WebP Quality Settings

November 26th, 2014
tech
One of the questions that came out of the UA-parsing discussion was whether pagespeed compresses WebPs too aggressively. In general PageSpeed tries to compress images to the lowest quality that will still look good enough, because then they'll use as little bandwidth as possible and load faster. By default its sense of "good enough quality" for JPEG is 85, and for WebP it's 80, [1] but are these equivalent?

For a test set I took the 668 full size JPEGs under jefftk.com/pictures that had a quality of at least 85. [2] I compressed them all as both JPEG and WebP at qualities 45, 50, 55, 60, 65, 70, 75, 80, and 85. To measure the image quality I used SSIM with SSIM.cpp. Here's a graph of quality against SSIM for WebP and JPEG, higher SSIM being better:

It looks like WebP at 80 is equivalent to JPEG at ~67, not JPEG at 85. Maybe there's something suspect with this SSIM implementation? Let's test with another tool, and be sure. This one is DSSIM, not SSIM, which means the units will be different and now lower numbers are better, but images with identical SSIM will also have identical DSSIM:

This is a very similar result, and these results suprise me. It sounds we do need to adjust our default quality settings. Luckily, it turns out webp now has a -jpeg_like flag:

-jpeg_like
   Change the internal parameter mapping to better
   match the expected size of JPEG compression. This
   flag will generally produce an output file of
   similar size to its JPEG equivalent (for the same
   -q setting), but with less visual distortion.
This doesn't sound like exactly what we want; we want to match visual distortion instead of size, but let's try it:

That works well enough that we sometimes can't see the JPEG dots because the WebP-jpeglike dots are immediately above them. Which is surprising: did the people calculating the compression curve use SSIM instead of file size to calibrate it?


[1] These can be configured with JpegRecompressionQuality and WebpRecompressionQuality Recompression Quality in the docs.

[2] This isn't a great test set for something that's compressing all sorts of images found all around the web, but it was already on my machine. When I get back from Thanksgiving I'll probably re-run this with a more representative set.

Recent posts on blogs I like:

Pay For Fiction

Against piracy

via Thing of Things February 29, 2024

When Nurses Lie to You

When the nurse comes to give you the flu shot, they say it won't hurt at all, right? And you trust them. Then they give you the shot, and it hurts! They lied to you. A lot of nurses lie to children about shots and blood draws. Part of it is they probabl…

via Lily Wise's Blog Posts February 28, 2024

How I build and run behavioral interviews

This is an adaptation of an internal doc I wrote for Wave. I used to think that behavioral interviews were basically useless, because it was too easy for candidates to bullshit them and too hard for me to tell what was a good answer. I’d end up grading eve…

via benkuhn.net February 25, 2024

more     (via openring)