Godot 4.0 Alpha 10 Brings Temporal AA

[aufkrawall]

That's true, but I mean that they should have implemented FSR 2.0 in the first place, as they may have to ditch their TAA as it would become useless if they implement FSR 2.0, as it works on any hardware and it's just plain better than any TAA out there.

I thought so too after seeing it in Deathloop. However, a rendered typical scene in Deathloop is very static. Then I saw it in God of War and just had to acknowledge that FSR 2.0 has severe issues with pixelation, shimmering and ghosting when lots of stuff is changing on screen, which basically is always the case in this game. I think it's too flawed to make game devs dump their own TAA, AMD badly need to improve it.

[binarybanana]

Another benefit of TAA is that it can resolve small details that are otherwise impossible to capture without that temporal component (or rendering at higher resolution, brute force style). That means you can use a somewhat lower rendering resolution and accumulate into a higher resolution framebuffer. Most games with resolution scale settings will do this. That's why dynamic resolution scaling works as well as it does, at least for the grmes that do it well. What's often missing is a final sharpening pass at the end or it's too weak to deal with the typical TAA blurriness. Personally, I like to inject "Filmic Anamorphic Sharpen" which really brings out the fine details. But ghosting is and remains an issue, especially at sub-60 FPS.

[Calinou]

That's true, but I mean that they should have implemented FSR 2.0 in the first place, as they may have to ditch their TAA as it would become useless if they implement FSR 2.0, as it works on any hardware and it's just plain better than any TAA out there.

FSR 2.0 itself is more expensive than traditional TAA (at a given input resolution), and it currently doesn't support being used as a pure antialiasing solution (without upscaling). AMD plans to allow using FSR 2.0 as a pure antialiasing solution in the future, but it will still be more expensive than TAA when doing so. However, when used as a pure AA solution, I believe FSR 2.0 should be able to provide better antialiasing quality and less blurring compared to traditional TAA.

I think traditional TAA algorithms will remain relevant for a while thanks to their lower GPU cost. There are a lot of low-end GPUs that can only reasonably handle traditional TAA, unless you're willing to drop the rendering resolution a lot.

That means you can use a somewhat lower rendering resolution and accumulate into a higher resolution framebuffer. Most games with resolution scale settings will do this. That's why dynamic resolution scaling works as well as it does, at least for the grmes that do it well.

I'd like to look into adding dynamic resolution scaling to Godot 4.x, but we're trying to stabilize the renderer for 4.0's release right now.

What's often missing is a final sharpening pass at the end or it's too weak to deal with the typical TAA blurriness. Personally, I like to inject "Filmic Anamorphic Sharpen" which really brings out the fine details. But ghosting is and remains an issue, especially at sub-60 FPS.

Godot 3.x has contrast adaptive sharpening (CAS), while Godot 4.0 has sharpening in FSR (called RCAS – it can't be enabled without FSR though). I have a pull request that adds CAS to Godot 4.0, but we need to compare its output to using RCAS alone. AMD does not recommend using RCAS alone compared to traditional CAS though, but I don't know exactly why.

Another thing that can be done to improve sharpness is to use a negative mipmap LOD bias when TAA is enabled. This can also be done with FXAA to a lesser extent. FSR already uses a negative mipmap LOD bias depending on the quality preset chosen, with lower-quality presets using more negative LOD biases.

As for combating ghosting, I think it would be interesting to adjust the TAA disocclusion threshold depending on framerate. This will result in less effective antialiasing at lower framerates, but this may be preferable to ghosting. I'm not sure if this can be done without having to reinitialize TAA every time it's changed. For instance, disabling TAA below a certain framerate isn't ideal because it will take a while to become visible again once you re-enable it. (You could however increase its blending factor to 1.0 to make only the new frame displayed, while also disabling per-frame camera jitter.)

In racing games with a chase camera, having a dynamically generated TAA mask can also help reduce ghosting behind the car when it's moving. I'm not sure how this could be exposed in a general-purpose enough manner though.

[piotrj3]

That's true, but I mean that they should have implemented FSR 2.0 in the first place, as they may have to ditch their TAA as it would become useless if they implement FSR 2.0, as it works on any hardware and it's just plain better than any TAA out there.

FSR isn't better then average TAA. The worst case of temporal upscalling is ghosting behind a car racing and that one we didn't from FSR yet.

[Quackdoc]

i've been thinking about the possibility of using FSR2 as a way to subsidize SSAA, won't be able to test it in any way until some games come out with it, but it could be an interesting use case, could be really good, or really bad.

[aufkrawall]

i've been thinking about the possibility of using FSR2 as a way to subsidize SSAA, won't be able to test it in any way until some games come out with it, but it could be an interesting use case, could be really good, or really bad.

There is diminishing return with FSR 2.0/DLSS in Deathloop in non-upsampling mode (yes, you can do that in this game by setting dynamic fps target to a low value like 30 that you exceed all the time) vs. upsampling. Higher render resolution also doesn't change the characteristics/issues the methods have, only mitigate them. Future better iterations should look better at 66% resolution than current ones at 100% in case they manage to greatly reduce motion artifacts (already somewhat true with recent DLSS 2.4.3 vs. old 2.1.x versions).

[Jahimself]

MSAA for me has always been the most convincing Anti aliasing method. X2 for instance in terms of performance/visual quality is unmatched. X4 start to cost a lot but is kind of the sweet spot. If you climb up to x8 then the cost/visual improvement becomes not as interesting as resolution upscale. The only thing I miss on linux is the DSR from Nvdia (not DLDSR, real DSR). On their menu they have a smoothing option that goes along with DSR, adjusting it to 6% really is the sweet spot, and no upscale method whatever they are called, not even the one used in modern hitman (which is the best in-game implementation of upscaling) can rival in picture sharpness with Nvidia DSR+6% smoothing. At least for my aspect ratio 16/10 24.1" 1920x1200p =>3840x2400p with DSR.

[rabcor]

TAA is garbage, sure it reduces aliasing but it blurs everything out so much that it actually looks worse with TAA on than with no antialiasing at all. (FXAA is the same, possibly even worse).

[Quackdoc]

TAA is garbage, sure it reduces aliasing but it blurs everything out so much that it actually looks worse with TAA on than with no antialiasing at all. (FXAA is the same, possibly even worse).

I disagree, most times I prefer it. it's my favourite AA technique outside of SSAA, (but use SSAA on a 4k monitor... yeah ill pass). but as I said, a lot of AA can come down to both scene and preference.

[aufkrawall]

but as I said, a lot of AA can come down to both scene and preference.

Not really. You need samples, and you need also samples for not just geometry, but also shading/everything with modern rendering. You only get those with traditional SSAA or temporal SSAA. Traditional SSAA is crazy expensive and its sample distribution usually laughably inefficient. Which in practice only leaves temporal SSAA as a viable option. If TSSAA looks worse in lots of scenes than other offered methods, that particular TSSAA implementation likely is far from being done optimally.
DLSS also can look better than "native" because TAA/TSSAA solutions offered by the games are often just bad. TSSAA of 2017's Assassin's Creed Origins is still much better than that found in many more recent games, it's apparently a too difficult matter for many devs. Which is why it's most welcome that there are now competing TAAU solutions of Nvidia, AMD and soon (*cough*) Intel.