A very low tech approach would be to create a video from the screenshots, and let the video codecs do this for you. Av1/x265 are quite mature in detecting differences between frames.
I agree that's not a bad starting point, but it should be possible to do a lot better than this because a screenshot will often be very similar to an earlier one that's not the previous one. For example, repeatedly switching between a few views. Video codecs are optimized for (a) streams that rarely go back to an earlier pattern, (b) minimal ongoing memory usage, and (c) realtime decompression, so they aren't going for quite the same use case.
Even h264, which is way over a decade old, can reference frames that are further back than the last one. And it supports lossless. Now the only thing to research is if those two features can be used together and how many frames back it can reference.
