Lossy samples where others are better

[gray copper]

AVIF has better subjective quality after recompressing JPEG-like artifacts at q=40; effort=10.

[gray copper]

AVIF is the best with q=40; effort=10

[native hawk]

If I may ask, what did you have in mind for the images?
Thumbnails, small previews, mobile viewing?

[gray copper]

No restrictions. Just comment your notes and conclusions (if any).

[gray copper]

AVIF has the lowest filesize at q=40; effort=10

[scenic latch]

why wouldn't you use lossless for that kind of image?

or better, svg with gzip network compression

[gray copper]

Lossy is good enough if it's not JPEG. JPEG has bad subjective quality and artifacts are easy to trigger.

Maybe we don't need lossless if image is popular on the web. We want to save bytes if some image is popular...

Pretty much all other modern formats perform much better than JPEG.

[native hawk]

Do you still have the original PNG from that test? I'd like to try some things with it too

Unless you were compressing that exact image with the jpeg compression

[gray copper]

No. I randomly stumbled upon it from a blog post. I'm not OP here...

[gray copper]

AVIF has the best subjective quality and the smallest filesize at q=40; effort=10. It's quite crazy how AVIF almost doesn't have artifacts...

[gray copper]

AVIF has the best subjective quality and the smallest filesize at q=40; effort=10.

[gray copper]

Both AVIF and WebP 2 are suboptimal in terms of visual quality at q=40; effort=10.

[native hawk]

Is every single image here re-compressing jpegs?

[gray copper]

AVIF has the best subjective quality and the smallest filesize at q=40; effort=10.

WebP 2 has the best subjective quality and 3rd filesize without clear visual downsides at q=40; effort=10. JPEG XL has minor problems with colors of a tree close to the protagonist. AVIF has problems with colors and artifacts around a tree in background.

[gray copper]

AVIF has the best subjective inscription text quality at q=40; effort=10 but also changes colors... JPEG XL and WebP 2 have problems with inscription text...

[gray copper]

AVIF has the smallest file size at q=40; effort=10. WebP 2 is second in terms of file size. JPEG XL is 33–40% bigger than AVIF and WebP 2.

[keen nymph]

Didn't expect to ever see that being used as test image

[faint lynx]

What do you do to calibrate quality to make statements like "33-40% bigger"?

[gray copper]

Good enough quality at 100% zoom (no pixel hunting). q=40 is just easy to remember for humans. q=40 and lower levels are typically ignored because JPEG was bad IMO. We shouldn't just ignore them with newer codecs. I'm choosing realistic images and illustrations to show this.

[native hawk]

q changes depending on the encoder, and the format, and the software... It's not really usable to just say "40" for both AVIF and JXL

[faint lynx]

"q=40" means wildly different things in different encoders, and even from image to image. Encoder quality scales are completely arbitrary and have no common interpretation in terms of actual visual quality.

[bright monolith]

Show us side by side comparisons of what it actually looks like + the difference in filesize

[hollow oar]

I'm aware of only one professional website that uses q=40 quality grade images and even them don't use it across the board -- it can be considered that 60-70 is the minimum quality

[scenic latch]

someone sent me a bunch of jpegs at 100 quality

[hollow oar]

I think MIT or MIT press website had 100 quality jpegs when I was looking at web quality in the past...

it can happen to the best of us 😛

[keen nymph]

At least it beats low quality jpg converted to png

[scenic latch]

oh god

[keen nymph]

Guess I'll just blame W3C for that

[gray copper]

Some faces are slightly better with AVIF at q=40; effort=10... AVIF has minor problems with the colors of text... AVIF has the smallest file size; 60 KiB versus 80 KiB with JPEG XL.

[scenic latch]

is that the source image?

[gray copper]

Yet another difficult sample at q=40; effort=10. AVIF, JPEG XL, and WebP 2 are suboptimal.

[faint lynx]

q40 is going to have artifacts. The useful range is q60-q98

[gray copper]

This is true for lower resolutions. This isn't true for Full HD and bigger. I don't like that users have to keep this in their heads. All codecs are problematic...

[native hawk]

Quality is the similarity to the original, having a bigger image doesn't suddenly make the problems go away, only blur them when you can't view them at 1:1 zoom

[faint lynx]

Codecs treat pixels the same way regardless of how large the image is, and I think that's the only sensible thing to do. You cannot just assume that large images will be downscaled, viewed from a larger distance, or on a denser display, so you can get away with more aggressive compression.

[bright monolith]

If a competing encoder produces high quality usable output at a certain bitrate, and you need -q40 to reach the same bitrate in libjxl, then it's a fair comparison... as long as the competitor you're comparing it to actually looks really good.

[gray copper]

Full HD monitors are majority. 1440 and 4K are way less frequent even in 2024... We can assume Full HD monitors for most users.

I don't remember how many megapixels Full HD has but medium smartphone cameras can get 12+ MP in 2024. This wasn't the case in the 2000s.

It's safe to assume the rescaling of 2-12 MP images for Full HD for most casual users in 2024.

Lower quality settings make sense for photos on the Internet... We should study how non-photo content behaves with previously ignored settings.

General recommendations for any resolution, usage, and any content aren't the best...

12 MP images with q=0.9 are definitely an overkill for Full HD monitors. This is true with the most popular codecs...

I don't think that I can cite any paper with similar statements... Few or nobody cares to factor everything... It's not that hard, really...

[scenic latch]

🧑‍🔬

[native hawk]

I feel I should point out, we have real world cases where using resampling is better than lowering quality. The same happens for bitstarved video content. High quality 720p is better than degraded and distracting 1080p

[bright monolith]

It would be nice if it always looked better than encoding at a lower resolution (and resizing to the same size to compare)

[bright monolith]

I wonder if I should revive this thread

https://afontenot.github.io/image-formats-comparison
I took a look at this comparison recently and I think some of these test images could help Jyrki tune things. In particular, the images where JXL does much better or much worse than the competition.

It's using libjxl v0.10.2 so it's not that old.

Obviously the most noticeable and obvious thing from a distance is the severely washed out colors in libjxl compared to other codecs, which I've mentioned before. The next most noticeable thing is the tendency to punish darker regions of the image especially hard compared to brighter regions.

Lastly, there is sometimes a lot of ringing artifacts compared to other codecs. Aside from that, it is extremely impressive and something needs to be said about how incredible libjxl performs at certain images compared to other codecs!

I have been looking at the "medium" bitrate image comparisons and I would especially like to mention, for example, "Sunset"

"Seikima" is another example of an impressive, overwhelming win for libjxl

"Seikima" and "Sunset" are overwhelming victories and very impressive. In fact they are so overwhelmingly excellent they really ought to be featured in press materials like on the jxl fan site

Double down on your strengths as they say

The most noticeable and most critical error is the severe desaturation that is noticeable even from a distance. Luckily it should not be that complicated to fix. Most noticeable in the paint in "Ohashi" and the flowers in "Abandoned factory" but also on the green dress in "Japan Expo"

An especially severe example of the aggressive punishment of dark shaded image regions can be seen in the shadowy struts of "Pont de Qebec"

Compare with avif on "medium" bitrate for both.

Anyways, hope this helps, Jyrki!

And whoever is running JXL fan sites, it would be a good idea to show off situations where jxl utterly outclasses the competition, like in the "Sunset" image especially. That one is amazing!

[bright monolith]

"White Dune" is another example of a job well done by libjxl

[native hawk]

The results from JPEG XL are complicated, because the quality setting actually dynamically chooses between normal mode and "modular" mode, so you're seeing two different modes of operation depending on what the quality level is.
Well that just isn't true

[bright monolith]

Yeah that sounded like news to me as well

[scenic latch]

Well technically it’s true, it switches to modular when quality is set to 100

[native hawk]

Though that only goes down to distance 1

[bright monolith]

"Kopenhahen," another good job and obvious win for jxl. The details on the red surface disappear with other codecs. Someone needs to put this comparison on the front page of the jxl fan site!

These are real good examples of what jxl does right compared to other codecs. Very real and compelling examples of the strengths that really ought to be on display on #website

"Source du Pecher" the stone exterior of the top of the leftmost building, very good

"Valee de Colca" has good stone texture too

The floors in "Mercado" too

The saturation issue is very noticeable in these photos however, also. It detracts and distracts from the otherwise very victorious images

Excessive ringing artifacts can sometimes manifest as strange chroma speckling. See "Clovisfest" for an example where libjxl falls flat on its face and libaom completely destroys libjxl here

I hope these examples can help @hollow oar when he's doing the tuning work. I think that's what this thread was originally for.

I think examples of where libjxl does outstandingly well are also important to keep an eye on too

To make sure they don't regress

And there are certain things that libjxl just utterly blows everything else out of the water somehow

[native hawk]

This thread was originally complaining that JXL looked bad at extremely low bpp compared to AVIF.
You're thinking of https://discord.com/channels/794206087879852103/1278292301038227489

[bright monolith]

Oh

Maybe something can be done at low bitrates though to limit ringing and strange color distortion that happens only at low bitrates

Although the strange color weirdness sometimes happens at high bitrate too

[hollow oar]

I'm starting on this on Monday, finally 😄

[native hawk]

Again, https://discord.com/channels/794206087879852103/1278292301038227489 was the main thread, but there are some images in here of extremely low quality too

[bright monolith]

I gave some examples of images where jxl is clearly doing incredibly well, too.