Sampling 0.5 + 0.5 * fade * sin( 2 pi a e^(5a) )

I first show uniform sampling of the image function, for comparison. Then I show jittered, regular, multijittered, and two types of adaptive sampling. I discuss the techniques at the bottom of the page.

Note: no gamma-correction has been done on these images.

Sampling Type 1 sample/pixel, hat filter 4 samples/pixel, hat filter 16 samples/pixel, hat filter 16 samples/pixel, tent filter 16 samples/pixel, circle filter 256 samples/pixel, hat filter 256 samples/pixel, tent filter 256 samples/pixel, circle filter

Uniform

Jittered

Regular

Multi-Jittered
Sampling Type Depth = 2, hat filter Depth = 4, hat filter Depth = 9, hat filter Depth = 9, tent filter Depth = 9, circle filter

Adaptive 1

Adaptive 2

Discussion

Uniform sampling is, as expected, terrible. I included it to show how much the other methods improve over it. Jittered isn't too bad, but has some annoying high-frequency noise except at high pixel resolution. Regular does a decent job for this image, but tends to smear out the high-frequency parts. Multi-jittered also has high-frequency noise.

The first adaptive method samples each pixel at the corners and subdivides the pixel if something interesting is going on in that area, i.e. the corners have different values. It produces some very strange artifacts and takes a long time to run. The second adaptive method jitters the corners and samples at the jittered points, then subdivides. It also takes a long time to run, but produces much better images than the first method. It has high-frequency noise but none of the artifacts of the first method.