▲ 309 points
back
113 comments
(Note: I used to be employed by Mozilla, and in that capacity I was the owner of Mozilla's image decoders. I've been disconnected from all decisions for almost a year, though.)
The main take-home here is that while Google's numbers all show WebP as being objectively better, the metrics they chose for comparison were relatively bad (i.e., some of them didn't take into account colour or didn't model colour correctly), and once you accounted for that the numbers were not nearly as good a story for WebP; in some cases, JPEG outperformed it.
The facts that (1) WebP was not terribly compelling technically, (2) JPEG is already supported by everything on the web, not to mention devices and mobile phones etc, and (3) there's still headroom to improve JPEG in a backwards-compatible way, meant that WebP was (and, it seems, remains) a non-starter.
Until JPEG supports transparency, it leaves a vast hole where a good lossy alpha enabled format is needed - namely icons. With the high resolution of mobile devices, using PNG for this use case is a huge waste. Regardless of the quality differences, WebP fills a major pain for mobile and web developers. I really think Mozilla should just support it.
There are a few ways of making fully backwards-compatible "lossy" PNG: http://pngmini.com/lossypng.html
You can have icon files 3-4 times smaller, and large photorealistic images 2 times smaller than the regular PNG.
Yeah, depending on the image putting a bmp into a zip is actually significantly smaller than a png. Well, bmp has no alpha channel, but just as comment to the size of pngs. pngs even uses the zip algorithm in a way that is supposed to be optimized for images, but apparently it is not. E.g. the tiles here are a lot smaller as bmp in a zip: http://panzi.github.io/mandelbrot/
Ok, it put them into one single zip and can't remember if it was solid or not, so it might be the cross-file compression that makes the major difference here.
Plus with chrome already supporting it, WebP has ~50% market share in most countries.
Now we'll have two similar but competing technologies and web developers will simply resort to the older formats.
How much of a gain do you get in WebP versus a properly optimized PNG for icons? I can't imagine it's a compelling difference.
I had read in the past that HEVC was probably the best for still image lossy compression. [1]
Does anyone have further information on this?
[1] http://en.m.wikipedia.org/wiki/High_Efficiency_Video_Coding#...
Best from a compression perspective, perhaps. There's the minor detail of it being patent-encumbered. There's no license pool available yet, so you couldn't pay for it if you wanted to. The proposed terms of the pool that is organizing (with some of the known patent holders refusing to join) would require up to $25mln/year for the video codec. No idea if you'd be able to get a better deal for still-images only.
Meanwhile, JPEG is free.
According to the study linked in OP, it's still true. https://people.mozilla.org/~josh/lossy_compressed_image_stud...
But is there headroom for JPEG to replace animated gifs? If I look at mobile social apps these days, animated GIFs eat up enormous amounts of data. JPEG doesn't seem to have an answer for this, and no one has yet, it appears, made the <video> element work for this use case.
> no one has yet, it appears, made the <video> element work for this use case.
Gfycat seems to be getting more and more popular or at least I'm personally starting to see a lot more of their links in place of imgur gif links.
Expanding on Ray's comment above, this is really the compelling use case for WebP: a better container that does it all. Right now we have PNG, GIF and JPEG for three different use cases: animation, transparency and lossy encoding.
I can't easily pick and choose which of these I want, which means that we end up with 20MB GIFs that could easily be 1/5 the size, and crazy hacks to get lossy images working with transparency.
This is particularly painful for HTML5 gaming in my previous experience. For one of my projects that involved a giant, high-res game board with transparency (Nick's online Pai Sho), I ended up manually slicing the board up, converting center pieces to JPEG and leaving the edges in PNG. The images are all pasted together to make the final, seamless experience. What a PITA!
> no one has yet, it appears, made the <video> element work for this use case.
4chan has, and it's significant not only because they have a lot of traffic, but also because their implementation has to be pretty solid -- 4chan users would love nothing more than to troll the administrators by breaking this.
http://blog.4chan.org/post/81896300203/webm-support-on-4chan
Yes, no one except for Twitter: http://mashable.com/2014/06/20/twitter-gifs-mp4/
Honestly, animated GIF (like all other image formats) is a terrible format for video-style stuff. It's fine for line art (well, sort of fine, anyways), but if you actually using it for video you really should be using <video>, because it takes into account temporal information. (gfycat is a service based around this already.)
It might be possible to do an APNG-style backwards-compatible animated JPEG, but it'll still be worse at it than video formats will be.
Thanks to the ACID3 test, all modern browsers support animation in SVG and svg in an img tag. Which means you can do extremely small vector based animations, or write an encoder that embeds jpegs and animates them like a film strip. You can have alpha transparency using a mask image via SVG filters. You can get really fancy and animate with deltas. The efficiency you lose by the base64 encoding is regained by gzipping.
APNG (https://en.wikipedia.org/wiki/Apng) is the answer, but Chrome doesn't support natively.
Webm's definitely been getting some traction as a replacement, if 4chan's support is anything to go by.
Twitter now routinely auto-converts animated gifs into embeded web video.
Could you conduct a detailed article and publish it please? I think it would be useful for people who are interested in this but don't have the background about JPEG or WEBP internals.
NVM I found what I was looking for:
http://people.mozilla.org/~josh/lossy_compressed_image_study...
It should be noted, since it is nowhere to be seen in this post, it breaks API and ABI while still presenting itself as libjpeg version 6 to the system, which is very evil.
Open Issues:
https://github.com/mozilla/mozjpeg/issues/67 https://github.com/mozilla/mozjpeg/issues/21
It is sad that Mozilla wont implement WebP format in addition to GIF/APNG due to political reasons (I see no valid technical reasons to not include this code, especially after the news about including code for DRM). They've even disabled comments in the WebP bug [1]. And this bug was already second, they've closed first one [2]. You still can write your comments in this bug (about APNG removal) - it is not closed yet [3]
[1] https://bugzilla.mozilla.org/show_bug.cgi?id=856375
We (as in browser vendors as a whole, which although I'm no longer directly part of I'm still around enough!) don't want to add stuff to browsers just because it's technically sound to include.
People propose plenty of things with technically sound code, but there's far more to developing a platform than including everything that's technically sound. You have to make decisions as to what you wish to include. Once a something becomes near universal (in terms of marketshare, not in terms number of browsers; see IE's peak for the clear examples there!) it becomes treated as a baseline; once it's there it's incredibly hard to remove. Every time something else gets added that's more code to maintain, more code to ensure correctness (and security!) thereof, and a higher barrier of entry to the browser market. As a result, the default answer to adding anything should be "no".
Did you not read the section in the linked article about how WebP is not measurably better than JPEG?
I wouldn't take the fact that a bug has been filed about removing APNG as evidence that it's going to happen. It definitely won't in the near future.
It's especially amusing since Firefox supports WebM, which is VP8 video in a Matroska container. WebP is a VP8 frame in a simple container.
How does this MozJPEG 2.0 encoding compare to JPEGMini? I've been using JPEGMini for all client projects, it results in optimized JPEGs that work on all platforms with minute differences, rarely visible to the naked eye, only when I subtract pre/post in Photoshop. Very rarely does it save less than 5% and for larger images (saved with PhotoShop, save for web) often 15-30% or even more.
Is the test set of images available?
jpegmini works by compressing an image as much as possible while keeping the distortion calculated by there particular metric below a threshold. To compare against mozjpeg you'd need to compress a bunch of images with JPEGMini and then compress those images with mozjpeg so that they were the same size. You could then compare the quality of the resulting images.
The test set of images is here: https://github.com/bdaehlie/web_image_formats/tree/master/im...
If you're interested, https://github.com/dwbuiten/smallfry is program similar to JPEGMini that uses mozjpeg as its backend.
http://people.mozilla.org/~josh/lossy_compressed_image_study...
mozjpeg does Trellis quantization, there is no detailed explanation about jpegmini, a guess is it optimizes the DCT similarly to how it is described in this paper:
jpegrescan AFAIK is the gold standard today.
https://github.com/kud/jpegrescan
This is all fine and dandy, but whatever happened to JPEG-2000 and wavelet-based compression? Still licensing/patent issues blocking wide adoption?
It's still competitive performanc-wise but adoption has been lacking. Patents were an issue but there were other problems, too: uncertainty about rights and the high complexity of the spec (which for awhile was, IIRC, $1500 simply to get a copy) meant that there wasn't much traction in the open-source world, which in turn meant that many image processing apps either didn't support it at all or used a slow library with limited compatibility and incomplete spec support. There are several commercial libraries available but they also had compatibility issues (largely resolved by now) and required negotiating licenses.
Mike Shaver had a good comment in the Mozilla feature-request ticket which is no doubt representative of many other projects' concerns:
https://bugzilla.mozilla.org/show_bug.cgi?id=36351#c120
None of this is insurmountable, of course, but it made a lot of people hesitant to invest heavily in it. I wrote about the long-term risks awhile ago:
http://blogs.loc.gov/digitalpreservation/2013/01/is-jpeg-200...
The situation is getting better – OpenJPEG is now actively developed (see http://www.openjpeg.org), supports many of the more valuable features (e.g. tiled decoding so you can avoid decoding an entire large image to extract a fragment), and is becoming an official reference implementation:
http://comments.gmane.org/gmane.comp.graphics.openjpeg/773
Recently ImageMagick and Pillow (the maintained fork of the Python Imaging Library) added support for JPEG-2000 using OpenJPEG and the same trend seems to be happening in other places. The big remaining challenge is browser support for non-Apple browsers but that's now possible, if somewhat grotesque, using Emscripten on OpenJPEG to render into a canvas tag.
It's very sad Mozilla does not even mention their own codec, Daala, and only compares to HEVC etc. I suspect it's severely lacks funding.
Daala development is proceeding at a rapid pace. We're just not ready to bring Daala into this conversation quite yet.
There's a progress report for Daala that makes a similar comparison. They claim to beat JPEG, VP8 and x264 but not HEVC (yet!) on still images.
Not sure how fair a comparison it is but FWIW I built it and ran it (./cjpeg) over a few photos and got ~18% savings in file size vs. imagemagick convert at quality=50%
5%? I don't know but I hoped for significantly more.
IMO the linked study is pretty much useless, because of enforced 4:2:0 chroma subsampling. I guess this is because some of the newer codecs were originally video codecs where 4:2:0 is standard, but for still images I expect much better. Photoshop, which is obviously widely used on the desktop to make JPEGs, only uses 4:2:0 for quality <= 50 by default.
This also makes me question the use of metrics to measure image quality.
I don't think it would make sense to compare some codecs at 4:2:0 and some at 4:4:4. Unless they can all do 4:4:4, you have to pick a subsampling that they can all do and compare them at that so it's like-for-like.
If your goal is to demonstrate 'JPEG can deliver better quality than this format by using 4:4:4, because that format can't do 4:4:4', then that's great, but it's a different observation than the one they're trying to make. (Also, I would be pretty curious about whether the different subsampling produces superior file sizes.)
Facebook's $60,000 donation is interesting. The article says that the average file size reduction was 5%. Given that Facebook has built datacenters capable of storing exabytes of photo data in "cold storage", I wonder how much money Mozilla just saved Facebook in storage space costs in relation to how much they donated.
It would be interesting to learn what the cost of CPU overhead vs. additonal cost of 5% cold storage is - but I bet the FB folks did that calculation.
Seems to be source only. Has anyone built win64 binaries for this that they'd like to share.
On that subject; say what you want about how clunky CMake is, but it's really nice to be able to build projects like this with VC++ or plain MinGW-w64.
Is there a command line tool available? I’d like to include it into my website build process.
Yes, building the source will produce (among other things) a command line application called "cjpeg", which can encode or re-encode.
- Alpha channel
- Animation
- Proper metadata
- File size
The list of things that WebP offers that JPEG, PNG, or GIF alone will never address.
If Google had done this, people would be in here complaining it wasn't open even though it was released under an open source license.