Before I address temporal supersampling, just a quick reminder on what aliasing is.
Aliasing is a problem that is very well defined in signal theory. According to the general sampling theorem we need to have our signal spectrum containing only frequencies lower than Nyquist frequency. If we don’t (and when rasterizing triangles we always will as triangle edge is infinite frequency spectrum, step-like response) we will have some frequencies appearing in the final signal (reconstructed from samples) that were not in the original signal. Visual aliasing can have different appearance, it can appear as regular patterns (so-called moire), noise or flickering.
Classic supersampling is a technique that is extremely widely used by the CGI industry. Per every target image fragment we perform sampling multiple times at much higher frequencies (for example by tracing multiple rays per simply pixel or shading fragments multiple times at various positions that cover the same on-screen pixel) and then performing the signal downsampling/filtering – for example by averaging. There are various approaches to even easiest supersampling (I talked about this in one of my previous blog posts), but the main problem with it is the associated cost – N times supersampling means usually N times the basic shading cost (at least for some pipeline stages) and sometimes additionally N times the basic memory cost. Even simple, hardware-accelerated techniques like MSAA that do estimate only some parts of the pipeline (pixel coverage) in higher frequency and don’t provide as good results, have quite big cost on consoles.
But even if supersampling is often unpractical technique, it’s temporal variation can be applied with almost zero cost.
Temporal supersampling theory
So what is the temporal supersampling? Temporal supersampling techniques base on a simple observation – from frame to frame most of the on-screen screen content do not change. Even with complex animations we see that multiple fragments just change their position, but apart from this they usually correspond to at least some other fragments in previous and future frames.
Based on this observation, if we know the precise texel position in previous frame (and we often do! Using motion vectors that are used for per-object motion blur for instance), we can distribute the multiple fragment evaluation component of supersampling between multiple frames.
What is even more exciting is that this technique can be applied to any pass – to your final image, to AO, screen-space reflections and others – to either filter the signal or increase the number of samples taken. I will first describe how it can be used to supersample final image and achieve much better AA and then example of using it to double or triple number of samples and quality of effects like SSAO.
I have no idea which game was the first to use the temporal supersampling AA, but Tiago Sousa from Crytek had a great presentation on Siggraph 2011 on that topic and its usage in Crysis 2 . Crytek proposed using a sub pixel jitter to the final MVP transformation matrix that alternates every frame – and combine two frames in post-effect style pass. This way they were able to increase the sampling resolution twice at almost no cost!
Too good to be true?
Yes, the result of such simple implementation looks perfect on still screenshots (and you can implement it in just couple hours!***), but breaks in motion. Previous frame pixels that correspond to current frame were in different positions. This one can be easily fixed by using motion vectors, but sometimes the information you are looking for was occluded or had. To address that, you cannot rely on depth (as the whole point of this technique is having extra coverage and edge information from the samples missing in current frame!), so Crytek proposed relying on comparison of motion vector magnitudes to reject mismatching pixels.
***yeah, I really mean maximum one working day if you have a 3D developer friendly engine. Multiply your MVP matrix with a simple translation matrix that jitters in (-0.5 / w, -0.5 / h) and (0.5 / w, 0.5 / h) every other frame plus write a separate pass that combines frame(n) and frame(n-1) together and outputs the result.
Usage in Assassin’s Creed 4 – motivation
For a long time we relied on FXAA (aided by depth-based edge detection) as a simple AA technique during our game development. This simple technique usually works “ok” with static image and improves its quality, but breaks in motion – as edge estimations and blurring factors change from frame to frame. While our motion blur (simple and efficient implementation that used actual motion vectors for every skinned and moving objects) helped to smooth edge look for objects moving quite fast (small motion vector dilation helped even more), it didn’t do anything with calm animations and subpixel detail. And our game was full of them – just look at all the ropes tied to sails, nicely tessellated wooden planks and dense foliage in jungles! :) Unfortunately motion blur did nothing to help the antialiasing of such slowly moving objects and FXAA added some nasty noise during movement, especially on grass. We didn’t really have time to try so-called “wire AA” and MSAA was out of our budgets so we decided to try using temporal antialiasing techniques.
I would like to thank here especially Benjamin Goldstein, our Technical Lead with whom I had a great pleasure to work on trying and prototyping various temporal AA techniques very late in the production.
Assassin’s Creed 4 XboxOne / Playstation 4 AA
As a first iteration, we started with single-frame variation of morphological SMAA by Jimenez et al.  In its even most basic settings it showed definitely better-quality alternative to FXAA (at a bit higher cost, but thanks to much bigger computing power of next-gen consoles it stayed in almost same budget compared to FXAA on current-gen consoles). There was less noise and artifacts and much better morphological edge reconstruction , but obviously it wasn’t able do anything to reconstruct all this subpixel detail.
So the next step was to try to plug in temporal AA component. Couple hours of work and voila – we had much better AA. Just look at the following pictures.
Pretty amazing, huh? :)
Sure, but this was at first the result only for static image – and this is where your AA problems start (not end!).
Getting motion vectors right
Ok, so we had some subtle and we thought “precise” motion blur, so getting motion vectors to allow proper reprojection for moving objects should be easy?
Well, it wasn’t. We were doing it right for most of the objects and motion blur was ok – you can’t really notice lack of motion blur or slightly wrong motion blur on some specific objects. However for temporal AA you need to have them proper and pixel-perfect for all of your objects!
Other way you will get huge ghosting. If you try to mask out this objects and not apply temporal AA on them at all, you will get visible jittering and shaking from sub-pixel camera position changes.
Let me list all the problems with motion vectors we have faced and some comments of whether we solved them or not:
- Cloth and soft-body physical objects. From our physics simulation for cloth and soft bodies that was very fast and widely used in the game (characters, sails) we got full vertex information in world space. Object matrices were set to just identity. Therefore, such objects had zero motion vector (and only motion from camera was applied to them). We needed to extract such information from the engine and physics – fortunately it was relatively easy as it was used already for bounding box calculations. We fixed ghosting from moving soft body and cloth objects, but didn’t have motion vectors from the movement itself – we didn’t want to completely change the pipeline to GPU indirections and subtracting positions from two vertex buffers. It was ok-ish as they wouldn’t move very abruptly and we didn’t see artifacts from it.
- Some “custom” object types that had custom matrices and the fact we interpreted data incorrectly. Same situation as with cloth existed also for other dynamic objects. We got some custom motion vector debugging rendering mode working and fixing all those bugs was just matter of couple days in total.
- Ocean. It was not writing to the G-buffer. Instead of seeing motion vectors of ocean surface, we had proper information, but for ocean floor or “sky” behind it (when with very deep ocean there was no bottom surface at all). The fix there was to overwrite some G-buffer information like depth and motion-vectors. However, still we didn’t store previous frame simulation results and didn’t try to use them, so in theory you could see some ghosting on big and fast waves during storm. It wasn’t very big problem for us and no testers ever reported it.
- Procedurally moving vegetation. We had some vertex noise based artist-authored vegetation movement and again, difference between two frame vertex position values wasn’t calculated to produce proper motion vectors. This is single biggest visible artifact in game from temporal AA technique and we simply didn’t have the time to modify our material shader compiler / generator and couldn’t apply any significant data changes in patch (we improved AA in our first patch). Proper solution here would be to automatically replicate all the artist created shader code that calculates output local vertex position if it relies on any input data that changes between frames like “time” or closest character entity position (this one was used to simulate collision with vegetation), pass it through interpolators (perspective correction!), subtract it and have proper motion vectors. Artifacts like over blurred leaves are sometimes visible in the final game and I’m not very proud of it – although maybe it is usual programmer obsession. :)
- Objects being teleported on skinning. We had some checks for entities and meshes being teleported, but in some single and custom cases objects were teleported using skinning – it would be impractical to analyze whole skeleton looking for temporal discontinuities. We asked gameplay and animations programmers to mark them on such a frame and quickly fixed all the remaining bugs.
Problems with motion vector based rejection algorithm
Ok, we spend 1-2 weeks on fixing our motion vectors (and motion blur also got much better! :) ), but in the meanwhile realized that the approach proposed by Crytek and used in SMAA for motion rejection is definitely far from perfect. I would divide problems into two categories.
It was something we didn’t really expect, but temporal AA can break if menu pops up quickly, you pause the game, you exit to console dashboard (but game remains visible), camera teleports or some post-effect immediately kicks in. You will see some weird transition frame. We had to address each case separately – by disabling the jitter and frame combination on such frame. Add another week or two to your original plan of enabling temporal AA to find, test and fix all such issues…
Wrong rejection technique
This is my actual biggest problem with naive SMAA-like way of rejecting blending by comparing movement of objects.
First of all, we a had very hard time to adjust the “magic value” for the rejection threshold and 8-bit motion vectors didn’t help it. Objects were either ghosting or shaking.
Secondly, there were huge problems on for example ground and shadows – the shadow itself was ghosting – well, there is no motion vector for shadow or any other animated texture, right? :) It was the same with explosions, particles, slowly falling leaves (that we simulated as particle systems).
For both of those issues, we came up with simple workaround – we were not only comparing similarity of motion of objects, but on top of it added a threshold value – if object moved faster than around ~2 pixels per frame in current or previous frame, do not blend them at all! We found such value much easier to tweak and to work with. It solved the issue of shadows and visible ghosting.
We also increased motion blur to reduce any potential visible shaking.
Unfortunately, it didn’t do anything for transparent or animated texture changes over time, they were blended and over-blurred – but as a cool side effect we got free rain drops and rain ripples antialiasing and our art director preferred such soft, “dreamy” result. :)
Recently Tiago Souse in his Siggraph 2013 talk proposed to address this issue by changing metric to color-based and we will investigate it in the near future .
Temporal supersampling of different effects – SSAO
I wanted to mention another use of temporal supersampling that got into final game on the next-gen consoles and that I really liked. I got inspired by Matt Swoboda’s presentation  and mention of distributing AO calculation sampling patterns between multiple frames. For our SSAO we were having 3 different sampling patterns (spiral-based) that changed (rotated) every frame and we combined them just before blurring the SSAO results. This way we effectively increased number of samples 3 times, needed less blur and got much much better AO quality and performance for cost of storing just two additional history textures. :) Unfortunately I do not have screenshots to prove that and you have to take my word for it, but I will try to update my post later.
For rejection technique I was relying on a simple depth comparison – we do not really care about SSAO on geometric foreground object edges and depth discontinuities as by AO definition, there should be almost none. Only visible problem was when SSAO caster moved very fast along static SSAO receiver – there was visible trail lagging in time – but this situation was more artificial problem I have investigated, not a serious in-game problem/situation. Unlike the temporal antialiasing, putting this in game (after having proper motion vectors) and testing took under a day, there were no real problems, so I really recommend using such techniques – for SSAO, screen-space reflections and many more. :)
Temporal supersampling is a great technique that will increase final look and feel of your game a lot, but don’t expect that you can do it in just couple days. Don’t wait till the end of the project, “because it is only a post-effect, should be simple to add” – it is not! Take weeks or even months to put it in, have testers report all the problematic cases and then properly and iteratively fix all the issues. Have proper and optimal motion vectors, think how to write them for artist-authored materials, how to batch your objects in passes to avoid using extra MRT if you don’t need to write them (static objects and camera-only motion vector). Look at differences in quality between 16bit and 8bit motion vectors (or maybe R11G11B10 format and some other G-Buffer property in B channel?), test all the cases and simply take your time to do it all properly and early in production, while for example changing a bit skeleton calculation or caching vertex skinning information (having “vertex history”) is still an acceptable option. :)