I've been experimenting with using bucket rendering instead of progressive rendering, and it's giving much better results for super-noisy things like depth of field. Statistically speaking, adaptive sampling (which you get when using buckets) should produce the same total error (noise) in an image, but with a more uniform spread. Pixels that are above the noise threshold will get sampled more, at the expense of the pixels that fall below the noise threshold. so by using bucket rendering in Corona, your less noisy areas will become more noisy, but your more noisy areas will clean up much faster.
Basically, using buckets produces a more uniform noise profile.
Also, cleaning up extremely heavy depth of field like that is basically not doable using brute force monte carlo rendering. The number of samples required is astronomical. You're better off filtering some of the noise in post, rather than trying to get perfect results in the render.
