Open gkjohnson opened 2 years ago
If this is related to RGBA floating point packing, I think I have an idea how to solve this floating point precision issue.
The issue is floating point precision and error resulting in artifacts like banding in gradients when performing math operations.
This three.js SSAA demo demonstrates the same kind of issue but with a fixed number of samples. The difference is that the pathtracer is performing a continuous average instead of fixed sample count.
I think Kahan summation can help with this. The thought came to me the other day but is also noted here.
https://twitter.com/BenHouston3D/status/1521962488807444481?t=gOGmOLpXPF55D4ihTdc9Ig&s=19
See banding in BG gradients