JPEG XL

So I don't see how this example disproves that `a[b]` is identical to `*(a + b)`.

ah, you said "if a has a pointer or array type"

1 usually has neither of those

Ah, you are right!

`a[b]` in C is literally just sugar for `*(a+b)`

in C++ it's more complex than that

Yeah, I was assuming C++ where classes complicate the matter.

('cause `operator[]` is a thing)

Is this fair? https://twitter.com/jonsneyers/status/1403679160502034435?s=19

I'm not sure what panorama views are but everything else likely relates more to the webpage/html than to the image format, no?

> jxl is excellent for options 1 and 2, great for option 3, okay for option 4. > okay for option 4. On photos, yes, but on low fidelity art content current jxl encoder is still noticeably behind AVIF/WP2 and even sometimes WebP

I meant option 4 w.r.t. jpeg/png

And yes, "okay" is noticeably behind "excellent" :)

Okay, I made some bindings to butteraugli for Rust and they work but I for the love of God am lost in C++

haha

what's wrong with it?

I mean, in this specific case, not in general 😛

Never made any "projects" in cpp, will have to give it time

if high pixel density means low quality and low bitrate per pixel, then possibly no with some reservations

that's more dependent on the application/browser/markup than the image format, no?

I consider that JPEG XL is a good choise in the 1 to 300 dpi range, 1200 dpi would benefit from different design solutions

🤔 https://github.com/libjxl/libjxl/issues/202

Encoder or format?

CLIC is a project that rates neural codecs

as a side effect it rates psychovisual metrics

they do the codec rating by human raters

so it is possible to check how well the metrics correlate with human opinion

JPEG XL has a possibility to plugin other psychovisual metrics for kitten and tortoise modes

if we learn that some other metric works better in a specific bit rate range, we can switch to that for that range

(all altogether)

the metrics have progressed a lot during the last 2-3 years, particularly the neural metrics in the low bpp area

(I had proposed George to file this issue for JPEG XL -- I had made a similar proposal for him to file issues against SSIMULACRA and DSSIM)

(I don't like that engineering and science are far apart -- engineering here using Butteraugli, DSSIM, SSIMULACRA, and science being 17 other metrics, with no intersection of engineering metrics, so I hope that this would let science and engineering to meet)

I think appeal metrics and fidelity metrics are fundamentally different things.

VMAF and many of the AI metrics are very much appeal metrics, designed to measure viewing comfort at low bpp

Butteraugli and HDR-VDP are fidelity metrics, designed to measure if humans can see a difference at high bpp (at low bpp of course they can)

Ssimulacra is a crude attempt at a compromise between the two

Appeal metrics are worthless in the high fidelity range, fidelity metrics are worthless in the low fidelity range

Lower pnorm butteraugli does extend its scope into medium fidelity but it's still fundamentally not the right approach if you want to measure appeal.

Appeal has much more of a semantic component: things like gaze prediction and expectations/viewer bias come into play, where denoising a sky can make a compressed image more appealing than the original if done right, etc.

Fidelity is more about the physical properties of the human visual system, appeal more about the mental processing that comes afterwards.

No-reference metrics (and no-refer nce or weak-reference (side by side) evaluation methodologies) don't make any sense for fidelity, while they do make sense for appeal.

Of course there is overlap between the notions, it's not a binary thing, and the appeal aspects (does the artifact annoy you?) are conditioned by the fidelity aspect (can you see the artifact?).

But any correlation study that considers "image quality" as a single entity across the bitrate spectrum is problematic imo.

I added a similar issue on SSIMULACRA and DSSIM, George was shy to add them himself 😄

You should create an image codec that just hashes the input, and uses it as a seed to generate a random pretty artsy image, storing just the seed and resultion in the encoded file. Super compact, high appeal, terrible fidelity.

just hash pixels and expand hash into jxl art

Duh. Why didn't I think of that.

This is obviously preferable.

Probably both format and encoder. I wasn't thinking much about the 1200 dpi range, and it's psychovisuals.

I see a lot of devs clamoring for using muted looped videos instead of animated images. Sounds good in theory, but in practice... For example, fifty 250x250 thumbs works just fine as GIF images. But if instead you're using fifty 250x250 webm videos as thumbs, browsers even on modern PCs are struggling to keep up.

I think it comes down to placement as well in general

You can't just have 50 animated anythings on the viewport at once

Animate on focus (YouTube is an example) or animate on load/scroll is better

Most ads nowadays too use video (usually h264 which is hwdec so it's pretty easy)

More than one animated thing on a website playing at the same time is typically bad web design imo

Unless you use a very old CPU/GPU most of those videos are handled in hardware

Nowadays it's GPU that does it

Do we know what in video decoding cost is related to number of videos and not just pixel count? Perhaps some simple work in optimizing the constant overhead in video decoding would solve the issues.

video coding is -- after all -- often far superior in bandwidth use in comparison to gifs and other animations, would be great to find how to make it workable

https://storage.googleapis.com/demos.webmproject.org/webp/cmp/2021_06_08/index.html#89-vezelay-basilique-2*2:1&AVIF-AOM=m&JXL=m&subset1

webp/webm team published a new comparison almost two weeks ago

which images/qualities are currently the biggest loss for JPEG XL vs. AVIF

LMFAO

<@!794205442175402004> <@!532010383041363969> <@!768090355546587137> <@!795684063032901642> <@!604964375924834314> <@!179701849576833024>

what a complete and utter dipshit of a spec

make sure you suck hard; google's teats are pretty dry these days

lmfao, nice try <@!604964375924834314>

enjoy making some fucks rich while you get breadcrumbs

man, living in this constant heatwave is really making everyone stupid

but let's update the jpeg spec so we can save some billionaires some hosting costs!

you think the hosting costs are the main benefit?

lmao

at least they put your daily microdose in the koolaid, am i right ladies?

this comment is spam

here come the oper simps

yeah, called it

what a complete shitshow of a marketing campaign

here comes daddy moneymaker

we're ready for your proclamation, pastor!

i'm good for some discussion, just keep it somewhat interesting please

tell me how we're all going to be fucked real good and long by daddy google, and how we're all gonna not choke on the CO2 their server farms are going to churn out for our benefit!

they're gonna save SOOOO MANY SINGLE DIGIT PERCENTAGE POINTS ON COSTS FROM THIS YESSSSSSSSS

*COUGH* FUUUUUCK YEAH IT'S SO GOOOOOD *HACK*

not sure what exactly your point is

Probably just some low effort trolling <:feelsbanman:853990457453707285>

are you okay? you seem a bit excited

yes, make sure to take your prescribed mood stabilizers. they're for your safety, <@!604964375924834314>

wouldn't want you ever emotional about absolutely anything, because you might face the permanent dread of an impending death handed down to you by those occupying your time with trivial compression matters

since all action is better done yesterday, make sure to take twice as many to catch up to all that missed time when you were too busy with those nasty feelings, right?

maybe you want to take that to <#806898911091753051> , <@!856560637986209862>

lmfao, he did it

he did the funny

weak. are you even trying

maybe you want to take this entire specification and do something altogether more better with your time: necking the fucks who are planning on starving your oxygen rich environment

after all, oxygen's such a nasty interactive chemical. you wouldn't want those harsh TOXINS in your breath, right?

should blackpill them about political issues around formats and AVIF and stuff but you're just pushing generic baits, idgi

<:BanHammer:805396864639565834>

yes, your specific brand of outrage is the ONE TRUE CAUSE, very online(tm) discord user <@!260412131034267649>

thank you for kind words

keep up the good work generating all that sweet data meat while your diet is reduced to ascetism to keep shareholders fat

honestly, don’t you have anything better to do with your time?

that’s kind of sad

says the motherfucker supposedly working on a REVOLUTIONARY techonlogy that will UTTERLY CHANGE THE WEB FOREVER

I will wait for something that is actually on topic

you all responded instantly

nah, I’m on vacation

but thanks for your concern

yes, vacation

IN YOUR WORK DISCORD

you are so hopelessly deluded

imagine having so much free time to join image format discord server and just throw generic baits. i sorta envy you

KEEP LOGGING ON AND MAKE SURE THAT BLACK MIRROR DOESN'T REFLECT ON YOURSELF ANY LONGER, RIGHT

<a:FrogPopcorn:821038588301017168>

imagine starting a fucking discord server for a fucking image format

Even I don't have that much time, damn

what exactly is your point? that Google is evil?

the average productivity skyrockets and this is the only thing left in the weakened mind of a completely subjugated class

but let's all get REAL EXCITED FOR NEW SHINY THINGS

NEW SHINY FOR THE NEXT HUNDRED YEARS MORTY

Yeah, shiny new tech is really exciting, with all the beautiful math and science 😛

HUNDRED YEARS NEW BARELY FEASIBLE DESTANDARDIZATION MORTY

If you want to do some marxist socio-economic analysis, I'm all for it, but let's maybe take it to <#806898911091753051> then

i turned myself into a bunch of adderall'd up devs morty

i'm fickle riiiiiiick

<@!794205442175402004> I think it would be right to ban him.

He does not seem to wan to have productive discussions.

the most common moronic discord user stereotype

prod your authority figures to restore your comfort

never face any conversation

LET'S TALK PAST EACH OTHER FOR TWENTY YEARS MORTY

Well one question HAHA, what is your interest being in this Discord server?

IT'S ABOUT THE DIALOG AM I RIGHT

if you paid a single fucking moment of attention and read what i said you'd have a clue

and yet you'd rather do the hackernews skim on literally everything you ever do

in the overwhelming anxiety of time robbed from you by those who profit from it

and it's my fault, right?

do you know how to switch channels in discord?

blame the ones sitting here with you in the cancerous fumes of this orgy of attention

It may not be your fault, but let's move to <#806898911091753051> please <@!856560637986209862> if you want to discuss more about this.

if you don't think this is on topic, you're not awake

don't post memes in general! back to my discussion about how the age of consent should be reduced to pre-term. ~ discord opers

so what is the thing you want to say about jpeg xl?

man i wonder if the eight pages of scrollback might have some hints

not really

they do not

but there's no git commit message so i'll just piss my pants and demand a from-first-principles rewrite instead!!

I sense some frustration

There is a git commit message history now however <:Thonk:805904896879493180> https://github.com/libjxl/libjxl/commits/main

this kind of retardation can only be maintained through a progressive rate of exploitation implemented over generations

your peers at your age were given resources beyond your wildest fucking dreams

you'll be working on your 2013 macbook that fell off of a fucking truck your entire life and think it's the grace of god to get PAID for the code you used to secure another dividend-paying stock you'll never be able to afford

and the moment someone offers you an upgrade and an office you'll atomize into ecstasy because of how starved you are for acceptance by the overwhelming triangulation of your only remnant social circles: coworkers

https://tenor.com/view/tevye-fiddler-on-the-roof-projection-gif-18152052

Well, projection much <:kekw:808717074305122316>

Anyway, HAHAHA, how is your job related to image/video/audio coding? What is your field of interest?

yes, dig out your 2014 conversation cointelpro techniques that were used to undermine your fucking elections

a teleoplexic spectacle just to replace the utter banality of television

came here for the automaton art but this is way better

my field of interest is exactly what this project's true goal is: the extraction of what a devout christian could only reason about through the term "soul"

I think it might be wise to take a little break, <@!856560637986209862>, and think a bit about what it actually is that you want to say

tfw jxl is literally satanism

you name the fucking newcomers "neophytes" in your desperation for purpose

you've been taught to generalize everything you can't understand so that your assumptions can never approach coming to terms with an author

to parasitize all dialog for your own emboldenment

neophyte literally means the same thing as newcomer

another generalization to meaninglessness

i can think of a few slurs that only mean "the color or mannerisms of a particular person," and if i said that to anyone with two fucking brain cells they'd meaningfully and rightfully disagree with the idea that my words don't carry meaning beyond my own immediate control

google is doing good work, that's why everyone pay them, you're just being jealous

which is why we fucking choose them, not lackadaisically piss them onto the concepts we mean to represent

possession of

whoops, fat hand.

You are quite funny, but I have to ask: why come here in the first place? I mean, it's not like you insulting us for being unproductive while you are being unproductive is really convincing.

possession of the doctrinal body by the devil, am i right ladies

too frustrated by truth to type huh

yes, smokers telling you to stop smoking are such UTTER RETARDS, RIGHT? HAHAHAHAHAHAHAHAHAHAHA

*huffs exhaust*

THIS IS KILLING ME

*injects gasoline*

DO NOT DO THIS

hahahahahahaha, hypocrites are just WRONG AND STUPID

No, I'm just saying that while you are saying some truths, you are mixing it with a lot of crap.

BETTER GET THE PUMP GOIN

this is only ok if you're a corporation

sed s/saying/backpedaling

googol big bwain. googol give me headpat. heckin doggo pupper is a cute and need imeg formot.

Another question HAHAHA: which image format do you think has the highest chance of winning the "format" war? JPEG-XL or AVIF?

googol want me make good doggo imog fomot.

me wouldn't wanna miss out on da fomot of a wifetime...

while they reduce that lifetime precipitously

this was the last warning, <@!856560637986209862>

your entire fucking planet will be uninhabitable in any present form by the end of your lifetime

(kicked)

jxl will spin on the remnant devices that supersede you until the power grid fails under its own unmaintained weight

google will pay no taxes nor licenses and you will have your house foreclosed

<:BanHammer:805396864639565834>

you're right we need psychiatrian

I don't mind some discussion but it has to be somewhat productive

Not lunatic ranting like that

ban him

Already banned Fabian.

So, this post is interesting: https://old.reddit.com/r/AV1/comments/o433uq/avif_decode_in_modern_browsers/

Why would he need to load AVIF images using WASM?

Firefox and Chromium both have dav1d for decoding.

iirc firefox not quite yet

another account appeared

HN frontpage naturally brings in some popularity, so yeah. shouldn't suspect all new accounts

I assume for Safari / old browsers? But wasm polyfills are not a great idea, better to just use jpeg/webp fallback

Yeah, only using one format for everything is not a good idea 😛

maybe in 3-4 years when all browsers support jxl it will be a good idea 🙂

(optimistic me speaking here)

https://twitter.com/nucknyan/status/1407559849991311361

Some twitter discussion later:

https://twitter.com/nucknyan/status/1407788560481067009?s=19

It's curious how many people express opinions against innovation (at least, relative to it's novelty).

Tbh I think here she just wants to get whatever next-gen codec (avif or jxl) universally supported so she can start using it. She'd probably prefer it to be jxl, but she'd rather have avif now then jxl later, which mostly just demonstrates how much people really do want to have better image compression on the web.

I answered to her

It's somewhat frustrating that typical reactions we get are either "jpeg is good enough, who cares about compression? Bandwidth is not an issue for images" or "avif has won and jxl is dead on arrival, end of story, you lost the race, avif was there first, they won"

Is there a way to make browsers prerender images that are currently hidden but might become visible any second? This would help to mask the low decode speed of JXL alpha.

Why don’t browsers just support both codecs and move on. They have an obscene amount of features anyway

That’s what I would do if I led the chromium project

And this was what Google did with Chrome before it became the dominant browser

Now that they are on top they can afford to be “slow” on stuff

First, "How about we try to improve AVIF encoders and decoders instead?" Then, "we need any format spec to be public so creative new approaches can be written against it" I agree with the philosophical argument of the latter, but this is still an ironic turn.

Facebook outage today related to jxl transition? Maybe they are getting ready for it

I don't think so.

ye, doubt it's related at all

never heard of z-lib. just how many of these libs there are

zlib is an implementation of the deflate compression algorithm. it's used by libpng.

correct. also there's zlib-ng which is p cool

oh lol i mixed up zlib and z-lib[rary]

Both are good resources

Oh great <:Poggers:805392625934663710> i see the iso pdf, but "chut" it's a secret 😄

Is it safe to speak about that in public? And say: "look it's open source code, spec are standardized and freeze, and you can READ the spec, use it NOW ! avif suck" 😄

It's only because some people don't want to use it because they don't have spec ... Not for creating an other encoder XD

Yes. AVIF's spec is basically on git, which is nice, but also means it's a moving target because making a change is easy (it's just getting a pull request merged).

So you can implement it, but then you better keep watching it because it's still evolving

Are additional channels implemented? (in the latest reference lib)

Partially. I don't think we have an api yet to get extrachannels besides alpha. But they do get decoded, it's just a matter of figuring out a good way to expose them. Spot color channels cannot be retrieved yet because of this, but the decoder will merge them with the RGB data it returns so at least you can see them in viewers.

ahh nice, i was reading this https://gitlab.com/wg1/jpeg-xl/-/issues/148 and this is similar requirements i have (storing additional per pixel lossless data)

We haven't priotized this because browsers don't need it, but maybe it's time we start adding that. <@768090355546587137> WDYT?

Would be great to have this working, JPEG XL is perfect for high performance desktop applications ... currently we're crawling along with webp on browsers 😄

I'm testing photoshop CC2019 what a *** joke, it does not support APNG to create animate picture, only gif ... adobe is a joke. I don't understand why people give us money.

to me jpeg xl needs at least 37 improvements like that 0,14% ringing one

to become acceptable to implement on squoosh

before the difference isn't actually visible

Why 37 improvements. Are you from future?

Yes.

He is indeed from the future <:Stonks:806137886726553651>

36.9703851 improvements, to be exact

Maybe CERN have some secret time machine research?

Sometimes it's hard to understand Japanese even with the translation, but here the point is roughly understandable https://twitter.com/a4lg/status/1408412785529131009

Like JPEG = for lossy and strange when it can encode and be good in lossless

shoudve been PNG XL

JXL

Jon's eXtra Lossless codec

Noooooo. Fxl is reserved to fabian. 😹

JXL is basically the combination of Jyrki's eXcellent Lossy and Jon's eXtra Lossless

Just kidding btw with those names, others have also contributed big parts of the codec

Luca, Lode, Sami, Jan, etc

jpeg x-3j2ls-et-al.

unrelatedly i'm trying to pick some branding colours and i'm wondering if there's like, any mathematical rule of thumb to give a reasonable chance the colours i pick can be represented reasonably accurately by a decent printer

i know it really depends on the individual printer which colours are representable

hence why i'm saying "rule of thumb" here

maybe one approach could be to find a bunch of printer ICC profiles, plot their gamuts, and see roughly where most of them overlap

https://photosauce.net/blog/post/what-makes-srgb-a-special-color-space

> So unlike every normal color space -- where you can calculate both the forward and inverse matrices to arbitrary precision by starting from the x,y values -- with sRGB you can only get the correct values by starting from the unambiguously specified forward matrix, rounded to exactly 4 decimal places.

> the whitepoint is intentionally not exactly D65, because it’s based on intentionally rounded colorants. This is not a matter of them being imprecise. Rather, they are precisely defined as being very slightly different from Rec.709. You can't calculate them 'more precisely' by going back to the x,y coordinates because they are already precisely and unambiguously defined in the spec.

i'm learning a lot today

not good

42% reduction is even not good

we need more

i had 5 mb to 1,8 mb size reduction

still not happy

even if the bpp is 2.923 bpp

yeah modular reductions are usually v good for animey kind of stuff

even if sourced from jpgs

nek is 37% not 63%

-37%

that's v good for lossless.

also if you knew english you'd notice that 63% is not reduction amount but fraction of original (63 + 37 = 100)

37% reduction sounds not really huge tho did you try slower speed

i suspect itd squeeze stuff out more at -s 9

huh so 42.3% reduction. kinda same as what fabian said. idk why one would consider that not good tho lol

something i could use for batch recompressing my downloads folder tho

like in background

just leaving it overnight

For m you want to use -Q

No -d

Or -q

tbh -q and -d should err or warn or something like that when they get forced -m

In my undestending you can use -d and -Q in the same image for mix content

ot: lighting looks really weird in that image 3d perspective wise

cannot unsee rip

q 95/96 not remember precisly

-d 0.5 == -q 95

d 1 is q 90

not criticising their work in general, its nice, just something what i noticed and makes me aaaaaaaa

-d will show after you type a q

but not the opposite

this is a bug

you could compress an image

is 5,5 mb

did it switch to vardct

try -Q

do -s 9 -q 100 -E 3

-m -Q 99 is rarely best then -m in my test

i'm uploading

who wants to try you need only 5/22 minutes of time

reply to this message

Not recommend use lossy modular for now.

-q 100 is lossless

do -s 9 -q 100 -E 3

I didn't added -m

do same command as me with newer build

with this image called windows 11

A.png

I'm waiting to receive the jxl

use latest commit and do -s 9 -q 100 -E 3 with that image i uploaded is only 6 mpx

it takes 22 minutes single core in a i3 330m

-E 3 and maybe -I 1

tell me also your processor and how many time it taked

don't run nothing in that time

only discord

no youtube music

v v big tortoise flags

don't do -I I

png pal7 Input file: E48AeCrUYAE0Fms.png | 454946 bytes Image: 1287x1800 pixels | 8 bits/sample | Indexed | 128 color(s) Delta filter: None Chunks: only critical

i gonna try v v big tortoise flags on my zen2

it's entry lvl zen2 but still shouldn't be too awfuly slow i think

``` cjxl -m -s 9 -E 3 -I 1 E48AeCrUYAE0Fms.png E48AeCrUYAE0Fms.png.ms9E3I1.jxl 62.68s user 0.11s system 100% cpu 1:02.17 total 454946 E48AeCrUYAE0Fms.png 260239 E48AeCrUYAE0Fms.png.ms9E3I1.jxl ```

so uhh 1 minute i guess, and i think further reduction is not that big

Calculations The percentage decrease was equal to -42.8%. 260239 is 57.2% of 454946. In absolute value, the decrease was -194707

ryzen 5 Summit Ridge 6c12t 20~30s > -m -q 100 -s 9 -g 3 -E 3 -I 1 --num_threads=12

Some lossless compress step can't optimize for multithreading, so I think best way is async and concurrent execute cjxl .

Like lzma2 use multithreading will reduce compression rate, so some people like use lzma or lzma 4x4.

i create a chart for that 🤓 (tab quality) https://docs.google.com/spreadsheets/d/1bTeraUXIl-nGM8c53IdURxmJbabX9eXqPZwVSynoH9U/edit?usp=sharing

I have seen printer drivers to do terrible things. Especially so for printing colors as a grayscale image.

Probably best to go for things that correspond to enough CMY ink but not too much - things look best if it's not too bright to require very sparse dots, and not too dark to require K to avoid too much ink coverage. At least that's my very unprofessional gut feeling 😅

Will do! 😄

Something like pure cyan should always look nicer (basically using the process colors as spot colors) than something like 10% cyan 70% magenta 20% yellow which will have more visible dot patterns

🤔 a clown business intentionally making their prints as spotty as possible

a clown-themed retro video game company so all the prints can be spotty and all the images on the website are pixel art with a small palette

(from https://www.reddit.com/r/woahdude/comments/o9ilda/only_the_stripes_are_colored_and_the_brain_does)

Nice demonstration of extreme color subsampling.

I first had to make it actually work 😅 https://github.com/libjxl/libjxl/pull/237

i'll add a test image for it soon

https://github.com/libjxl/conformance/tree/master/testcases/cmyk_layers

nice

Efficient Rendering of Heterogeneous Poly-Disperse Granular Media > We address the challenge of efficiently rendering massive assemblies of grains within a forward path-tracing framework. https://studios.disneyresearch.com/2016/11/11/efficient-rendering-of-heterogeneous-poly-disperse-granular-media/

reshade program detected the track and lost license + watermark

Hi, where can I find bitstream code specification?

Officially, from ISO.

The reference software is an executable specification

I have been playing a little with Nix (https://nixos.org/download.html) yesterday and wrote this small `shell.nix` with which `nix-shell` can drop you in a shell with all jxl dependencies installed (including clang): ```nix { pkgs ? import (fetchTarball "https://github.com/NixOS/nixpkgs/archive/21b696caf392ad6fa513caf3327d0aa0430ffb72.tar.gz") {} }: pkgs.llvmPackages_12.stdenv.mkDerivation { name = "jxl-dev-shell"; buildInputs = [ # Core build tools. pkgs.cmake pkgs.ninja pkgs.which # For building the documentation. pkgs.asciidoc pkgs.python3 pkgs.doxygen pkgs.graphviz-nox # Optional but makes the command-line tools more useful. pkgs.pkg-config-unwrapped pkgs.giflib pkgs.libavif pkgs.libjpeg # This is libjpeg-turbo, can be replaced with pkgs.mozjpeg if desired. pkgs.libpng pkgs.libwebp pkgs.openexr # Optional. For the plugins. pkgs.gdk-pixbuf pkgs.gimp pkgs.gtk2 # Optional and very large. For building the comparison tools. pkgs.extra-cmake-modules pkgs.libsForQt5.qt5.qtbase pkgs.libsForQt5.qt5.qtimageformats pkgs.libsForQt5.qt5.qttools pkgs.libsForQt5.qt5.qtx11extras ]; } ```

if leaving everything in, it uses quite a lot of disk space (~2 GB)

but at least it’s convenient

and reproducible thanks to the pinning at the top

What is the proper way to use ccache when building libjxl?

passing `-DCMAKE_C{,XX}_COMPILER_LAUNCHER=ccache` to cmake should do it

Thanks

so i have discovered that some grayscale(1 channel) jpg transcoded to jxl images are being decoded as rgb green(through api when out format is 4 channel) and completely black using djxl or output 1 channel

anyone know if this is a bug or a feature?

I doubt that it's a feature

who do i complain to

asking it here is pretty okay already i think though you can make an issue on libjxl github if you want to

include image this happens with when doing so

well it's more easily reproducible than i had thought

$ cjxl test.jpg test.jxl

well anyway i found that converting it to grayscale png then to lossless jxl is somehow significantly more efficient than just the lossless transcode of grayscale jpg

converting jpg will involve lossless jpeg recompression mode which is fully reversible into original jpg. converting png wont do that it'll have more choices but u wont be able to undo that action and get back byte-by-byte identical file

it sounds like a bug is in jpeg recompression mode

i'll take the smaller jxl so 🤷‍♂️

actually nevermind i just encoutered it being bigger

oh that's if it's not bugged

if you give cjxl `-j` flag it will not do lossless jpeg recompression stuff

is that supposed to fail for jpg input

i think it shouldn't?

works fine

<:thinkRaging:608369431730585642>

cjxl'ing that file indeed makes it green i wonder why

?

djxl fully restores file though

oh i see

the preview wasnt working

when i djxl it, it's all black

what full command you use

hmm it's only if i djxl to png

that wont do though because i need this to work with the api

ah, that doesn't do jpeg reconstruction

i guess ill hold off on transcoding the grayscale jpgs

make issue on github.com/libjxl/libjxl if youre not too lazy (or busy etc)

damn it it's not just grayscale

yeah will do now

if youre encoding manga kind of stuff, probably `-j` (if youre encoding from jpg) and `-m` can help

was -j just added yesterday or something

$ cjxl -j 33.jpg 33.jpg.jxl JPEG XL encoder v0.3.7 [AVX2,SSE4,Scalar] Failed to read image 33.jpg.

huh which version are you using

i built from source which i pulled yesterday

no it should def work.. can you share image or it fails for everything?

well i get the error for that test image i posted earlier

lemme try

``` % cjxl -j -m test.jpg test1112.jxl JPEG XL encoder v0.3.7 [AVX2] Read 2400x3456 image, 225.4 MP/s Encoding [Container | Modular, lossless, squirrel | 26-byte Exif], 6 threads. Compressed to 8810 bytes (0.008 bpp). 2400 x 3456, 2.70 MP/s [2.70, 2.70], 1 reps, 6 threads. Including container: 8888 bytes (0.009 bpp). ```

works here

wtf

did you not have libjpeg compiled in

maybe not

why does my stuff only show AVX2 and yours others too tho..?

Mine has [AVX2,SSE4,Scalar] too, maybe your CPU doesn't have these instruction?

well turns out you can't do transcoding without libjpeg

still larger tho

Your build might have compiler optimizations tuned for CPUs with AVX2, knowing that AVX2 supersedes AVX to SSE.

where is -j documented?

aom moment right here

`-v -v --help` shows everything

who thought that was a good idea

not me

well i guess i have to do both -j and not -j and compare for each

unless that's already a thing

hidden inside -v -v

tbh black and white manga stuff should really be something like png but some people think otherwise i guess

https://github.com/libjxl/libjxl/issues/248 https://github.com/libjxl/libjxl/pull/250 Works fine after this

great it was totally this

🤦 i was a few hours too early

hah 9 hours ago

how to fix this?

by assigning a profile to the monitor using colord (there is a `colormgr` command to do that, or various GUIs), ideally one created by profiling the monitor, but it’s also possible to create one from the monitor’s EDID

doing it from the EDID is not perfect but likely better than nothing

Error gone away after installing color-kde package. But it still crashes 😩

I have one jxl in one folder and another jxl in another folder. Selected both folders in flicker-test tool. Am I using it wrong?

a crash seems like an extreme reaction

if you build it with CMAKE_BUILD_TYPE=Debug (or RelWithDebInfo) and run it in gdb, does it seem to give a useful stacktrace?

(can just edit it in `CMakeCache.txt` and rebuild, it should take the new value into account)

(or you can use a separate build folder)

sharpness

for %i in (C:\Users\User\Documents\bil*.jpg) do cjxl -j -q 98.36 -s 7 --progressive --faster_decoding=1 --use_new_heuristics --epf=2 %i %i.jxl

06052021

deblocking

for %i in (C:\Users\User\Documents\deblocking3\*.jpg) do cjxldeblocking -j -I 0.881 -s 1 -q 91.732 --gaborish=0 --use_new_heuristics %i %i.jxl

libjxl v0.3.7-171-gc12aec2 win_x64 2021.06.28

It needs to recognize jpg jpeg JPG JPEG and PNG png

do you know how to do

also you need to share sharpdevelop or visual studio project full folder with the solution .sln and the debug

and i'll make a backup copy on my github profile

can anyone help?

who wants to do?

<@456226577798135808> do you know how to code this?

sorry for pinging

Sorry, it's too hard for me ¯\_(ツ)_/¯

I just stumbled across this awesome site http://graficaobscura.com/

it's barely been updated since the 90s, but its full of neat computer graphics info

oops I meant <#806898911091753051>. close enough...

`Reading symbols from ./flicker_test... (No debugging symbols found in ./flicker_test) (gdb) `

I built it using, `cmake -DCMAKE_BUILD_TYPE=Debug -DBUILD_TESTING=OFF -DCMAKE_BUILD_TYPE:STRING='None' -DCMAKE_INSTALL_PREFIX:PATH='/usr' -DJPEGXL_ENABLE_BENCHMARK:BOOL='false' -DJPEGXL_ENABLE_FUZZERS:BOOL='false' -DJPEGXL_ENABLE_PLUGINS:BOOL='true' -DJPEGXL_ENABLE_VIEWERS:BOOL='true' -DJPEGXL_ENABLE_GIMP_SAVING:BOOL='ON' -DJPEGXL_FORCE_SYSTEM_BROTLI:BOOL='true' -DJPEGXL_FORCE_SYSTEM_GTEST:BOOL='true' -DJPEGXL_FORCE_SYSTEM_HWY:BOOL='true' -DJPEGXL_WARNINGS_AS_ERRORS:BOOL='false' -DCMAKE_C{,XX}_COMPILER_LAUNCHER=ccache -DCMAKE_C_COMPILER="clang" -DCMAKE_CXX_COMPILER="clang++" -Wno-dev -B jxlbuild -S libjxl -G Ninja`

I have tried building without ccache. still it doesn't make debug symbols.

the setup is ready

i need only to code the code

there is a `-DCMAKE_BUILD_TYPE:STRING='None'` in there, maybe try editing `CMakeCache.txt` and rerunning ninja?

this way, no risk of overriding

please i want this program

i will insist forever

`#0 0x00007ffff65f38a8 in ?? () from /usr/lib/libQt5Core.so.5 #1 0x00005555555805fe in QHash<QString, QHashDummyValue>::findNode (this=0x7fffffffb000, akey=..., ahp=0x7fffffffab9c) at /usr/include/qt/QtCore/qhash.h:946 #2 0x00005555555804af in QHash<QString, QHashDummyValue>::insert (this=0x7fffffffb000, akey=..., avalue=...) at /usr/include/qt/QtCore/qhash.h:773 #3 0x000055555558002d in QSet<QString>::insert (this=0x7fffffffb000, value=...) at /usr/include/qt/QtCore/qset.h:222 #4 0x000055555557d9a2 in QSet<QString>::QSet<QList<QString>::iterator, true> (this=0x7fffffffb000, first=..., last=...) at /usr/include/qt/QtCore/qset.h:66 #5 0x000055555557a8a8 in jxl::FlickerTestWindow::FlickerTestWindow (this=0x7fffffffdab8, parameters=..., parent=0x0) at /mnt/data/libjxl/tools/flicker_test/test_window.cc:70 #6 0x000055555556d11a in main (argc=1, argv=0x7fffffffdcf8) at /mnt/data/libjxl/tools/flicker_test/main.cc:16`

uh, very strange

does this fix it? ```patch diff --git a/tools/flicker_test/test_window.cc b/tools/flicker_test/test_window.cc index 96dbb88a..3112e08f 100644 --- a/tools/flicker_test/test_window.cc +++ b/tools/flicker_test/test_window.cc @@ -67,10 +67,12 @@ FlickerTestWindow::FlickerTestWindow(FlickerTestParameters parameters, auto originalImages = QSet<QString>::fromList(originalFolder_.entryList()); auto alteredImages = QSet<QString>::fromList(alteredFolder_.entryList()); #else - auto originalImages = QSet<QString>(originalFolder_.entryList().begin(), - originalFolder_.entryList().end()); - auto alteredImages = QSet<QString>(alteredFolder_.entryList().begin(), - alteredFolder_.entryList().end()); + const QStringList originalFolderEntries = originalFolder_.entryList(); + QSet<QString> originalImages(originalFolderEntries.begin(), + originalFolderEntries.end()); + const QStringList alteredFolderEntries = alteredFolder_.entryList(); + QSet<QString> alteredImages(alteredFolderEntries.begin(), + alteredFolderEntries.end()); #endif auto onlyOriginal = originalImages - alteredImages, ```

Yes it fixed the crash.

thanks for reporting this, it’s fixed in the repository now

Does flicker test work on jpg images? It just closes after I click start.

strange, from reading the code, it seems that it should at least try

ah, no, I missed the critical line

apparently it’s `png` only

So it works only with PNG images... Thanks.

I'm trying to add noise synthesis to `jxl_from_tree`. What's the difference between `noise_level` and `intensity` in `struct NoiseLevel`?

I actually have local code doing that, but I don’t quite remember why I didn’t create a PR for it

there was something I was not satisfied about

What is the image compression codec that you are recommending?

jpeg xl and webp2 are my favourite

webp2 image have very few sentiment if not all

jpeg xl is about quality

different stuff

the bottom end is not very easy to get right with so few LUT entries but I think this looks rather encouraging

This noise is not very random, you can add more low and high frequencies?

nah, you can't actually control that

we spent some effort tuning the noise and IIRC this is the best we could come with for actual photographs - adding it to a gradient is likely not very representative of actual images 😛

yeah, this is kind of a worst-case scenario 😁

It looks like a linear noise, it's impossible to have a Gaussian noise? like this photoshop gaussian noise.

it's uniform noise with a box filter

(well, a subtracted box filter)

<@!604964375924834314> how does it look like on a real gradient?

If you need a clean gradient.

ugh, I can still see the banding on it...