Protip: Do not use progressive JPEGs to save disk space or network transmission speed. If you want progressive loading, ok, but otherwise, progressive jpegs require disproportionate amounts of CPU effort to decode vs. the savings. E.g. on mobile you're paying with power.
When I measured, filtering was about the same cost as inflate. I could pipeline them to hide the latency. It’s true that inflate was usually the bottleneck in that case, but it still saves power to not have to defilter.