This is cool code but can someone explain why we'd need this? Do we not have a common lossy format that supports transparency? My primary concern is that people have an expectation of what PNG means and this completely subverts that.
We actually don't, WebP can do this, but it's by no means common yet. This is actually a reasonable use of PNG still, it's optimizing the contents for the lossless algorithm. You might be able to take the difference/error and put it in another image and layer them to make it give back the detail.
PNG isn't being subverted. The lossage happens before PNG is applied.
What you are getting is a lossless copy of an image that has been cleverly preprocessed to be easier for the PNG algorithm to compress while not losing noticeable visual quality. The result is good compression and transparency support (which JPEG doesn't have).
Yeah I only asked because I agree with the above poster that we can't consider WebP common yet but was unsure if I was forgetting something. I can see the utility in this then, at least as a stopgap.
I've researched lossy GIF too: http://pornel.net/lossygif but even lossy version wasn't better than PNG.
LZW is just a very poor compression. Even best case is taking ridiculous amount of bits (you can only add 1 byte to previously used pattern, so it's sequence of symbols for 1+2+3+5+6+etc. pixels resetting every 4k iterations), so there isn't even room for improvement.
