TEMPORAL ANTI-ALIASING like this would be awesome in Elite Dangerous.

The problem with rendering a larger image, and then scaling it down (as ED Supersampling does, as I understand it) is, it take MORE video memory to hold that large extra video buffer i.e. 4K at 2.0 super-sample, required a 4K and 8K image buffer. You render to 8K and then resize to 4K and display. If you think about it, this ONLY gives 4 more sample per pixel. So to me, it does not seem efficient. And it also uses a LOT of extra Video RAM.

If however you render to 4K, but render multiple samples per pixel, then you could possible use 8 or 16 samples, with a similar speed. But using 4x or 8x mem PER pixel, BEFORE those samples were combined and the result used for the pixel. You are, kind of, rendering 8k or 16ks worth of image, but its more memory efficient, as the extra data is thrown away once combined, per pixel.

Hey, at the end of the day, Im not a hard nut GFX pipe coder. So what I am saying may not make sense. Im just guessing to some extent.

You know, thinking about it, re-using samples for a given surface point from an earlier frame, actually makes a lot of sense. Some shading, like Ambient Occlusion, and Diffuse shading, by its nature, is generally directionally independent. This basically means, the shading on a chalk sphere looks the same, whether looking at it from its left, right, or above. indeed, if you have combined multiple diffuse samples for a pixel, and then do so on the next frame. Combining those samples would be relatively cheap.

However, reflective shading, IS very much dependant on your view. Small changes in view, can mean large changes in the reflection, especially on curved surfaces. However, again, these samples from previous frame ARE quite cheap to keep, and would tend to reduce dark/bright flicker when combined, as they average out the shift over time.

So I guess Temporal AA may make more sense than it seems at first. As its kind of recycling previous, expensively aquired samples on subsequent frames, getting better value PER sample.
 
Last edited:
TAA would have way less of a performance hit than EDs supersampling option but offer better results than MSAA x4.
 
With EDs massive draw distances you risk losing a lot of details completely if you over do the AA.

AA, always "loses" detail, is not the right way to think about it...

AA reduces apparent sharpness of detail, as jagged (high contrast) pixels appear "sharper" than the blurred appearance of AA pixels. But in truth, there is actually MORE detail in AA pixels, as they contain multiple samples of the geo there, which is closer to reality than a single sample... See my post below.
 
Last edited:
How Uncharted 4 does it for the interested. Not quite simple as just turn it on. . :)

(200mb pptx)

http://advances.realtimerendering.com/s2016/s16_Ke.pptx

It's from this year's SIGGRAPH.

More presentations on various things if you scroll right down.

http://advances.realtimerendering.com/s2016/

Thanks for the link! :) Interesting read.

Edit: I'm more and more impressed by the Alias Isolation TAA mod. It works wonderfully and adds back the immersion that was lost with the fugly jaggies.
 
Last edited:
Yeah, now that I have seen it, Im a bit obsessed with it. Need to put some time aside to read through both articles on how this is achieved.

Last night I was flying into a planet side city. This is on a 4K screen, with SMAA on. And all I could see with the city at 10k was twinkle twinkle twinkle, as the bright pixels in its window aligned with the pixel grid, and then did not, from frame to frame.

Id love to see it just resolve as a blurry set of lights in the distance.

Lets all push for TXAA in a future point release. I was actually looking forward to 8K screens, thinking that as pixels get smaller, their scintillation will be less distracting. But I can see now that TXAA, especually when ramped up, gives a much better solution. Not saying that 8K wont be a bad thing, one day :)
 
Oh god, if it's possible please!!!

The black white flickering are such an annoyance, yes you learn to live with it but it's not good.

Agreed!

- - - - - Additional Content Posted / Auto Merge - - - - -

Yeah, now that I have seen it, Im a bit obsessed with it. Need to put some time aside to read through both articles on how this is achieved.

Last night I was flying into a planet side city. This is on a 4K screen, with SMAA on. And all I could see with the city at 10k was twinkle twinkle twinkle, as the bright pixels in its window aligned with the pixel grid, and then did not, from frame to frame.

Id love to see it just resolve as a blurry set of lights in the distance.

Lets all push for TXAA in a future point release. I was actually looking forward to 8K screens, thinking that as pixels get smaller, their scintillation will be less distracting. But I can see now that TXAA, especually when ramped up, gives a much better solution. Not saying that 8K wont be a bad thing, one day :)

Oh God! I was hoping my impending upgrade from 1080p to 1440p might help with this matter... And there's you complaining of it at 4K :(
 
  • Like (+1)
Reactions: NW3
Oh God! I was hoping my impending upgrade from 1080p to 1440p might help with this matter... And there's you complaining of it at 4K :(
I also have a 4K monitor and still found the pixel crawl to be highly visible. Unfortunately, I can't use Super Sampling at that resolution (or even at 2560x1440) and maintain a reasonable frame rate on my nVidia GTX 980 Ti card.

I'm currently using SS=2.0 at a custom resolution of 2272x1278 which yields a framerate of ~80-50 fps. It looks decent, but is a bit blurry and still suffers from minor pixel crawl.

I'd really like FD to spend some time improving the AA in ED.
 
Last edited:
I also have a 4K monitor and still found the pixel crawl to be highly visible. Unfortunately, I can't use Super Sampling at that resolution (or even at 2560x1440) and maintain a reasonable frame rate on my nVidia GTX 980 Ti card.

I'm currently using SS=2.0 at a custom resolution of 2272x1278 which yields a framerate of ~80-50 fps. It looks decent, but is a bit blurry and still suffers from minor pixel crawl.

I'd really like FD to spend some time improving the AA in ED.

Well, you've just saved me £300-700...

I was going to buy a 1440p monitor & Nvidia 1070 in a hope of improving fidelity such as this... Maybe I should rethink...
 
Well, you've just saved me £300-700...
I was going to buy a 1440p monitor & Nvidia 1070 in a hope of improving fidelity such as this... Maybe I should rethink...

Higher resolution is always an improvement in quality (well, until reaching visible limits at 400+ dpi). Maybe check if some large computer chain has monitors setup that you can check out - of course testing Elite will be difficult.
 
Well, you've just saved me £300-700...

I was going to buy a 1440p monitor & Nvidia 1070 in a hope of improving fidelity such as this... Maybe I should rethink...
I'm sorry I burst your bubble, but at least I saved you from spending a lot for a disappointing result. I think a GTX 1080 could handle 2560x1440 with SS=2.0, assuming you have a decent CPU. It's possible that a 1070 might be sufficient, especially if your monitor is G-Sync compatible. I saw framerates between 40 and 60 fps; most of the times it was in the 50s.

If FD doesn't improve AA, I'll probably wind up buying a "GTX 1180" next year, so that I can play at 4K w/ SS.
 
Last edited:
Got an answer from the TAA mod author:

"Hey! I'm glad you liked the mod :)

Some people have asked me to apply the same techniques to other games, most notably Rise of the Tomb Raider; I'm afraid that my answer in all of those cases is: I don't really have the resources or motivation. If I were an avid player of any of those games, I would consider that. Otherwise however, I don't want to spend my personal time on doing so. It's a pretty tedious process, and would take me a few weeks for every game.

If Elite: Dangerous is still under active development, the devs should just do it. Temporal AA should be pretty simple for the developers to implement. The most difficult part is the temporal reprojection (aka finding velocity vectors) of complex animated objects, such as animated characters and vegetation. In the case of Elite, it's all just rigid objects as far as I know. It shouldn't take the developers more than a week to implement a basic version of Temporal AA, especially that open source implementations thereof exist, and considering that they have implemented a gamma of other post-process anti-aliasing methods."
 
Havent peeps tried forcing anti aliasing in the Nvidia control panel?

That doesn't change the fact that MSAA is not compatible with deferred rendering. You can try force it through Nvidia inspector or even control panel, nothing will happen.
 
Havent peeps tried forcing anti aliasing in the Nvidia control panel?

I gave it a try, multiple settings but it doesn't seem to change anything in the game (at least to my eyes). For giggles bumped it up to 32x on one of the AAs and didnt hit frame rates so assume its not working.

Does it work for you?

Maybe its my system
 
Got an answer from the TAA mod author:

If Elite: Dangerous is still under active development, the devs should just do it. Temporal AA should be pretty simple for the developers to implement. The most difficult part is the temporal reprojection (aka finding velocity vectors) of complex animated objects, such as animated characters and vegetation. In the case of Elite, it's all just rigid objects as far as I know. It shouldn't take the developers more than a week to implement a basic version of Temporal AA, especially that open source implementations thereof exist, and considering that they have implemented a gamma of other post-process anti-aliasing methods.

^^ Well, Frontier? Please?
 
I watched the Siggraph slide set, on their implantation of TXAA. It was very interesting. The basic take-aways are....

* TXAA does take quite a lot of preparation in your shaders. Its not a switch on, and it works.
* It works better when frame rates are high

And its this point that is THE BIG THING

* It allows for techniques, that are too expensive to calculate on a single frame.

So, good quality AA, is too expensive when calculated on a single frame. Say, using 4 samples. See my post for an explanation of what samples are. But basically, imagine a real view, out of a window. And you want to render that as a pixel image. So you place a grid over the window, with 1920 holes on its X axis and 1080 on its Y. as you look through each grid hole, in turn, you take ONE sample, and thats represented by an infinitely small line drawn from your eye (centre) to the centre of the grid, and onwards to the view beyond. Where that line intersects your view geo (in a real scene, that might be a hill 10 miles away!), thats the colour you must record (that area of the hill, visible through the one grid square, might actually contain trees, a house, sheep, purple heather, you name it....). A real eye would receive ALL the light coming through that grid, and use its average. But a computer ONLY sees the colour of a single point on the viewed surface.

That single colour, DOES NOT represent the full complexity of what can be seen through the grid, and this is the cause of aliasing.

If you fired four views through that grid, each slightly offset from each other, and combine those values, your closer to reality. But, it still looks quite pixelated, Aliasing is still an issue.

And this is where the magic happens. TXAA, uses a pre-prepared pattern, of pseudo random view offsets, ordered in such a way, that as you take more samples, each sample gets as much NEW information as possible. If you used a real RANDOM set of sample points, there is always the danger that one or two samples would actually be pretty much exactly the same. A waste of time... imagine these offsets as being on a compass. With real random, you might get N,S,E,E,SE,E,W,N,SW. With pseudo random, you'd instead get N,S,E,W,SE,NW,NE,SW. It looks random, but is not, its hand selected, to avoid repetition of previously sampled areas.

So, with this set of pre-arranged samples, you can do 4 samples on frame 1, 4 on frame 2, and so on. And each set of 4 samples builds on the quality gained in the previous frames 4 samples. Genius.

Now, you can use relatively low sample counts per frame. Well, really you use the max that the target fps can afford. But you dont aim for final quality on that one frame. Then you rely on the same number of samples on subsequent frames, to up the quality, incrementally. In the Siggraph paper, they talk about using ONLY the previous frame for this, but my guess is, they actually use more.

Now the downsides.

TXAA relies on samples of pixels (view grids) from previous frames. So...

* What if there were NO previous frames (starting rendering)
* What if some geo visible now, was not in a previous frame i.e. it was off screen, or occluded by another object.

To get around this, various poly-fillers were used in their code (the really clever stuff).

Another thing to bear in mind. As the camera moves, the geo you are rendering in this frame, may have been in another position in the previous frame. So, you cant just get 4 samples in grid 100 (pixel 100) on this frame, and get the other 4 samples from the same grid in the previous frame, its moved. So to accommodate this, a motion buffer is generated by the shaders, and this gives the direction, and distance to move the view ray, to find the same geometry on a previous frame. Its this ability to look back in time, and accurately find the same point on a surface (and its sample packet) over and over that IS TXAA keystone. Again, Genius.

So, going back to some of my points at the start of this post. Because TXAA relies on previous frames samples, if there are MORE frames in a shorter period of time, those shared samples are more coherent (similar). And this is better. As such, using TXAA is more practical in 60fps, than 30, and would be even better at 90fps (good news for VR :) )

The other point, is that if you wanted to calculate better Ambient Occlusion, or Subsurface Scatter, or bounce lighting, you can. These effects would be too noisy to calculate in a single frame. But what about if you had 10 frames to spread that calc over? With TXAA, this is what you have. Sampled effects (take your pick) that are too noisy in 60fps, single frame samples, CAN be achieved when done over 10 frames at 60fps. At 90fps, you could even do less samples per frame (less work), but use more frames, or use the same samples, over more frames, to get better quality.

TXAA, IS the future.
 
Last edited:
Back
Top Bottom