5 ms·

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 ad

by Paul_S 5y ago

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.

exclipy 5y ago
By golly, you're right. eg. https://eclipseo.github.io/image-comparison-web/#us-open&WEBP=m&JXL=m&subset1 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)
- torgard 5y ago
  I'd say JXL Medium is slightly better than WebP Small - compare e.g. the racket, face, and left hand. But it's also slightly worse than WepB Medium, e.g. with the corner of her mouth.
  - LispWizard 5y ago
    JXL's design operating point is "no visible compression artefacts". 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.
    - jonsneyers 5y ago
      ^^^ This. 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.
    - torgard 5y ago
      Great point!
    - pavon 5y ago
      This blog, linked from the original article goes into this idea in detail, describing how JXL was focused on improving compression of high fidelity images, and has not focused as much on the appeal of highly compressed images. https://cloudinary.com/blog/what_to_focus_on_in_image_compression_fidelity_or_appeal https://cloudinary.com/blog/what_to_focus_on_in_image_compre...
- maeln 5y ago
  Although the Moscow demo[0] look better (in my eyes) in jxl than wepb. 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=m&JXL=m&subset1 https://eclipseo.github.io/image-comparison-web/#moscow&WEBP...
  - espadrine 5y ago
    One case where JXL is very noticeably more accurate is here: https://storage.googleapis.com/demos.webmproject.org/webp/cmp/2021_06_08/index.html#meg-20110701-japan-expo-02*3:1&WEBP2=m&JXL=m&subset1 https://storage.googleapis.com/demos.webmproject.org/webp/cm... 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.
JyrkiAlakuijala 5y ago
You can get a more up-to-date comparison from the WebP team's comparison page: https://storage.googleapis.com/demos.webmproject.org/webp/cmp/2021_06_08/index.html#us-open-tennis-2010-1st-round-046&WEBP=m&JXL=m&subset1 https://storage.googleapis.com/demos.webmproject.org/webp/cm... 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.
pavon 5y ago
Yeah, I agree that webp looks better for many of the images. For most of them it comes back to the subjective issue that webp (and VP9) choose to err on the side of blurring detail they can't encode, while jpeg XL (and x264, etc) try to keep all the detail they can at the expense of more artifacts. There are very vocal proponents of both approaches, and personally I think it varies by the image content. 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.
- astrange 5y ago
  > Yeah, I agree that webp looks better for many of the images. For most of them it comes back to the subjective issue that webp (and VP9) choose to err on the side of blurring detail they can't encode, while jpeg XL (and x264, etc) try to keep all the detail they can at the expense of more artifacts. There are very vocal proponents of both approaches, and personally I think it varies by the image content. 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.
- JyrkiAlakuijala 5y ago
  The amount of artefacts should be substantially reduced in JPEG XL in the next round of comparisons. 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.

by Paul_S 5y ago

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.

exclipy 5y ago

By golly, you're right. eg. https://eclipseo.github.io/image-comparison-web/#us-open&WEBP=m&JXL=m&subset1 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)

torgard 5y ago
I'd say JXL Medium is slightly better than WebP Small - compare e.g. the racket, face, and left hand. But it's also slightly worse than WepB Medium, e.g. with the corner of her mouth.
- LispWizard 5y ago
  JXL's design operating point is "no visible compression artefacts". 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.
  - jonsneyers 5y ago
    ^^^ This. 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.
  - torgard 5y ago
    Great point!
  - pavon 5y ago
    This blog, linked from the original article goes into this idea in detail, describing how JXL was focused on improving compression of high fidelity images, and has not focused as much on the appeal of highly compressed images. https://cloudinary.com/blog/what_to_focus_on_in_image_compression_fidelity_or_appeal https://cloudinary.com/blog/what_to_focus_on_in_image_compre...
maeln 5y ago
Although the Moscow demo[0] look better (in my eyes) in jxl than wepb. 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=m&JXL=m&subset1 https://eclipseo.github.io/image-comparison-web/#moscow&WEBP...
- espadrine 5y ago
  One case where JXL is very noticeably more accurate is here: https://storage.googleapis.com/demos.webmproject.org/webp/cmp/2021_06_08/index.html#meg-20110701-japan-expo-02*3:1&WEBP2=m&JXL=m&subset1 https://storage.googleapis.com/demos.webmproject.org/webp/cm... 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.

JyrkiAlakuijala 5y ago

You can get a more up-to-date comparison from the WebP team's comparison page: https://storage.googleapis.com/demos.webmproject.org/webp/cmp/2021_06_08/index.html#us-open-tennis-2010-1st-round-046&WEBP=m&JXL=m&subset1 https://storage.googleapis.com/demos.webmproject.org/webp/cm... 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.

pavon 5y ago

Yeah, I agree that webp looks better for many of the images. For most of them it comes back to the subjective issue that webp (and VP9) choose to err on the side of blurring detail they can't encode, while jpeg XL (and x264, etc) try to keep all the detail they can at the expense of more artifacts. There are very vocal proponents of both approaches, and personally I think it varies by the image content. 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.

astrange 5y ago
> Yeah, I agree that webp looks better for many of the images. For most of them it comes back to the subjective issue that webp (and VP9) choose to err on the side of blurring detail they can't encode, while jpeg XL (and x264, etc) try to keep all the detail they can at the expense of more artifacts. There are very vocal proponents of both approaches, and personally I think it varies by the image content. 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.
JyrkiAlakuijala 5y ago
The amount of artefacts should be substantially reduced in JPEG XL in the next round of comparisons. 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.