I did some reading of the PNG specification and it may be possible to take advantage of Adam7 interlacing (8 passes total) and scanline filtering (several methods) to write out an image where the scanlines of 'known to be blurry' areas contain only 1 pass of image data and 7 passes of highly-compressible scanlines with filters that generate blur at decode-time.
Doing this formally at scale would require the compositor and the encoder to cooperate, as the encoder would benefit greatly from having access to both the 'blurred' and 'unblurred' areas without the blur filter having been applied to the former, as it could then construct low-fidelity, low-bandwidth, visually pleasing blur for the 'blurred' segments.
This exceeds my ability to write PNGs by hand and it certainly exceeds the bounds of what most people think 'an encoder' should be capable of doing, but at least it presents a path forward. I'll post to HN someday if I ever somehow manage to do this.
Doing this formally at scale would require the compositor and the encoder to cooperate, as the encoder would benefit greatly from having access to both the 'blurred' and 'unblurred' areas without the blur filter having been applied to the former, as it could then construct low-fidelity, low-bandwidth, visually pleasing blur for the 'blurred' segments.
This exceeds my ability to write PNGs by hand and it certainly exceeds the bounds of what most people think 'an encoder' should be capable of doing, but at least it presents a path forward. I'll post to HN someday if I ever somehow manage to do this.