To combat temporal AA blur, my idea was to use a higher resolution history buffer, but then I saw in Brian Karis’s slides that they’ve already considered this and it was too expensive (slide titled Blurring: reprojection diffusion; https://de45xmedrsdbp.cloudfront.net/Resources/files/TemporalAA_small-59732822.pdf ). Aside from the resolution, a lot of the added the expense seems to come because you can no longer do everything in place. Instead of reading the history into the main render target, you have to do an upscale step, apply the TAA, replace the history (cheap swap?) then downscale back into the render target.
So, because I’m wanting this for VR, an idea i’m thinking of now to only do this for the green channel (OLED pentile already downsamples red and blue at the display anyway, so TAA blur doesn’t matter as much in those channels), and only do it for a fairly small inset rectangle with the result feathered into place to avoid any sharp cutoff (VR lens distortion already downsamples the periphery, so again, TAA blur doesn’t matter as much there).
With only one color channel and using a smaller inset center region this additional history buffer should be a good bit smaller than the existing one even at double resolution.
The basic approach would be:
Crop and upscale the green channel (with nearest sampling/pixel doubling?) of the center region of the main render target off to another buffer, then apply TAA using the higher-res green history buffer (applying TAA uses bilinear or bicubic sampling, this is where the higher resolution actually should help things and result in less blur as the accumulated history starts having more subpixel accuracy, at least for the static scene (moving stuff is I think projected in using a half res velocity buffer)). Next apply the normal TAA process to the main render target and then finally feather in and downscale the high res replacement buffer into the green channel.
The main thing I’m concerned with is that TAA uses YCOCG filtering for deghosting in a bunch of ways that I don’t really understand, could it even continue to work in just the green channel? Could I make the normal TAA pass keep using that heuristic and just apply the work to the higher res green buffers simultaneously with it’s own work? Would it destroy the benefits?
Any thoughts? Is this workable? Terrible idea?