This is even more interesting (full disclosure - I vibecoded a lib for it lately - lowzag) when you take into consideration working set restrictions. In almost

This is even more interesting (full disclosure - I vibecoded a lib for it lately - lowzag) when you take into consideration working set restrictions. In almost every case, downscaling-while-decoding is going to beat decode-then-scale on performance - merely keeping the uncompressed image in RAM is going to cost you, when you can scale it down in L1 or even vector registers.