Very nice article. BUT it's as if there's a couple of paragraphs that have deleted in the middle. They say absolutely nothing about HOW they are doing perceptual image compression. Maybe they're using SSIM (structural similarity) values, which is standard in image compression literature? In fact they don't even explicitly say they are doing perceptible image compression, but I assume that's the whole point of the article.<p>JPEG appears to be too aggressive at compressing areas of an image with a low imformation content. Look at the first 4 examples: the people and other detailed ojects look fine in all four images, but there's a huge difference in the near-flat red regions, because too many bits have been cut out. Really, it's the JPEG compressor that needs tuning rather than the q-level or other settings.<p>That might be the reason that some users loudly complained about the more aggressive compression when their testing showed that on average people could only notice it 3% of the time. Probably that means the compression became noticeable for a few images or parts of images.