most cases, the 720p framebuffer would be scaled up by the display. Better to output 1080p, the most common display resolution.
 
Who has a 720p display these days? In most cases, the 720p framebuffer would be scaled up by the display. Better to output 1080p, the most common display resolution.
I have a 720p display connected to my Xbox. Not everyone buys a new TV every 5 years.
 
I think FXAA looks better than MLAA, just because it handles subpixel aliasing better, it does blur the image somewhat but I don't mind that as it looks more realistic than harsh polygonal edges, jaggies and shimmering everywhere.
 
I think FXAA looks better than MLAA, just because it handles subpixel aliasing better, it does blur the image somewhat but I don't mind that as it looks more realistic than harsh polygonal edges, jaggies and shimmering everywhere.

IMHO I don't understood in what exactly is better FXAA to MLAA, I have read a lot of enthusiastic comments about that & of course like MLAA in the capture seems good but from the video is an another story... surely in motion don't gives the same artifacts of MLAA but from a bunch of videos in the linear edges just not work, but even in a lot of differents situations seem almost not work compared to MLAA. I guess at the end maybe FXAA is very helpful more to support MSAA or MLAA works, but alone not convice me so much.
 
IMHO one thing that should be taken into account is that in most of these pix/videos we aren't seeing driver level optimizations.

That said, I'm much more impressed with MLAA than FXAA. MASS isn't the answer to everything itself. It doesn't touch a lot of alpha, so you end up with hideous vegetation while all the regular geometry is nice and smooth at 8xMSAA or greater.

I believe the answer lies between these solutions, in some sort of combination perhaps.
 
Who has a 720p display these days? In most cases, the 720p framebuffer would be scaled up by the display. Better to output 1080p, the most common display resolution.

Yeah, most people have 1080p displays, so you'd get scaling artifacts, but the image would be supersampled with FXAA. It sounds kind of weird to render, downsample and then upscale, but you should get higher quality AA than regular MSAA, from what I understand. Not an expert in any sense, so one of the more tech oriented people could probably explain why this might be a good idea. They are basically suggesting supersampling but with a much smaller memory and bandwidth penalty.
 
Status
Not open for further replies.
Back
Top