For every pixel-group, the denoiser now keeps the match-count-throttle
matches with the lowest SAD, and flood-fills each until the resulting
region is at least match-size-throttle pixel-groups big, then applies
that region right then and there. This new algorithm is a lot simpler
than the previous one, but seems to work a lot better.
Also, a nasty bug was fixed in the way the extent of applied regions
are removed from the search-window cells. VERY nasty bug. It
technically produced broken data.
Image quality has improved DRAMATICALLY. Performance has increased from
anywhere between 20% (for normal video) to 35% (for pathological video).