UPDATE: The author has commented elsewhere that his 45KB images weren't 580x363, but 1920x1200. I've added full resolution 1920x1200 examples.
The Real world analysis of google’s webp versus jpg over at English Hard is getting a lot of attention today, comparing Google's new WebP image compression to JPEG compression, using a few "real world" images.
While the post presents the image compression quality visual results in PNG (with accompanying numeric error level analysis) for fairness, the JPEG compression images' visual presentation is seriously flawed. Setting aside that a numeric noise ratio is known to be a poor metric for human perception of image quality, what you see in this post is even worse.
This is what the post shows for a 44.6KB "webp" image:
This is what the post shows for the 44.5KB "jpg" image:
I have never seen banding like this produced by a JPEG compressor, no matter how bad the compressor. Was this really true?
I used a 1920x1200 PNG image source, opened it in Photoshop CS5, chose "Save for Web and Devices", resized to 580 x 363, and set a quality of 75. The resulting file is 43.6 KB:
This doesn't show the dramatic concentric rings of banding. It certainly looks nothing like the alleged JPEG shown in the blog. To check if the conversion to PNG would damage the JPEG as shown in the blog, I opened this JPEG in Mac's native Preview app and saved as PNG:
I'm not the only one who noticed this issue. Several others called out the problem. The author hasn't yet fixed his post, saying essentially "you're not looking at the big files I'm looking at".
Yesterday, a review by x264's lead developer, H.264 and VP8 for still image coding: WebP?, concluded, "You’d have to be nuts to try to replace JPEG with this blurry mess as-is." I don't know if the webp images in this blog post show a blurry mess, but who can say if their representation is any fairer than the badly broken JPEG results?
It seems to me that when presenting a visual comparison of image compression, the visual itself really should be representative.