The sample count map for 0.04 error threshold and 4-256 samples.

It seems to get the most out of adaptive sampling, we want our sample limit high enough that we rarely hit it, but not too high to be unreasonably slow. We also want our error threshold to be balanced such that we arent spending way too many samples on spots that dont matter but spots that need higher samples can spend them.Eyeballing it, I rendered this in 18 seconds with 4-256 samples and an error threshold of 0.04. It looks much better than the supposed 0.01 error threshold and rendered faster as well. Id argue it even looks better than the 0.005 error threshold image and is WAY faster than that one.

0.005 error threshold. Definitely hitting our 64 sample ceiling more frequently.

Moving to 0.005 error threshold, took 35 seconds. Seems we are only marginally better but our times are only marginally worse!

2-64 samples, 0.01 error threshold

Adaptive sampling, 2-64 samples at 0.01 error threshold: 23sec.

For fun: 100 samples took 2 min 31 seconds. Not much visual improvement over the 10 samples. Adaptive sampling incoming.

10 samples, took 15 seconds. Quite nice, very little aliasing.

5 random samples, took 8 seconds. HUGE improvement to aliasing.

For reference, this is the single sampled original.