pex-gl
Reflection map prefiltering improvements
Current implementation suffers from two major issues: a) due to lack of multiple scattering metals get darker and they get rougher - solved by bumping reflection color based on roughness b) bright spots cause fireflies due to insufficient filtering - solved by sampling different mipmap levels
Both are described here https://bruop.github.io/ibl/
First implementation in a branch is promising.
Before (see dark metals on the right and very pale transition to plastic going down left column)
Getting blurry faster, and bright metals. Matches reference from Satin
The issue of sparkles persists
and could be fixed by adding the following (from here and here) to prefilter-from-oct-map-atlas.frag.js#LL93
if (NoL > 0.0) {
// Based off https://developer.nvidia.com/gpugems/GPUGems3/gpugems3_ch20.html
// Typically you'd have the following:
// float pdf = D_GGX(NoH, roughness) * NoH / (4.0 * VoH);
// but since V = N => VoH == NoH
float pdf = D_GGX(NoH, roughness) / 4.0 + 0.001;
// Solid angle of current sample -- bigger for less likely samples
float omegaS = 1.0 / (float(NUM_SAMPLES) * pdf);
// Solid angle of texel
float omegaP = 4.0 * PI / (6.0 * imgSize * imgSize);
// Mip level is determined by the ratio of our sample's solid angle to a texel's solid angle
float mipLevel = max(0.5 * log2(omegaS / omegaP), 0.0);
prefilteredColor += textureCubeLod(s_source, L, mipLevel).rgb * NoL;
totalWeight += NoL;
}
but.. the code is for cubemaps... which could be converted to octahedral maps if we find solid angle formula for them and improve mipmapping. So we either implement Improve reflection probes seams #321 or switch to cubemaps that can be seamless and mipmapped in webgl2.
but.. the code is for cubemaps... which could be converted to octahedral maps if we find solid angle formula for them and improve mipmapping. So we either implement Improve reflection probes seams #321 or switch to cubemaps that can be seamless and mipmapped in webgl2.
Would that impact #7 ?
Possibly as having prefiltered cubemap would allow us to convert it to SPH like e.g. BabylonJS does.
If we move to cubemap mipmaps then we could reconsider atlas packing to the following as it would be half the size.
