I'm not so sure about it, but lossy compression algorithms are usually "non deterministic" and what I mean by that is that there is a target format to encode to, but you have some variation in how you're turning the original data into that

So there's a lot of room in optimising the process