Is it just me or do all the comparisons linked from the page show that webp actually looks better?
Still, the JPEG recompression feature might be what helps adoption.
Still, the JPEG recompression feature might be what helps adoption.
eg. https://eclipseo.github.io/image-comparison-web/#us-open&WEB... The player is missing the corner of her mouth in the JXL version.
JXL Medium (32KB) is about the same quality as WebP Small (19KB)
Note also that JXL is still being worked on. We have no information also about which encoder was used (I am assuming the reference encoder, since it is the only one that I know off right now) and which version.
Edit: the Citroën demo is also a clear win for JXL.
[0] https://eclipseo.github.io/image-comparison-web/#moscow&WEBP...
The singer’s left hand has wrinkles in the original image that disappear in WebP2.
Overall, WebP2 and especially AVIF are really good at very low bitrates (<1 bit per pixel), but unlike video, images on the Web will always be shown at the smallest bitrate necessary to be indistinguishable from the original; there, JXL tends to show all the details at a lower bitrate.
But it's also slightly worse than WepB Medium, e.g. with the corner of her mouth.
Most people do not want to have visible compression artefacts on the images they put on their web pages. JXL starts from this premise and tries to answer the question: "How small can we then make the image?"
Care must be taken when trying to compare the performance of image codecs by increasing compression density until there are very visible compression artefacts and then evaluating whether A's or B's artefacts look worse: If both A's and B's artefacts are so bad that one would not want to put such an image on one's website, such an experiment gives no insight on what one would pick for images that one would actually put on one's website.
Figuratively speaking, if I buy a shirt, my main criterion is that it looks good in good condition, and not that it still looks good if I put a coffee stain on it.
So, before comparing codec quality at compression levels where artefacts show, always ask yourself: At that level of visual quality, would I actually want to put either of the two options on my website? Now, it is of course tempting to compare "away from the actual operating point", because it is just so much easier to do comparisons if there are very visible differences. Comparing near-identical images for quality is hard. Doing this over and over again in a human rater experiment is exhausting. But that's then answering the actual performance questions that need to be answered.
Comparing artifacts at 0.2 bpp is tempting because the artifacts are big there. But it's like buying a car based on how it performs when you are using only the first gear.
https://cloudinary.com/blog/what_to_focus_on_in_image_compre...
Internet uses jpeg qualities that average around 2-3 bpp today, and an improvement in compression density through a new format that compresses 50 % better would push it down to 1-1.5 bpp. The comparison tool displays the bitrates when you hover on the images.
For all these examples, I was comparing at Small (as with Tiny they were both bad enough but in different ways that I often couldn't decide which was least bad). For the Abandoned Factory and Panthera Tigres, I think the extra detail of JXL looks better than the blurring of WEBP. On the otherhand, I think WEBP looks cleaner than JXL for Buenos Aires, Reykjavik, B-24 Bombers without loosing significant detail. And Avenches is a mix as JXL looks much better than WEBP for the trees and tile roof, but has worse chroma artifacts near the edges of the hat and clothing.
But that isn't the whole story, as for some of the images WEBP seems to both preserve more detail, and have fewer artifacts, such as Air Force Academy Chapel, Pont de Quebec, Steinway and White Dunes. What all these cases seem to have in common is a smooth gradient adjacent to sharp detail. WEBP seems to do a much better job of dealing with that boundary by blurring the smooth part, put preserving the sharp lines.
And if you bump up to WEBP2, the number of cases where it both preserves more detail and has fewer artifacts than JXL increases significantly.
In VP9/WebP this wasn't a "choice" so much as they were optimizing for good looking marketing graphs instead of pictures. You get blurry images if you target a metric like PSNR instead of actually looking at your output. x264 does have a few different tunings, the film one will try to turn detail to noise and the animation one won't.
Last week I submitted the 'ac strategy decision tree traversal' change to libjxl, which reduces ringing artefacts substantially, often by 60-70 %. It does a diligent job by choosing the best combination of small integral transforms (there are 10 different 8x8 transforms) instead of just satisfying with a large transform.
This change also improves on image quality, sharpness and truthfulness, it doesn't just blur the artefacts away.
This JPEG XL quality improvement has not yet landed in online comparison tools.