Re: Halton

I don’t remember what bias I used exactly. I might have used -1.5 or even -2.0. Quality wise was somewhere between the OGSS/RGSS and Ground Truth. And it was doing a straight average of the samples for the resolve.

And while there was certainly some slightly better temporal stability, the sharpness wasn’t significantly better likely due to only using a box filter. And the performance difference was very significant on the PS4 & Quest, enough to make the option infeasible. Even on PC it was a big enough performance hit for me to drop testing of it immediately. So I didn’t really get too deep into it.

For the same reason I didn’t try any other higher tap sample patterns. Though the 3, 2 Halton I used AFAIK is a MSAA 8x pattern used on some hardware, and is part of why I used it.

In the real world use cases I’m also far less strict about the bias I use, often tweaking it per material depending on the font / icon. Most of the time I think I used values between -0.75 ~ -0.85 with the RGSS as it was a good mid-way between sharp and aliasing. Much better than doing nothing, but avoiding some aliasing that can show at -1.0. I know some people have used a bias of -1.5 with the RGSS shader because they liked the extra sharpness!

Re: Cubic filtering

You can very closely approximate cubic filtering with 4 taps. But that really only works for magnification. As soon as you start to fade to another mip level, the cubic sample positions are no longer valid as the approximation is leveraging the linear filtering within a single mip level, meaning anisotropic and trilinear filtering would need to be disabled. Or you’d have to do cubic filtering on two mip levels and do the blend between them manually. This would mean 8 samples minimum, which as noted above was way too slow. So no, I did not pursue cubic filtering at all.