Open miikkas opened 7 years ago
As a first step, it would be nice to measure FLIF lossy images with their pyscho-visual model Butteraugli.
The scripts to run lossy benchmark are here: https://github.com/FLIF-hub/benchmarks/tree/master/lossy
Just need to replace dssim
with butteraugli
.
I still have to see results demonstrating that Butteraugli and/or Guetzli are actually better than, say, DSSIM and MozJPEG. I'm not very impressed yet – both are aimed at very-high-quality encoding (Guetzli cannot do less than "quality 84"), so it's a bit hard to evaluate. The evaluation in https://arxiv.org/pdf/1703.04416.pdf is not very convincing to me – they beat libjpeg in human scoring 75% of the time, but MozJPEG probably does at least as good as that.
Getting smaller file sizes than other JPEG encoders when the criterion for comparison is based on the same metric they are using internally in the encoder, that's not very hard to do. The question is if that metric is actually any good, and I haven't seen much data on that.
See also: http://ericportis.com/etc/guetzli_test/
Google just announced Guetzli, a new JPEG encoder claiming 35% smaller file sizes. I think it would be prudent to add results for Guetzli-encoded images to the comparison charts. Also, could there be something in this algorithm worth trying out in the lossy FLIF compression?