SSRNode: Improve quality with TAA.

[Mugen87]

Description

Combining SSR with TAA currently exhibits noticeable flickering on SSR reflections. I've prepared a live example based on the SSR demo but with a special material setup. The flickering is more subtle on very rough surface so the materials now use a zero roughness and a metalness value of 1. You can easily see it on areas with specular highlights.

https://rawcdn.githack.com/Mugen87/three.js/c5c8b69023091cbdc9cb80bf97dff54070fda817/examples/webgpu_postprocessing_ssr.html

I've already added a simple experimental temporal filtering in this example which reduces SSR banding artifacts but unfortunately it does not mitigate the flickering. The idea of Mugen87@c5c8b69 is to vary the step sizes based on some sort of noise similar to other raymarching techniques.

I've read that some SSR implementations undo the TAA jitter to perform SSR in a stable coordinate space. However, I'm not sure yet how this is supposed to work. Any insights and recommendations are welcome^^.

/ping @zalo

[zalo]

The SSGI stabilizes by repeating the same jitter pattern every 6 frames; perhaps there’s some non-repeating jitter (or non-integer synchronized jitter) in the SSR that causes it to repeat across more than 6 frames?

[Mugen87]

In the current SSR implementation (not the one from the above commit), there is no jitter applied as well as no changing noise/randomness over time. When combining SSR with TAA, the only jitter comes from the view-offset encoded in the projection matrix. This jitter affects not just the beauty (the main scene pass) but depth and normal buffers as well in order to keep the MRT data in sync for each frame.

The problem is now that SSR seems to react very sensitive on this jitter. Especially high-frequency, specular areas can rapidly change in SSR reflections and the TAA is not able to stabilize the image. I'm still studying the related work but it seems one common approach in the industry is to compute SSR in a stable space. Unfortunately, undoing the jitter in the projection and inverse projection matrix isn't sufficient. Since we want to reuse the depth and normal buffers computed with the beauty pass, the sampled data (viewZ, view position and view normal) must be stabilized as well. I'm not sure how to do this, yet.