AI language models can exceed PNG and FLAC in lossless compression, says study | Ars Technica
AI language models can exceed PNG and FLAC in lossless compression, says study | Ars Technica
Is compression equivalent to general intelligence? DeepMind digs up more potential clues.
PNG compression is an absolute joke so that's not worth anything - JPEG-XL and WebP Lossless beat PNG by a longshot. PNG also requires intense optimization to be truly compressed, and I'm not sure to what extent they tried to optimize their PNG results. As an example, the lossless comparison chart I uploaded was 380KB initially, and I was able to optimize it to 280KB using OxiPNG. The PNGs in the lossless comparison chart were properly optimized.
Beating FLAC is interesting, but FLAC compression hasn't really been updated in a long time TMK - I wonder if there's traditional gains left on the table with modern compression techniques?
A copy of the original paper can be found here.
Yeah, PNG is overly complex, slow to encode/decode, and despite all that, it atill doesn't compress the picture very well
I think i heard that one of the creators of the format said he didn't have any experience with compression, and they more or less just threw things at the wall to see what stuck
If you want the mediocre compression level of PNG, but waaaay faster, i can recommend QOI
QOI should actually be worse than JPEG-XL in both speed and size, if you turn JPEG-XL's effort level down to a speed-oriented level like
-e 1. I wrote a quick script to compare them, and this is the output from 10 random (fully-optimized) PNGs I had: https://pastebin.com/raw/nSkBSePBI have a lot of cores on my CPU so you probably want to be looking at the
usertime metric, but consider that parallelization is a strong and intentional benefit to JPEG-XL.The results are not exactly scientific since I only picked 10 PNGs, but it shows a general trend.
