TXAA Anti-Aliasing Makes Its Debut In Latest Update For The Secret World

[SirPauly]

Well, don't forget that he's playing at 1440p where AA usually carries big penalty.

Also, SLi may not scale as well based on the temporal nature of the feature!

 

[Akantus]

Also, SLi may not scale as well based on the temporal nature of the feature!

From what I have heard, SLI scales really bad in this game even without TXAA. (Unless it's been fixed)

 

[sontin]

He says, reduction of sharpness comes with the removal of aliasing, but SGSSAA proves him wrong. I too think he is trying to sugarcoat the severe shortcomings of TXAA. I don't blame him - he wrote the damn thing and works for Nvidia. Still, it leaves a very bitter taste.

SSAA without LoD adjustment blurs the picture, too. And to get the same result you need more than 4 samples which makes SSAA for a Single-Card user useless.

 

[BallaTheFeared]

SGSSAA > SSAA yes?

 

[boxleitnerb]

SSAA without LoD adjustment blurs the picture, too. And to get the same result you need more than 4 samples which makes SSAA for a Single-Card user useless.

As far as I know, that is not true. The LOD adjustment adds detail/sharpness that is not there when using MSAA. So it is not blurry, but less detailed than it could be without the adjustment.
Anyway, everyone in their right mind would use the LOD adjustment. And don't forget the compatibility bits to fight blur.

4xSGSSAA is perfectly fine, no need to go to 8x. If I get 90% of the smoothness and none of the blur, fine by me.

SGSSAA > SSAA yes?

SGSSAA is a subset of SSAA. SSAA is just a general term. I think when you say "SSAA" you mean OGSSAA, right? Then you are correct. SGSSAA > OGSSAA in quality/perf.

 

[BallaTheFeared]

<---- Noob

 

[ViRGE]

The cinematic look may not be for everyone, personally welcome the feature but since it is using multi-sampling; it would be also welcomed to see a multi-sampling + transparency ability as well, for the gamers that really enjoy sharper images.

Agreed. It's a cool technology and if NVIDIA wants to work on it that's fine - there are definitely a lot of people who would find this an improvement. But it shouldn't be allowed to become the de-facto method of AA, we're already having that problem with FXAA.

 

[TimothyLottes]

it shouldn't be allowed to become the de-facto method of AA

Don't worry, TXAA was never designed to replace all the other AA methods, but rather provide something different.

it would be also welcomed to see a multi-sampling + transparency ability as well, for the gamers that really enjoy sharper images.

The transparency MSAA is basically selective super-sampling. This is something I highly recommend to developers to do manually as an option to increase quality in general (especially with alpha-tested geometry). This stuff shouldn't just be a forced driver option IMO.

He says, reduction of sharpness comes with the removal of aliasing, but SGSSAA proves him wrong. I too think he is trying to sugarcoat the severe shortcomings of TXAA. I don't blame him - he wrote the damn thing and works for Nvidia. Still, it leaves a very bitter taste.

The physics of the problem proves otherwise. For example, take pixel-width green text on a black background. When this text is aligned to pixels, you get a sharp image like the text on your desktop. When the lines of the text get shifted by 1/2 a pixel, then it is blurry. If the text is moving very slowly you get a flickering on-and-off of blur, this is temporal aliasing. You cannot have ultimate sharpness and no temporal aliasing at the same time.

In the film industry, they filter out details which would lead to temporal aliasing, to insure the image is stable under any amount of motion. This is done so your eye doesn't realize that you are looking at pixels which would break the perception of reality, and so that DVD or BluRay can be resized without showing aliasing. If you are pissed at TXAA (which you don't need to use if you don't want to), then I'd highly suggest also bitching at the entire film industry, on the grounds that every time you see a movie you are getting shafted in sharpness.

BTW, I don't suger-coat anything, that's not my way, I say it how it is. If I didn't personally like TXAA and want to use it in games that I play, I would have never pushed it out.

He says TXAA is more expensive than MSAA but refers to it as a little

How about some numbers: the majority of the cost associated with TXAA is just the cost of in-game MSAA. To put this in perspective for a game like Battlefield3 which has relatively optimized MSAA support for deferred rendering, on the **mid-range mobile** 650M at 1280x800 I measured a BF3 scene at 18ms/frame with no-AA. Simply turning on 2xMSAA adds 4.4 ms/frame, or using instead 4xMSAA adds 9.2 ms/frame. Other games with deferred rendering are much worse than BF3, and games with forward rendering are much better. The cost of TXAA over just MSAA on this GPU at the same resolution is 0.67 ms/frame for 2x, and 1.34 ms/frame for 4x.

ATI's Temporal AA was dependent on rapid succession of other frames to create the "effect" of AA. So one frame really couldn't show what's happening.

Yeah that is correct. ATI's prior temporal AA simply changed the MSAA sample position each frame without blending frames. This effect required fast refresh rates paired with slow pixel switching displays. Current displays switch too fast for this to work well. Things like excessive film grain will hide the flicker, but it wouldn't apply in the general case.

TXAA's temporal works differently, the current version of TXAA is not doing a sub-pixel jitter like Crytek or some forms of SMAA (which blends two frames) or ATI's prior (which doesn't blend frames). I have some more R&D to do before I'm happy with the blending in TXAA before I use the sub-pixel jitter to improve quality on still frames.

So for now TXAA still frames look like MSAA with a different and larger filter kernel, the real effect of TXAA is in motion.

I am not saying this is traditional or conventional super-sampling.

Yeah TXAA is not conventional super-sampling. Certainly I'd rather just super-sample everything on a rotated grid, just do RGSSAA, but the cost of that is not practical in many cases. So TXAA does the next best thing, and uses samples across time. So in motion, when the temporal aliasing is bad with MSAA, you get some non-conventional form of SSAA with TXAA.

Note at some point the LOD bias could also be applied to TXAA just like RGSSAA. This is something I'm looking to improve upon on later TXAA releases once I get the jitter to the point where I like it (then you'd have 2 shaded samples/pixel on still images too, so LOD bias starts to make more sense).

Its cell shading style really begs for AA from my experience.

Games like BL2 which are effectively designed for the 360/PS3 console, apply cell shading as a post process. Traditional MSAA techniques don't work for these cell shaded games, because the outline shader re-introduces any aliasing that would have been removed by the MSAA resolve. There might be ways to correct this by using the non-resolved MSAA depth during the outline post process, but that is an R&D project which I don't believe anyone has had time to solve yet unfortunately.

Mod note: we've confirmed that this is really Timothy Lottes. Welcome Timothy
-ViRGE

 

[Siberian]

It looks like Crytek and Unreal engines will be using this, so their should be a ton of game support soon.

 

[Grooveriding]

Timothy,

Thank you for joining and addressing some of our thoughts on TXAA. Hopefully some of us running rough shod over some aspects of TXAA was not offensive I appreciate new methods of anti-aliasing and if TXAA did not introduce the amount of blur it does, I would be more interested in using it.

Your comments on BL2 sort of lead me to the impression that we are going to only see post-AA methods available for the game via options. There are some compatibility bits for forcing more 'pure' anti-aliasing options for the original Borderlands, hopefully these will be effective in the second one. As I recall turning off the outline shader allowed these to work well in the first game.

Thanks for joining and sharing more on TXAA.

 

[Gordon Freemen]

Don't worry, TXAA was never designed to replace all the other AA methods, but rather provide something different.


The transparency MSAA is basically selective super-sampling. This is something I highly recommend to developers to do manually as an option to increase quality in general (especially with alpha-tested geometry). This stuff shouldn't just be a forced driver option IMO.


The physics of the problem proves otherwise. For example, take pixel-width green text on a black background. When this text is aligned to pixels, you get a sharp image like the text on your desktop. When the lines of the text get shifted by 1/2 a pixel, then it is blurry. If the text is moving very slowly you get a flickering on-and-off of blur, this is temporal aliasing. You cannot have ultimate sharpness and no temporal aliasing at the same time.

In the film industry, they filter out details which would lead to temporal aliasing, to insure the image is stable under any amount of motion. This is done so your eye doesn't realize that you are looking at pixels which would break the perception of reality, and so that DVD or BluRay can be resized without showing aliasing. If you are pissed at TXAA (which you don't need to use if you don't want to), then I'd highly suggest also bitching at the entire film industry, on the grounds that every time you see a movie you are getting shafted in sharpness.

BTW, I don't suger-coat anything, that's not my way, I say it how it is. If I didn't personally like TXAA and want to use it in games that I play, I would have never pushed it out.


How about some numbers: the majority of the cost associated with TXAA is just the cost of in-game MSAA. To put this in perspective for a game like Battlefield3 which has relatively optimized MSAA support for deferred rendering, on the **mid-range mobile** 650M at 1280x800 I measured a BF3 scene at 18ms/frame with no-AA. Simply turning on 2xMSAA adds 4.4 ms/frame, or using instead 4xMSAA adds 9.2 ms/frame. Other games with deferred rendering are much worse than BF3, and games with forward rendering are much better. The cost of TXAA over just MSAA on this GPU at the same resolution is 0.67 ms/frame for 2x, and 1.34 ms/frame for 4x.


Yeah that is correct. ATI's prior temporal AA simply changed the MSAA sample position each frame without blending frames. This effect required fast refresh rates paired with slow pixel switching displays. Current displays switch too fast for this to work well. Things like excessive film grain will hide the flicker, but it wouldn't apply in the general case.

TXAA's temporal works differently, the current version of TXAA is not doing a sub-pixel jitter like Crytek or some forms of SMAA (which blends two frames) or ATI's prior (which doesn't blend frames). I have some more R&D to do before I'm happy with the blending in TXAA before I use the sub-pixel jitter to improve quality on still frames.

So for now TXAA still frames look like MSAA with a different and larger filter kernel, the real effect of TXAA is in motion.


Yeah TXAA is not conventional super-sampling. Certainly I'd rather just super-sample everything on a rotated grid, just do RGSSAA, but the cost of that is not practical in many cases. So TXAA does the next best thing, and uses samples across time. So in motion, when the temporal aliasing is bad with MSAA, you get some non-conventional form of SSAA with TXAA.

Note at some point the LOD bias could also be applied to TXAA just like RGSSAA. This is something I'm looking to improve upon on later TXAA releases once I get the jitter to the point where I like it (then you'd have 2 shaded samples/pixel on still images too, so LOD bias starts to make more sense).


Games like BL2 which are effectively designed for the 360/PS3 console, apply cell shading as a post process. Traditional MSAA techniques don't work for these cell shaded games, because the outline shader re-introduces any aliasing that would have been removed by the MSAA resolve. There might be ways to correct this by using the non-resolved MSAA depth during the outline post process, but that is an R&D project which I don't believe anyone has had time to solve yet unfortunately.

If TXAA is so great then why does it make the image quality worse at the same time making it better.

 

[Silverforce11]

It looks like Crytek and Unreal engines will be using this, so their should be a ton of game support soon.

Yeah, as long as they actually include proper MSAA as a core AA mode, then great, more options for all.

The last thing I would want as a PC gamer is to build high-end rigs, to run at the highest texture resolution and quality, then have the final image blurred so you lose all the definition. It's going backwards.

If it will be possible to tweak TXAA further like custom FXAA injectors, so it can provide great AA and crisp images, then perfect!

 

[boxleitnerb]

The physics of the problem proves otherwise. For example, take pixel-width green text on a black background. When this text is aligned to pixels, you get a sharp image like the text on your desktop. When the lines of the text get shifted by 1/2 a pixel, then it is blurry. If the text is moving very slowly you get a flickering on-and-off of blur, this is temporal aliasing. You cannot have ultimate sharpness and no temporal aliasing at the same time.

In the film industry, they filter out details which would lead to temporal aliasing, to insure the image is stable under any amount of motion. This is done so your eye doesn't realize that you are looking at pixels which would break the perception of reality, and so that DVD or BluRay can be resized without showing aliasing. If you are pissed at TXAA (which you don't need to use if you don't want to), then I'd highly suggest also bitching at the entire film industry, on the grounds that every time you see a movie you are getting shafted in sharpness.

BTW, I don't suger-coat anything, that's not my way, I say it how it is. If I didn't personally like TXAA and want to use it in games that I play, I would have never pushed it out.

Thank you for your reply. Fair enough.
But I would like to add, that with the film examples on your blog, there are still way more surface details that don't get washed away. Basically with TXAA, you could reduce texture quality by one or two levels and wouldn't even notice it I presume. I again have to point to SGSSAA with proper LOD adjustment. It may not be THAT smooth in motion, but very close. And very sharp and with little to no loss of surface detail and even additional detail due to the negative LOD.

 

[Akantus]

Note at some point the LOD bias could also be applied to TXAA just like RGSSAA. This is something I'm looking to improve upon on later TXAA releases once I get the jitter to the point where I like it (then you'd have 2 shaded samples/pixel on still images too, so LOD bias starts to make more sense).

I'm now much more optimistic about TXAA, because I know our voices are heard and there is still work that will be done.
When developers listen and comunicate like this, then I know it's going to be ok. :thumbsup:

 

[f1sherman]

So Timothy will look up into this, and perhaps hack LOD bias into TXAA,
or apply some other trick regarding sharpness.

And the cost of TXAA is basically the cost of corresponding MSAA.

Great new indeed!

 

[Grooveriding]

Thank you for your reply. Fair enough.
But I would like to add, that with the film examples on your blog, there are still way more surface details that don't get washed away. Basically with TXAA, you could reduce texture quality by one or two levels and wouldn't even notice it I presume. I again have to point to SGSSAA with proper LOD adjustment. It may not be THAT smooth in motion, but very close. And very sharp and with little to no loss of surface detail and even additional detail due to the negative LOD.

This is the main issue with TXAA, it's simply way too blurry and washed out. Considering the FXAA screenshots show a less blurry image, you would think there would be some way to get rid off all that smearing of the image when you use TXAA.

 

[SirPauly]

Don't worry, TXAA was never designed to replace all the other AA methods, but rather provide something different.


The transparency MSAA is basically selective super-sampling. This is something I highly recommend to developers to do manually as an option to increase quality in general (especially with alpha-tested geometry). This stuff shouldn't just be a forced driver option IMO.


The physics of the problem proves otherwise. For example, take pixel-width green text on a black background. When this text is aligned to pixels, you get a sharp image like the text on your desktop. When the lines of the text get shifted by 1/2 a pixel, then it is blurry. If the text is moving very slowly you get a flickering on-and-off of blur, this is temporal aliasing. You cannot have ultimate sharpness and no temporal aliasing at the same time.

In the film industry, they filter out details which would lead to temporal aliasing, to insure the image is stable under any amount of motion. This is done so your eye doesn't realize that you are looking at pixels which would break the perception of reality, and so that DVD or BluRay can be resized without showing aliasing. If you are pissed at TXAA (which you don't need to use if you don't want to), then I'd highly suggest also bitching at the entire film industry, on the grounds that every time you see a movie you are getting shafted in sharpness.

BTW, I don't suger-coat anything, that's not my way, I say it how it is. If I didn't personally like TXAA and want to use it in games that I play, I would have never pushed it out.


How about some numbers: the majority of the cost associated with TXAA is just the cost of in-game MSAA. To put this in perspective for a game like Battlefield3 which has relatively optimized MSAA support for deferred rendering, on the **mid-range mobile** 650M at 1280x800 I measured a BF3 scene at 18ms/frame with no-AA. Simply turning on 2xMSAA adds 4.4 ms/frame, or using instead 4xMSAA adds 9.2 ms/frame. Other games with deferred rendering are much worse than BF3, and games with forward rendering are much better. The cost of TXAA over just MSAA on this GPU at the same resolution is 0.67 ms/frame for 2x, and 1.34 ms/frame for 4x.


Yeah that is correct. ATI's prior temporal AA simply changed the MSAA sample position each frame without blending frames. This effect required fast refresh rates paired with slow pixel switching displays. Current displays switch too fast for this to work well. Things like excessive film grain will hide the flicker, but it wouldn't apply in the general case.

TXAA's temporal works differently, the current version of TXAA is not doing a sub-pixel jitter like Crytek or some forms of SMAA (which blends two frames) or ATI's prior (which doesn't blend frames). I have some more R&D to do before I'm happy with the blending in TXAA before I use the sub-pixel jitter to improve quality on still frames.

So for now TXAA still frames look like MSAA with a different and larger filter kernel, the real effect of TXAA is in motion.


Yeah TXAA is not conventional super-sampling. Certainly I'd rather just super-sample everything on a rotated grid, just do RGSSAA, but the cost of that is not practical in many cases. So TXAA does the next best thing, and uses samples across time. So in motion, when the temporal aliasing is bad with MSAA, you get some non-conventional form of SSAA with TXAA.

Note at some point the LOD bias could also be applied to TXAA just like RGSSAA. This is something I'm looking to improve upon on later TXAA releases once I get the jitter to the point where I like it (then you'd have 2 shaded samples/pixel on still images too, so LOD bias starts to make more sense).


Games like BL2 which are effectively designed for the 360/PS3 console, apply cell shading as a post process. Traditional MSAA techniques don't work for these cell shaded games, because the outline shader re-introduces any aliasing that would have been removed by the MSAA resolve. There might be ways to correct this by using the non-resolved MSAA depth during the outline post process, but that is an R&D project which I don't believe anyone has had time to solve yet unfortunately.

Mod note: we've confirmed that this is really Timothy Lottes. Welcome Timothy
-ViRGE

Precious information, thanks!

 

[SirPauly]

And the cost of TXAA is basically the cost of corresponding MSAA.

That's the key to me and its anti-aliasing abilities with-in a moving screen are really impressive.

 

[blackened23]

That's the key to me

With this statement being in reference to TXAA having a similar cost to MSAA, I don't buy it.

#1) There is no proof of that being true, words on paper can say this but I just can't see it happening in a game. In other words, words are cheap, nvidia needs to prove it by refining the performance of TXAA.

#2) The performance hit of TXAA in TSW is outrageous. I thought the chart on the geforce.com website was somewhat funny because it suggested no performance loss with TXAA, and thats quite not true. ~15 fps loss during stills and 35-40 fps loss during combat on my end compared to FXAA HQ.

#3) with MSAA you can "build" on it with SGSSAA or TrSS.

I don't see 2x MSAA causing anywhere NEAR the performance hit that TXAA 2x-4x causes for me in TSW. The game is basically unplayable with 680 sli, and that is ridiculous - during heavy combat with 2-3 mobs and spell effects the game basically crawls. I certainly run a ton of games with MSAA at 2560 resolution and none of them are unplayable, really the only ones I can see suggesting that the performance hit being the same as MSAA are yourself and mostly ones that haven't even used TXAA yet (quite a few GTX 400 owners!). During combat the game slows to a crawl at times with even 2x TXAA. I can't fathom MSAA 2x causing that kind of performance hit. I can run crysis 2 and metro 2033 with MSAA just fine FFS.

I saw on Mr. Lottes' Blog that SLi scaling isn't ideal with TXAA, maybe that is the case, I don't know. I feel like this feature should be usable on systems with 2 of the fastest graphics cards available, I hear that even at 1080p - users are reporting poor TXAA performance (ideal being 60 fps). For such a feature I think 2x 680 sli should be up to the task, otherwise something is wrong. Having more options is great, but i'd like it a lot more if it were usable.

 

[SirPauly]

Why don't you stop offering words and showcase some examples.

 

[BFG10K]

SSAA is already capable of delivering what TXAA does, without the blurring. With it I see perfect image quality in many games I play, and during movement there’s not a single pixel out of place.

Heavily zoomed screenshots usually show no significant blur over MSAA, or even slight increases in sharpness from OGSS. With the extremely rare titles that show visible blurring with SGSS, adjusting the LOD makes the game sharper while still maintaining less aliasing than edge based schemes.

As for adopting what movies do, movies are generally capped at 1080p and 24-30FPS. Shall we bring that into PC space as well? Movies also have more realistic “textures” and “lighting” than real-time graphics because they’re pre-rendered, so they can afford to trade some of that detail to reduce aliasing.

Maybe if all games sampled from 16K x 16K textures at a minimum, TXAA wouldn’t be so bad.

 

[blackened23]

Why don't you stop offering words and showcase some examples.

Showcase examples of what? I've already shared pics.

 

[blackened23]

more action with TXAA:

I enlarged the MSI afterburner overlay so you can view the framerate.

 

[SirPauly]

Showcase examples of what? I've already shared pics.

Examples of FXAA, TXAA 2 and TXAA 4 disparity in frame-rate! I have found the frame-rate fine with x2 TXAA at 1080P with a GTX 670 so far.

 

[SirPauly]

more action with TXAA:

I enlarged the MSI afterburner overlay so you can view the framerate.

With FXAA HQ, x2 TXAA and TXAA 4!