> (Spoilers: it also allocates 60+ GB of RAM for intermediate results, so consider reducing the image size before running it on your own machine)

The input algorithm is essentially written in SSA form, and it's easy to analyze when each "register" stops being used; then you can drop the arrays after their last use. Turns out the number of live registers never exceeds 160 at any point in time, and with this one addition the max memory usage drops to barely above 1GB.