How JPEG XL compares to other image codecs

Posted by bentocorp 10/23/2024

How JPEG XL compares to other image codecs(cloudinary.com)

156 points | 71 comments

Uncorrelated 10/27/2024|

Articles about the merits of JPEG XL come up with some regularity on Hacker News, as if to ask, "why aren't we all using this yet?"

This one has a section on animation and cinemagraphs, saying that video formats like AV1 and HEVC are better suited, which makes sense. Here's my somewhat off-topic question: is there a video format that requires support for looping, like GIFs? GIF is a pretty shoddy format for video compared to a modern video codec, but if a GIF loops, you can expect it to loop seamlessly in any decent viewer.

With videos it seems you have to hope that the video player has an option to loop, and oftentimes there's a brief delay at the end of the video before playback resumes at the beginning. It would be nice if there were a video format that included seamless looping as part of the spec -- but as far as I can tell, there isn't one. Why not? Is it just assumed that anyone who wants looping video will configure their player to do it?

terribleperson 10/27/2024||

Besides looping, video players also deal kinda badly with low-framerate videos. Meanwhile, (AFAIK) GIFs can have arbitrary frame durations and it generally works fine.

mikae1 10/27/2024|||

> GIFs can have arbitrary frame durations and it generally works fine.

But we shouldn't be using animated GIFs in 2024.

The valid replacement for the animated GIF is an animated lossless compressed WebP. File sizes are are much more controlled and there is no generational loss when it propagates the internets as viral loop (if we all settled on it and did not recompress it in a lossy format).

michaelt 10/27/2024|||

Most modern video container formats support arbitrary frame durations, using a 'presentation timestamp' on each frame. After all, loads of things these days use streaming video, where you need to handle dropped frames gracefully.

Of course, not every video player supports them well. Which is kinda understandable, I can see how expecting 30 frames per second from a 30fps video would make things a lot simpler, and work right 99.9% of the time.

Pikamander2 10/27/2024|||

> With videos it seems you have to hope that the video player has an option to loop

<video playsinline muted loop> should be nearly as reliable as a GIF in that regard.

The one exception that I've found is that some devices will prevent videos from autoplaying if the user has their battery-saver on, leading to some frustrating bug reports.

eviks 10/27/2024|||

How does GIF require support for looping as opposed to it being just a player implementation no different from any other format?

acchow 10/27/2024||

GIF format includes a flag inside it to indicate how many times (or forever) to loop the video.

HEVC does not have such a flag. Quite unfortunate

eviks 10/27/2024||

Interesting, it seems like this isn't part of the spec per wiki:

> Most browsers now recognize and support NAB, though it is not strictly part of the GIF89a specification.

So I guess the players for the new video codecs could do something similar and agree on some metadata to be used for the same purpose?

codetrotter 10/27/2024||

> Most browsers now recognize and support NAB

The “NAB” was introduced in Netscape Navigator 2.0, 50 million years ago.

The phrasing with “most browsers” and “now” is a bit weird, on part of whoever wrote that part of the Wikipedia article.

Every major browser that I know of has supported animated gifs since forever.

Any browser that doesn’t is probably either a non-graphical browser in the first place, or one that has like five people using it.

eviks 10/27/2024||

You're doing a similar phrasing weirdnesd as the wiki

> Every major browser

There are only 2 major ones

And this brings us back to the main point - this is no "format" issue, Chrome could just as well support some metadata field as "loop for n" for the newer video files, and the situation would be the same as with NAB when Safari adds it.

teruakohatu 10/27/2024||

There are two major browser engine lineages but at least four of 'major' browsers (>5% market share) and a number of minor browsers.

troupo 10/27/2024|||

There are three major browser engines. Chrome derivatives cannot be counted as a separate browser (yet, or ever? remains to be seen)

eviks 10/27/2024|||

5% is minor, not major.

mort96 10/27/2024||

Come on, this is ridiculously pedantic. They used "major browser" to exclude WiP and hobby project to avoid pedants coming in and saying "uh Ladybird doesn't support looping gifs yet" or whatever. But I guess there's no pleasing the pedants.

eviks 10/27/2024||

You've missed both points:

- the 5% point you're responding to doesn't refer to the original "Every major browser"

- the original reponse just highlights that there is no non-pedantic difference between "most browsers" and "Every major browser", so that was the start of the anti-pedantism battle

mort96 10/27/2024||

No, this is ridiculous. The claim isn't about "most browsers" but the ones people actually use. You know, Firefox and those based on Chromium and WebKit. There could be 100 browser hobby projects out there which don't support NAB, "most browsers" would then not support NAB, but pretty much everyone would still be using a browser which supports NAB.

"The major browser engines" is a commonly used phrase to refer to Chromium, WebKit and Gecko (and formerly Trident and Presto). You're willfully misudnerstanding it. Please stop.

This will be my last response in this thread, this conversation is absolutely ridiculous.

eviks 10/28/2024||

> Come on, this is ridiculously pedantic. They used

"most browser" to include the major ones, it's a commonly used way of expressing it

> But I guess there's no pleasing the pedants.

who come up with imaginary examples like this:

> There could be 100 browser hobby projects

DidYaWipe 10/27/2024|||

The same thing stood out to me. With the popularity of animated GIFs, it's disappointing and ridiculous for a new Web-friendly image format to omit at least a simple multi-image/looping facility.

As for your question about video looping: Nothing prevents that, although I don't know of a container format that has a flag to indicate that it should be looped. Players could eliminate the delay on looping by caching the first few frames.

Dwedit 10/27/2024||

JPEG XL is three different codecs in one.

There is a lossless ultra-packer for existing JPEG files. It's completely reversible, you can get byte-for-byte identical JPEGs back.

Then there is "VarDCT" mode, which acts like JPEG, lossy Webp, or video codecs.

Then there is "Modular Mode", a completely different kind of codec that has different kinds of compression artifacts than JPEG-like codecs. The compression artifacts you see tend to be more like sections becoming more pixelated, or slight color differences. Strong edges don't have ringing artifacts. Modular mode mainly is used for lossless compression, but also allows lossy compression.

pizza 10/27/2024||

Technically it also had a fourth :^) [0] but it was spun out into a separate project of its own, jpegli [1]: JPEG but it uses some tricks from JPEG XL. These include spatially adaptive quantization, quantization matrices that better preserve psychovisual detail, more efficient color spaces, and also HDR (10+ bit depth) support [2].

[0] https://github.com/libjxl/libjxl/tree/main/lib/jpegli

[1] https://github.com/google/jpegli

[2] https://opensource.googleblog.com/2024/04/introducing-jpegli...

chronogram 10/27/2024||

Pretty good news! I imagine that it'll take a while before libjxl and jpegli won't both supply a cjpegli binary so that'll be mildly annoying at the start, but hopefully this way it'll be adopted quicker so it'll accept more input formats and image software will switch over to jpegli native export instead of using the libjpeg compatible controls.

It's a really excellent software. Its default output quality is storage quality, while the file size is acceptable for mobile data and cloud storage of pictures in most countries. It producing progressive pictures by default still helps when quickly swiping through a whole album of vacation pictures stored on cloud storage, and its progressive output actually reduces size rather than add to it. And it's compatible with everything so now I just throw everything lossy I produce through its default settings until JXL becomes natively supported in Chrome and Windows.

JyrkiAlakuijala 10/27/2024||

Thank you for such a kind praise for Jpegli! This made my day.

lifthrasiir 10/27/2024|||

The lossless JPEG recompression is a combination of VarDCT and some additional metadata. In fact, VarDCT should be considered as a (very large) superset of JPEG1 compression. The distinction between VarDCT and Modular in JPEG XL is relatively clear, but in reality VarDCT would still use modular encoding for various data anyway so it is hard to consider one without another. (Compare with Opus, which also uses two main mechanisms but mix them so well that they can't be really separated.)

Joel_Mckay 10/27/2024||

Interesting, similarly the only part of the JPEG 2000 standard I encountered in the wild was for the lossless part of the standard.

I guess Tiff is still the hot mess it always has been too... lol =3

SG- 10/27/2024||

I believe JPEG XL is now supported in macOS 15 and iOS18.

edit: previous discussion about it https://news.ycombinator.com/item?id=41598170

heliographe 10/27/2024|

It’s supported, but unfortunately no 3rd party APIs yet. It’s a bit surprising they wouldn’t ship them on launch to encourage adoption.

I make photography/camera apps and would like to support JPEG XL natively (without having to rely on 3rd party code) so I hope it’s something they add soon!

troupo 10/27/2024||

> It’s a bit surprising they wouldn’t ship them on launch to encourage adoption.

Because Apple is all in in HEIC/HEIV. The "high efficiency" codecs that require up to a second on an M1 Pro to render an image. A comparable image in PNG renders instantly

illiac786 10/28/2024||

They might have been all in on HEIC, yes, but they definitely are not any more if they are introducing JPEG XL.

I rather think they are doing a cautious test run, and decided that API was for a later phase. (decision which I do not agree with btw)

troupo 10/28/2024||

Oh great. Another image format no one really supports, that requires hardware decoders, and will probably take 2 seconds to decode on a modern supercomputer

illiac786 10/28/2024||

I’m a big fan of XKDC #927 [0] but if we followed your position, we would never introduce any new standard, ever.

https://xkcd.com/927/

troupo 10/28/2024||

I'm bot a big fan of codecs that seem to only benefit a few megacorps

illiac786 10/28/2024||

please provide some details on why you think JPEG XL only benefit megacorps – if you say it like this, it just sounds like trolling.

couple of counter arguments on my side:

* JPEG XL heavily reduces storage requirements both for lossless and lossy compression * JPEG XL allows to reversibly compress old JPEG to further reduce storage requirements * JPEG XL is patent unencumbered * JPEG XL support very high definition pictures (jpeg does not), native HDR support, higher number of bit per pixel, etc. etc. * Google wants JPEG with gain map to support some form of HDR, as well as now introducing XYB color coding, etc. etc. They clearly are against JPEG XL

jonsneyers 10/27/2024||

This article is from 2020. I think so far it has aged reasonably well. But you might be interested in my more recent articles too. You can find those here: https://cloudinary.com/blog/author/jon_sneyers

morpheuskafka 10/27/2024||

I found this interesting note in the article:

> HEIC and AVIF can handle larger [than 35MP, 8MP respectively] images but not directly in a single code stream. You must decompose the image into a grid of independently encoded tiles, which could cause discontinuities at the grid boundaries. [demo image follows].

The newest Fujifilm X cameras have HEIC support but also added 40MP sensors--does this mean they are having to split their HEIC outputs into two encoding grids?

It seems like the iPhone avoided this, as 48MP output is only available as a "ProRAW" i.e. RAW+JPEG, which previously used regular JPEG and now JPEG-XL, but never HEIC.

firecall 10/27/2024||

Will JPEG XL support return to Chrome?

Apple supporting it surely has to be a signal to begin wider adoption!?

SoothingSorbet 10/27/2024||

I don't know if Google has changed their position, but Mozilla is willing to accept a memory-safe (read: Rust) JXL decoder [1].

If it becomes used in Firefox, maybe there's a chance that Google would see the benefit in picking it up?

[1] https://github.com/mozilla/standards-positions/pull/1064

GoblinSlayer 10/27/2024||

>100,000 lines of multithreaded C++

>potentially introduce memory safety vulnerabilities across the myriad of applications that would eventually need to support it

Right, new code, new vulnerabilities.

RedShift1 10/27/2024|||

Not unless Chrome starts having serious marketshare issues.

theandrewbailey 10/27/2024|||

I recently wrote a script that encodes an image to fall within a size range[0]. After toying with it, I noticed that smaller AVIF files are completely fine for web use, but identically sized JPEG XL files are not. Given ubiquitous browser support for AVIF[1], unless JPEG XL gets much better at smaller sizes, I reluctantly agree that Chrome's call to drop JPEG XL is the right one.

[0] https://gist.github.com/theandrewbailey/4e05e20a229ef2f2c1f9...

[1] https://caniuse.com/avif

chronogram 10/27/2024||

In what environment do you work where you need such low quality images though? In my web environment I only want the highest quality I can get at a reasonable size and I've never been interested in slightly less awful looking tiny images. In another comment I wrote about using jpegli at its default distance of 1 for everything and being happy with that size, so maybe I work in a completely different environment to you.

theandrewbailey 10/27/2024||

A normal one like everyone else I guess. No need to waste bandwidth and storage if you don't have to. If an image looks good to me, I'll try going lower until it doesn't look good, then go back one step. I've been surprised many times by just how low it can go and still look good. That script I wrote defaults to what I've settled on. (AVIF between 0.5 and 1 bpp at 1 megapixel, increasing or decreasing by square root of total pixels in image, plus JPEG fallback.)

If I change my mind, I keep high resolution originals of everything to do it again.

jmb99 10/27/2024||

Apple’s supported JPEG2000 for years, but no one else does. I imagine the same will be the case for JPEG XL.

cpeterso 10/27/2024|||

Safari dropped support for JPEG 2000 in Safari 18 last month: https://webkit.org/blog/15865/webkit-features-in-safari-18-0...

dvhh 10/27/2024|||

One issue is that format support detection was iffy, compared to Jpeg XL where people knows to use the accept header to declare support format.

Animats 10/27/2024||

JPEG XL is supposed to be a progressive mode. Can you read a lower resolution from the file by reading only part of the file, as you can with JPEG 2000? Is there a header which tells you how much file to read for the desired resolution?

lifthrasiir 10/27/2024|

Yes, but kinda complicated.

You have to first read the image header, an optional ICC profile and finally a portion of the first frame. This first frame might actually be a preview generated by an encoder, but should be fine for our purpose and it's not hard to seek to subsequent frames anyway. The frame itself contains its own header and all offsets to per-frame sections ("TOC"), while there is always one LfGlobal section that contains the heavily downscaled---8x or more---image in the modular bitstream, even when the frame itself uses VarDCT.

Any higher resolution would require some support from the encoder. The prime mechanism relevant here is a version of the modified Haar transform named Squeeze, which generates two half-sized images from one source image. As each output image is placed to distinct sections, only one out of two output images is needed for low-fidelity decoding. If the encoder didn't do any transformation however (often the case in VarDCT images), then all sections would be required regardless of the target resolution.

Therefore it is technically possible and in fact libjxl does support partial decoding by rendering a partial bitstream, but anything more than that would be surprisingly complex. For example how many bytes are needed to ensure that we have at least 8x downscaled image? This generally needs TOC, and yet a pathological encoder can put the LfGlobal section to the very end of frame to mess with decoders (though no such encoder is known at the moment). Any transformation, not just Squeeze, has to be also accounted to ensure that all of them will produce the wanted resolution once combined. Since the ICC profile and TOC already require most entropy encoding stuffs except for meta-adaptive trees, even the calculation of the number of required bytes already needs about 1/2--1/3 of the full decoder in my estimate from building J40.

That said, I'm not very sure this complexity could've been radically reduced without inefficiency in the first place. In fact I've just described what I wanted when I started to build J40! I think there was an informal agreement that the ICC profile could have been made skippable, but you still need all the same stuff for decoding TOC anyway. Transformation is a vital part of compression and can't be easily removed or replaced. So any such tool would be definitely possible, but necessarily complicated, to build.

Animats 10/27/2024||

OK, so it's not a feature of mainstream JPEG XL implementations.

It's a standard feature of JPEG 2000, but JPEG 2000 decoders are rare, and either expensive or slow and buggy.

gforce_de 10/27/2024||

Something seems wrong im this article. The side-by-side comparison shows 4 formats:

  · Original PNG image (2.6 MB)
    · Name "high_fidelity.png", but in fact 298.840 bytes and format: JPEG
  · JPEG XL (default settings, 53 KB): indistinguishable from the original
    · Name "high_fidelity.png.jxl.png", but in fact 3.801.830 bytes and format: PNG
  · WebP (53 KB): some mild but noticeable color banding along with blurry text
    · Name "high_fidelity_webp.png", but in fact 289.605 bytes and format PNG
  · JPEG (53 KB): strong color banding, halos around the text, small text hard to read
    · Name "jpeg_high_fidelity.jpg", but in fact 52.911 bytes and format JPEG

The comparison does not make any sense, everything is just wrong. Also when encoding the large original PNG image to AVIF, it has only 20.341 Bytes with no visual change, see: http://intercity-vpn.de/files/2024-10-27/upload/

throwaway0665 10/27/2024||

PNG is lossless so they are using it to display compression artifacts of formats that may not have had wider browser support.

eviks 10/27/2024|||

> high_fidelity.png.jxl.png

means the original was

- loss-converted to JXL, measured as 53k

- then losslessly converted to PNG to be displayed on the website

spider-mario 10/27/2024|||

52 911 bytes instead of 53 kB is really not that far off.

And your AVIF is certainly not without visual changes. The colours are off and there is visible ringing.

gforce_de 10/27/2024||

you are right, i changed it to "acceptable visual changes".

morpheuskafka 10/27/2024||

> "jxl.png", but in fact 3.801.830

I guess that is because the noise from the lossy encoding creates more entropy, that then has to be losslessly encoded as PNG, which pushes the files size above the original?

theandrewbailey 10/23/2024||

(2020)

JyrkiAlakuijala 10/27/2024|

JPEG XL is seeing strong industrial adoption where image quality matters: professional and prosumer image acquisition (as a replacement, with huge benefits, of traditional raw imaging in iPhone 16 Pro) and processing and storage reduction with Digital Negative, ProRAW and medical imaging with DICOM. Clinics with archiving and telemedicine needs benefit massively.

More comments...