andrews05

Thanks. Sadly though, the new palette sorting did not improve this image.

Following up here on behalf of @adamreisnz: Thanks @manast, turned out that was indeed the problem so should be easy for us to resolve. The issue can be closed.

Thinking about this further, I realise this should probably be on the caller to wrap it if appropriate and not zopfli's responsibility at all. Also, I found the BufReader incurs...

I don't think it's actually about buffering. I just happened to try the BufReader initially but wrapping the &[u8] in *any* Read implementer seems to resolve the issue, such as...

Ah, the copy change sounds like a likely suspect. For reference, the greatest difference I've observed so far was 0.8% on a ~350k (compressed size) png.

While I don't know anything myself, I have noticed libdeflate's code is super well documented - could be helpful to read over it. I think it borrows a lot of...

Yeah, I'm not too worried about performance really, zopfli is always going to be horrendously slow 🤣. I just came across crc32fast and thought it might be useful here. Although,...

Nice! I measured this around 3-4% faster on one file I tested. > When dealing with near-optimal algorithms like Zopfli, one must keep the [no free lunch theorem](https://en.wikipedia.org/wiki/No_free_lunch_in_search_and_optimization) in mind....

Yeah, I totally understand. Wish I knew more about compression myself 🙂

> Another issue is present with XYB: the colors appear significantly different between the default 4:2:0 and the manually defined `chroma_subsampling=420`. My guess is that it’s using YUV subsampling rather...