JPEG XL

the 5 MB upper bound is likely to be more limiting

(for an image that takes up the whole 4096×4096, it’s about 2.38 bpp)

found it! https://x.com/NolanOBrien/status/1281639466688380928

ohh i thought you meant the file size

we need a `cjpegli -q twitter` 🙂

so basically up to 2236x2236 (or anything else that is 5 megapixels) the limit is 8 bpp, and for a larger image the limit is just 5 MB

`-X Conform output file to X/Twitter limitations` `-T Conform output file to Telegram limitations` `-D Conform output file to Discord limitations (Extremely large images)`

I don’t think there is a 5 MP limit?

The 5 MPx limit is because the Max is 8 bpp

but if you have less than 8 bpp, you can have more pixels than that.

So say that you get your desired quality at 2bpp, then you can have 20 MP

I remember seeing certain images that look one way up close and different from afar: for the close up image only the high frequency information is displayed, and for the afar image only the low frequency image is displayed. And these images are blended together into a single image, and the effect will work. I was just watching a video about how JPEG works and now I'm wondering if you could take two JPEGs and replace only the high or low frequency information of one with that of the other, to end up with an image that also exhibits this effect. Initial thought is that 8x8 pixel blocks are already pretty high frequency in terms of this effect, so maybe with JPEG XL you could create bigger blocks and make it work

Well, the limit is always 4096 x 4096 regardless of what you throw at Twitter, so not *that* many more pixels

Reminded me of this https://discord.com/channels/794206087879852103/1042307875898408960/1285572825561825393 I used the same bug to 'extract' the HF image from the DC by adding a 2048 black block before the real image

still, up to 16 MP isn’t so bad

phone pictures are typically well within that limit, at least when binned if quad bayer

Surprisingly legible (Don't mind the fork I stuck in his crotch....)

Also explains why that screenshot is about 60% larger than usual, a lot of high frequency detail

`652 bytes` can anyone make a smaller JXL lossless?

no. libjxl doesn't yet have the optimizations ported over from libwebp

you would need to do it by hand or write a new jxl encoder that has similar optimizations

You could make it smaller by reducing the amount of entropy in the image and making the pixels a little bit more predictable.

Reducing unique colors and making the patterns in the image more precisely consistent.

Not lossless

But the original image has a lot of randomness in it that makes it less compressible

“Not only is it lossless I also removed the noise” /s

you could try `-e 12` from Monad's fork https://github.com/monad0/libjxl/tree/cpu-heater needs to be built from source tho

oh boy...

ooo nice!

(but that's not lossless)

from Monad (on JXL server)``` JPEG XL encoder v0.11.0 850f6cdf [AVX2,SSE4,SSE2] Encoding [Modular, lossless, effort: 12] Trying 760320 configurations, please be patient ...```e11: 0.492s, 1779 B e12: 18m10.482s, 1752 B

still looks good for my purposes

Funny enough lossless noise reduction has been attempted for the purpose of image compression

Like noise reconstruction?

Nope! Has to do with signal deconvolution https://www.sciencedirect.com/science/article/abs/pii/S0923596515001630

How is it reversible ?

I believe it's an averaging filter, like a more advanced version of the type of prediction filters employed by png

https://link.springer.com/article/10.1007/s11042-019-08371-w

But noise is random, you can’t just get rid of entropy data

Okay the smallest jxl file I could produce without making it by hand is 743 bytes

You know what is the craziest thing though? I got it from libjxl 0.8, NOT libjxl 0.11 **-e 12**

Here are the numbers: ``` libjxl 0.8 --- 743 bytes libjxl 0.9 --- 774 bytes libjxl 0.10 -- 766 bytes libjxl 0.11 -- 760 bytes libjxl -e 12 - 756 bytes ```

I waited over 20 minutes for `e 12` to finish... Only for libjxl 0.8 `e 10` to beat it in like 1 second

<:kekw:808717074305122316>

Here is the e12 file... <@794205442175402004> <@263300458888691714> any ideas why this file is bigger?

Probably different palette order that happens to work better on that image

<:Stonks:806137886726553651>

https://discord.com/channels/794206087879852103/794206170445119489/1288275096502407268 I haven't studied this topic much myself, that after adding 1 pixel, it began to compress 9 bytes better, but as if there is a connection here. There are a lot of anomalies when compressing very small images. It's like CJXL cut off some ways. like local minimum

It is certainly not exhaustive, not even looking for a local minimum.

From your testing, `-e 10` is far from exhaustive. But first you forgot one crucial step, optimizing your png inputs. When I reduced the colorspace to 1-bit grayscale, both images became 196 bytes. So in the end, the dot didn't really matter. As I've shown in my testing, even `-e 12` with 760320 iterations isn't the catch-all people make it out to be. A smarter encoder is the logical step forward.

huh, still loses to WebP

WebP makes some very smart decisions for paletted images. In theory Jpeg XL beats WebP (and it usually does). But libjxl's encoder isn't as smart and mature for this kind of content.

I've seen quite a few cases where WebP beats Jpeg XL for pixel art and the like

I know it’s been over 3 years since you touched this but do you know what licence it’s under ?

I think it's clear that it was put in the public domain, do you need anything more than that (e.g. relicensing to MIT-0 or similar)?

Nah public domain is fine , that means anyone can do anything right?

I'm always surprised by how good Real-ESRGAN can be to restore shitty JPG artifacts

sure, it looses a bit of clarty for the background pattern but still, all others solution I tried still had JPG artifacts in the output (waifu2x, jpeg2png)

You know what might help for really bad images like that? Shrink the image first to eliminate high frequency distortion introduced by jpeg block artifacts

maybe someone should train a model that is specifically trained for decompressing JPEGs

i think there are some ESRGAN models for that

Some? They are super plentiful. It's almost like the 'Hello World' for training AI models.

I wonder how complex an FPGA version of fjxl would be

In about 20 weeks, we're supposed to plan and execute our own FPGA project, and I thought making an encoder could be pretty fun.

I guess the risk is if it turns out not to be practical for one reason or another

maybe a safer bet would be something that can be incrementally improved

I was thinking about that the other day. If cameras use hardware encoders that means you're stuck with that encoder forever until you buy a new camera. An FPGA implementation could be a middle ground between needing an expensive CPU for software encoding and never benefiting from future improvements with a hardware decoder. Idk what the current state of DSL cameras is though, it was just a passing thought

FPGAs are hella expensive 😵‍💫

hm, yes. If I manage to get ahead of schedule, I'll explore the idea beforehand to see if it may be viable, I guess.

Otherwise there's always the tried and true, implement your own CPU.

It seems to me like mass producing the same FPGA chip for all your cameras which can support a bunch of formats programmatically would be cheaper than developing hardware encoders for different formats?

An FPGA needs to be MANY times bigger than the equivalent dedicated chip, and draws a lot more power. Doesn't make sense unless you truly need a lot flexibility or if your product is low volume enough that a custom chip would be too expensive to develop.

But all cameras use custom camera chips already, so.

Fun fact: Nvidia's G-SYNC module (now discontinued and replaced by a custom chip by Mediatek) used an FPGA that cost them $500 per module!

> An FPGA needs to be MANY times bigger than the equivalent dedicated chip Huh, never knew. I always assumed the FPGAs I was working with were bigger just because they didn't really need to be small

You can get small FPGAs. But an equivalent fixed circuit would be much smaller still.

today, it's over $3000 per unit on digikey.

you can find cheaper ones now

lots of... OK ones flooding the market. no idea if they would be good enough for jxl, peolly would be

did you tested some of them and could recommend one pls ?

It's been awhile since I've been in the AI upscaling space

Let me dig through some discord servers real quick

everyone uses span or plksr these days

so tested 8 models here <https://openmodeldb.info/?t=compression-removal+jpeg+restoration+input%3Aimage+scale%3A1> many don't handle text correctly (my use case was a quite hard-compressed file): https://cdn.discordapp.com/attachments/673202643916816384/1294756926797905931/original.png https://cdn.discordapp.com/attachments/673202643916816384/1294756927170941050/jpg_40-60_by_Alsa.png https://cdn.discordapp.com/attachments/673202643916816384/1294756927502286848/jpg_40-60_by_BlueAmulet.png (the best one imo) (this is a small part of the image) the best model for my needs was <https://openmodeldb.info/models/1x-JPG-40-60>

original pic: https://cdn.discordapp.com/attachments/673202643916816384/1294757248513343563/original.png

Oh man! I forgot how fun AI upscaling is. Sorry, I haven't been able to do any AI research yet. I was hospitalized 😅

Now I have to catch up on 2 late homework assignments and email my professors. I'll take a look at these when I have the time.

no worries lol

I always thought upscale meant "bigger output" but they say upscale even for 1x <:Thonk:805904896879493180>

This is the best output I got, with some small tweaks to get the red bg to pop out a bit more since AI loves to smooth it over

I used this AI model: https://discord.com/channels/547949405949657098/579685650824036387/1244906401495515148

Btw this discord server has a lot of good models/information if you want to dip your toes more into AI stuff

There might be newer versions of the model I used but this more of an example of the quality of stuff being made nowadays

Here is a comparison

Typically I find that a well-trained 4x model downscaled to 1x, beats 1x models quite regularly. So I just took the best anime 4x model I had laying around on my computer and gave it a go.

I don't have access to that channel which server is it ?

but indeed, impressive result, thank you [⠀](https://cdn.discordapp.com/emojis/853506500088692747.webp?size=48&quality=lossless&name=pepelove)

I'm really new to this "AI upscale" world just got my hands trying a few stuffs on chaiNNer other than that, I have almost no knowledge lol

https://discord.gg/cpAUpDK

After some more tweaking, here is my final result. I manually merged the best qualities of 2 AI outputs: 4x_umzi_digital_art_rplksr_v1 (best at handling sharp lineart) and 4x_IllustrationJaNai_V1_DAT2 (best at handling the red bg & certain small details)

all we are saying is give jxl a chance

may I know how you do that?

The AI processing part was easy with ChaiNNer. The hard part was deciding what details looked "better" in either image. I then manually painted out details I didn't like. These difficult cases usually require a human touch. It took around half an hour.

oooof

at the JPEG meeting last week, we discussed a lot about AIC-3: https://jpeg.org/aic/aic3.html

one of the things we're discussing is to better clarify, and to extend the scope of AIC-3

the illustration we used before was a bit too simplistic, and not really correct. AIC-2 does not cover a range, it's a method to determine a specific quality point (visually lossless in the viewing condition of flickering)

This is the new figure (currently under discussion, WIP) that we'll use to illustrate the AIC-3 goal and scope

Fidelity oriented please...

https://arxiv.org/abs/2410.09501

Congratulations on your publication :)

Are the jpeg meetings available online?

nope

https://www.reddit.com/r/GalaxyS21/comments/13ez3wh/how_to_view_hdr_photos/ldusu78/ Most knowledgable JPEG XL adversary

> HEIF files are still usually 50% smaller than JPEG XL files at the same resolution/ color gamut/ image fidelity <:galaxybrain:821831336372338729><:galaxybrain:821831336372338729><:galaxybrain:821831336372338729><:galaxybrain:821831336372338729><:galaxybrain:821831336372338729><:galaxybrain:821831336372338729><:galaxybrain:821831336372338729><:galaxybrain:821831336372338729><:galaxybrain:821831336372338729><:galaxybrain:821831336372338729><:galaxybrain:821831336372338729><:galaxybrain:821831336372338729><:galaxybrain:821831336372338729><:galaxybrain:821831336372338729>

Watch me beat HEIF with JPEG Legacy

fella gives no number that can *resist* scrutiny but insists you must not believe your lying eyes

in other circles many people are defaulting to e10 for large images and consequently claiming avif is far faster for more acceptable results. maybe all efforts above e8 should be developper-only toggles with warnings meant to deter their use - or effort adapts to the actual image resolution (ie slightly less costly conversion techniques if its 30+ mpixels, slightly more expensive if its small images)

its crazy capturing 2K+resolution screenshots at e10 while youre playing too

Not to forget that e1 is faster than most PNG screenshot implementations and will result in smaller images

When people see a big number, they'll use it. It's why a few have bluescreened their computer by hearing of effort 11 and using it without any regard of the expert option setting

e12 when? 🤔

I'm not immune to this, I went through the pain of building a custom libjxl fork when I heard effort 12 was a thing 😂

maybe integrations need better documenting? distance is the setting that makes the real difference, effort just tries being slightly more efficient storage-wise by e7 is still the best, especially for lossless jpg->jxl

Now: https://discord.com/channels/794206087879852103/794206087879852106/1292756506449543200

e7 is under 50ms for sub 20 megapixel images but can last a minute for e10-11 for barely 5 kilobytes less

I'm assuming lossy but even then that's not right. Oh, unless you mean JPEG transcoding

I honestly wonder where he got his infos from. Nearly every sentence is wrong

do it

I'll do a quick debunking but feel free to add/correct anything after

2am here :P

just call him a liar and leave it there lmao

lmao

actually, I wonder...

Wouldn’t AVIF be better than HEIC?

correct. conversions already guarantee a 20% filesize reduction at almost all efforts, giving up quasi-instant conversions for an extra few kbs saved is nuts for anything but say a comic chapter's first image(s)

I use e8 because I don’t care about speed lol

Yes

I had a thing typed out about bitdepths but simplicity is sometimes best... And now we've both posted a few mintes apart xD

Still waiting for a proper way to convert Apple HEIC with a bunch of aux data and a gainmap into a native HDR JXL with extra channels storing the extra data

Like HEIC on iPhone sometimes adds depth maps and segmentation maps

Adjusting effort based on megapixels like lossless or lossy based on file type has been talked about for quite a while, but since people will just crank it up anyway, maybe effort 10 should be made expert too so that multithreading can't be disabled unknowingly

lol

Please do, I am not able to access reddit, but this is brazenlu wrong

Make e1 the default so everyone will be amazed at blazing fast encodes 😎 /s

Maybe e6?

I agree about e10 being expert

I also had that typed out "You are right about JXL being fast though, we can hit 300 MP/s"

heck yeah

Is this just lossless or what about VarDCT?

Well HC was talking about transcoding, I was thinking lossless but everyone else probably means VarDCT

Me out here always only thinking about lossless

I still want lossy modular tbh

Similar speeds to England ?

Similar speeds to Lossless Modular?

No, about half

So really slow!

https://tenor.com/view/lordfibo-lordzfebz-gif-25361200

Wow, how can people sound so confident while being so totally wrong. I suppose the Trumpification of discourse is spreading even to image format discussions now...

Hello, I am a jxl noob phone photographer requiring some help I was trying to convert some RAW DNG pictures taken from my phone which could output an HDR gamut but still coupled with an SDR color space so I guess I would still have to export in 32bit anyways ? I noticed that compression did not work in 32bit mode so I was wondering if this is an expected behavior or a lack on feature from Affinity Photo or could it be something else ?

What exactly do you mean by "compression did not work"? Do you have slider for effort?

By moving the quality slider I could not see any difference regarding file size

Even at quality 100%? Can you export an image at quality 0 and one at 100 and share them here?

I'll give it an other try just to make sure

are you changing the quality at both locations in the screenshot or just one of them?

So my original file is a 23MB DNG picture, and I can confirm exporting in JXL 32bit will give me a 42MB picture regardless if I export with the slider at 0% or 100%

I am not sure what 'both locations' implies here

i was refering to the quality slider at the top and the quality drop down at the bottom

Oh these two seems to be always in sync regardless which one I use (Edit:I've tried it once again and the slider would effect both sections of the tool, they are synced up)

Try actually exporting, that value is just an estimate

The 32-Bit pixel format means you're trying to export Float JXL files instead of 16-bit Integers that DNG uses. Try changing that first

I smell AI

I did, and the predicted size was accurate in every cases

Indeed compression works when using the 16-Bit format. But I thought those DNG files were in a 32 Bit format. Thanks

It might be forcing lossless for 32bit

Bitdepth can be confusing because some use the total and some use per channel (3 x 8 bit = 24 bit, 3 x 32 bit = 96 bit)

It'd be strange, but maybe, and then this would be causing it https://github.com/libjxl/libjxl/issues/3511

depends

Can I use the embed.moe service on other Discord servers to preview my JPEG XL photos? I assume the creator is in this server

I thought I'd ask for bandwidth reasons

yes

Cool, thanks :)

I'm making a viewer for an obscure image format. I discovered a bad quirk of Windows 11's round corners. They cut-off a part of client area inside the window.

analogies I did not expect to run into

(trying to see if I can learn enough to slightly clean up this concert recording https://youtu.be/NDooA7NgA-Y?t=20m35s – I love the piano at the start of Part Two)

fft is about waves optics manipulate light waves checks out

their example to demonstrate temporal vs. frequency precision is actually quite nice

I'm by no meaning an audio mastering expert, but maybe one of the variants is acceptable. NR 1-4 in the names means passes of noise reduction https://we.tl/t-Q9qqlGgfiW P.S. If only they did not do dynamic noise shaping, and left a second with only noise inside..

my attempt https://youtu.be/d5tD1bDZkHY

Removed residual hiss https://we.tl/t-GHXYGcm9eg

https://news.ycombinator.com/item?id=41982698

https://www.androidauthority.com/google-apple-hdr-photo-standard-3495035/

I wonder how many ISO standards became purely standards in name only.

I feel like out of all formats, using a jpeg to store extremely dark images is the worst possible one...

Here's an easy way to make HDR photos work better on Android and iOS JPEG XL adoption! <:JXL:805850130203934781>

>>> "The challenge for apps is that there are multiple ways to encode gain map metadata in an image file. Google’s Ultra HDR, which uses Adobe’s encoding scheme, is one of them, but Apple has its own version. That means that platforms like, say, Instagram would need to not only handle both schemes but also potentially convert between them if they wanted both Android and iOS devices to see any given HDR photo the same way. Fortunately, there’s a better solution to this problem, one that has already been adopted by both Google and Apple: **the ISO 18181 standard**

I would rather have an heic than a jpeg with gain map

If only 😢 😭

Apple uses 8-bit heic with a gain map so you get the worst of both worlds

Seriously? wtf... :(

When are gain maps going to be replaced by parametric functions?

Is parametric function like PQ and HLG?

It’s the worst, still trying to find a reliable way to convert “HDR” HEICs to any other format with hdr metadata

No, those are transfer curves. I mean a tone mapping method that is adjustable with parameters

Ah, so like a custom polynomial?

https://ssoready.com/blog/engineering/truths-programmers-timezones/ > `America/Nuuk` does daylight savings at -01:00 (yes, with a negative)

ಠ_ಠ

The world can be a weird place sometimes.

Daylight savings time is evil

And it's absurd that everyone plays along with such a ludicrous and sadomasochistic game

We just need ISO 8601

I love daylight savings time, it's horrid when you go to do construction and it's still really dark, always better when daylight savings kicks in and now we can start work in sunlight again

But... Nothing is stopping you from just changing the time you show up to work. That's what you're doing anyways, you're just doing it in a far less convenient and more idiotic way with DST

Unless the value you get out of showing up to work an hour later is so low that it's not worth changing your schedule for without everyone else changing their clocks for you.

Have you ever done computer programming?

yes

Ever done date handling?

yes

And you still like DST

I also live like a normal person who has a life, and worked jobs such as construction and shop/fabrication work

it's good in theory, but having everyone keeping in sync is still important

deliveries, shopping etc

though most people don;t have a sleep schedule worth anything any more, so shopping isn't really an argument any more, just be open 24/7 [av1_kekw](https://cdn.discordapp.com/emojis/758892021191934033.webp?size=48&quality=lossless&name=av1_kekw)

It's easier for everyone to keep in sync if you don't force people to set their clocks back and forth, that's just sadistic.

I can't believe that people obey and comply with that just because they're told to do it

People are way too accustomed with just going along with what they are told to do, even if what they are being told is abusive and evil

I mean, you don't have to, as you said but in reverse, just don't do it.

I don't change my clocks lol

But everyone else still does.

I keep the clock in my car on normal time for example

But all my online devices automatically change and everyone else complies and uses stupid-time

So it doesn't really matter that much what I do if the majority of people just do whatever they're told to do

It still inconveniences me a lot

maximising sun hours is pretty important for all sorts of people and jobs, and keeping sync is also pretty important, when people get out of sync it makes everything else painful

But it would inconvenience us all a lot less if people ignored what they were told to do and think for themselves more

screw it, lets just all use UTC

i give up

like even basic things such as shopping become actual issues

I don't mind that idea either, at least it doesn't change

But the dumbest thing in the world is forcing people to fiddle their clocks back and forth

would make international travel easier

Instead of just, I dunno, actually starting your day in alignment with the natural world using your own decision making brain

No one has to force everyone to be inconvenienced in order to benefit from sunlight

We're supposed to have brains :(

Sorry for rambling, I'm just sick of everyone always following orders

No one takes responsibility for their own actions and their own decisions

No one is even conscious that they're choosing not to make a choice

People in general are afraid to own their own actions and decisions.

Like when the AVIF/webp team lead Jim Bankowski doesn't say "I decided" to remove jxl support from Chromium, instead he uses the word "we" without defining who that even is. 😂

Despite him being the head of the team responsible for deciding that

honestly would be wild, while we are at it, return to proper sunrise and sunset cycles. boss: "I want you to be at work half hour past sunrise" Sunrise: 6:32.48 you: 7:03.00 boss: Fired.

well in UTC it would probably be like 10:00 for sunrise lol

I doubt, if your boss is that cool, that you would be fired :)

I am still sad that the French revolution failed to make decimal time a thing.

https://en.m.wikipedia.org/wiki/Decimal_time

10 hours per day, 100 minutes per hour, 100 seconds per minute.

Then you can do stuff like 1.56789 days equals 1 day, 5 hours, 67 minutes, 89 seconds. Quite convenient.

that makes literally zero sense. Daylight savings kicks in as we go from winter to summer, that is, when the sun is rising earlier and earlier anyway

im talking about the other way, daylight savings time itself is necessary, so if you got rid of the change you would keep it to be a "permanent daylight savings"

eh?

why not use regular time year around?

there is nothing necessary about daylight savings

the point of daylight savings is to make sure you maximize daylight during working hours, so you would use that as the basis

but summer already has way more daylight

prioritizing productivity.

??

how?

the vast majority of production stuff happens in "summer" times, farming, construction, fabrication etc.

yeah but because there's already way more daylight during the summer, you're not actually gaining anything

and professions for which daylight is important, they get up and work with the sun regardless of what the clock says...

Spain is in the same timezone as Poland, and I can tell you that people in Poland and Spain don't get up to work nor go to bed at the same wallclock hours.

in countries where people sync up work periods it matters a lot.

also > yeah but because there's already way more daylight during the summer, you're not actually gaining anything may be true for locations close to the equator, but as you get further north or south the differences start to become a lot more apparant

literally the opposite of the truth :p

it's up north where summer has crazy amounts of sun

I live 60°N and for a decent while, we don't even have *night time*

it doesnt matter for summer but it matters for everything else

for nothern hemisphere

surely youve seen this chart

https://images.squarespace-cdn.com/content/v1/5bfc8dbab40b9d7dd9054f41/1550725976319-CL4T63M326DXBGFQPW55/daylight-saving-time-explained_5096f7937aee4.png

yeah I don't really see the point from that chart

there will be too little sunlight during working hours in jan-mar/nov-dec

the only other way to solve it is to move working hours

eh? But that's during the winter, which is regular time.

That's what I want to have year round

it's obviously more different the higher/lower you go

Winter time all year round makes a lot of sense. I would prefer that too.

that's the china approach

The EU was supposed to abolish DST but it kinda fell off the agenda when COVID hit

Though it was up to each country whether they wanted permanent winter time or permanent summer time.

for me there would be ~~too much light in the summer~~ no light in waking hours of winter

actually this is just ownedf

here is how it looks for me

The justification for DST has always been pretty flimsy. The main argument was that it allowed saving some energy for lighting since in summer you get more daylight after work. Maybe that was a reasonable-ish point back when we used incandescent light bulbs...

i'm mainly seeing it as sunlight when you wake up

hence the working/productivity

There are a lot of issues with permanent DST in general. The biggest one just being safety. I believe that was an actual issue In the past in usa

yeah it would still be pitch black at 8am here

Swedish winter is fun

pitch black when you go to work

pitch black when you go home

Sometimes I wonder if it wouldn't make more sense just to split daylight savings into halfs or quads. So clocks change 30 minutes or 15 minutes at a time. a 15 min change would be fairly unobtrusive to sleep schedules.

permanent winter time, I say

DST is just a form of gaslighting, ironically

we just convince ourselves to do everything one hour earlier because “omg it’s already [H+1] o’clock”

this kind of makes me want to make one of those “stop doing maths” memes but for DST

“wanted to wake up one hour earlier anyway for a laugh? we already had a tool for that: it’s called ‘setting your alarm clock one hour earlier’ ”

Permanent winter time, and if needed change working/school hours. In any case it makes little sense to have fixed working hours all year round, at least for work that requires daylight (like construction) or that has a big seasonal component (like farming).

having to change hours

call it daylight savings

https://en.wikipedia.org/wiki/Swatch_Internet_Time

Not everyone, for the last 20 years the amount of places with it decreased significantly.

I’m going to start a political party for the abolishment of DST in Australia

There's always some nonsequitor excuse to keep dst for some reason

The only reason such an idiotic idea ever took off the ground is because it was forced upon people with the excuse of "there's a war, we need to conserve electricity" or some stupid crap

War justifies everything

do you wanna go to war, Balakey?

(reference: https://youtu.be/Dd7FixvoKBw)

it took me too long to realize DST was referring to daylight savings time and not discrete sine transform considering this is <#794206087879852106> lol

It's pitch black at 8am right now and dst closing tomorrow will fix it

Maybe instead of changing the clocks back and forth and inconveniencing everyone, people could just change their own schedule if it's convenient to do so.

that's even more inconvenient

I'm more used to changing schedules so for me mentally it's better without dst

dst is already automatic having to manually change alarms and schedules is worse

I still have devices that require manually changing the time when dst starts/ends

If dst would be abolished, it would have been fixed today already 😉

but then the sun would be up too early during the summer

We should go back to waking up earlier in summer and sleeping longer in winter, imo.

we should keep summer time imo DST in winter -> go to (7h20 am) and come back from (17h30) work in the dark no DST in winter ->come back with sun yeepeee

I just hate not having the sun at home

Anyone know some lossy transforms that exist (that benefit compression) which are asymmetrical, in the sense that they take more computation to encode than to decode?

(And not a “fractal” method)

the only thing that comes into mind are the usual CfL implementations, in that you'll need to try out multiple alpha values (and sometimes beta) to get the one with the smallest error

decoding just involves using the provided values, so that's significantly faster

Hmm yeah but that’s different than like a “compex DCT” encode and a “simple DCT” decode which was what i had in mind

yeah, I don't know of any asymmetrical-in-time DCT

the asymmetry usually comes from the encoder trying out multiple transform sizes/types

or different quantisation

Yeah which is why i said “and not a fractal method” lol because that’s the king of trying out different things

maybe neural net AI (TM) encoders have the answer?

`sb-qp-sweep` my beloved <:KekDog:805390049033191445>

You know what will really boost JPEG XL adoption? JPEG XL needs to be an ✨✨ AI ✨✨ first format. We need to integrate ✨✨ generative AI ✨✨ and ✨✨ machine learning ✨✨!

modular but add transformers and relu somewhere

gen AI context trees?

now you're onto something

data-driven predictors

AI assised MA trees has been an idea for a while

I'd like to propse a new extension to the standard, JPEG XL + AI, it's just like JXL, but there is a metadata slot that contains a prompt to 🌈 generate 🌈 an "enhanced" version of the image when forwarded to my AI model, only €20 per prompt ||Nothing in this proposal shall be considered an offer of product or technical services, a gurantee of quality, or a warranty of fitness for any purpose or of merchantability||

Sorry but I don't think I'm going to invest, for these main reasons: 1. Not enough 🚀🚀🚀 rocket emojis 🚀🚀🚀, how do I know if it's going to be 🚀blazingly fast🚀? 2. How can you use ✨AI✨ without the ✨✨✨sparkle emojis✨✨✨? 3. Not written in 🦀rust

<:rust:854096515727753266>🦀<:rust:854096515727753266>🦀<:rust:854096515727753266>🦀

My bad, I'm never gonna get VC funding for the future of ~~my wallet~~ image encoding at this rate

Here, let me fix it for you 🚀🚀🚀 JPEG XL IS NOW AN AI FIRST IMAGE FORMAT I would like to propose a new 🚀 **blazingly fast, super secure** ✨ AI based ✨ extension to JPEG XL, written in <:rust:854096515727753266> Rust! 🦀 This extension adds a metadata slot that uses ✨AI✨ to generate an enhanced version of the image when forwarded to my AI model, only €20 per prompt. It's 🚀blazingly fast and 🔒super secure (trust me).

VCs: 🤑🤑🤑💸💸💸

<@794205442175402004> what do you think of our proposal?

can we have the $3 billion of funding to set this up? Please????

Needs more buzzwords. Where's the VR/AR in this? And the blockchains?

FINE (your demands mean i had to raise prices a bit) 🚀🚀🚀 JPEG XL IS NOW AN AI FIRST IMAGE FORMAT I would like to propose a new 🚀 **blazingly fast, super secure** ✨ AI based ✨ extension to JPEG XL, written in <:rust:854096515727753266> Rust! 🦀 This extension adds a metadata slot that uses ✨AI✨ to generate an enhanced version of the image when forwarded to my AI model, only $200.00 per token in prompt per pixel per image generated. It's 🚀blazingly fast and by using revolutionary **BLOCKCHAIN TECHNOLOGY** your photos are 🔒 *super secure* (trust me). This technology also allows you ✨JPEG XL with AI ✨ images to be compatible in the AR VR Metaverse Spatial Computing Next Generation Oculus Vision Pro Max Ultra virtual reality dimension!!! Oh, and it runs in the !! CLOUD !! as a SaaS distributed edge computing matrix network protocol!

You gave me a flashback to a VR game with a cryptocurrency no one used, until one day it's value increased by 10000x bringing a swarm of cryptobros to the door while everyone ran to sell sell sell

We use ethereum gas to provide the AI processing, in order to ensure the *consensus* in how your image should look <:YEP:808828808127971399>

It also allows you to automatically create an NFT of your image <:YEP:808828808127971399>

Maybe I should go visit Hello Games and give them a stern talking to....

They uploaded 8K PS5 screenshots for the last patch notes on their site, even though there's no fullscreen button and they can only take up half of the screen at most

they’re just waiting for jxl in chrome before taking the step to lossy

(just in case: I don’t actually know that, I’m just joking)

Progressive 206 Partial Content request

In the release log where you choose what update to look at, they have preview images underneath the paragraph describing it... All 4K images, 10MB+ each, no lazy loading, no lube https://www.nomanssky.com/release-log/

this bloody hurts

The one at the top has the 8K images

hhhhhhhhhh ```File: No-Mans-Sky-PS5-Pro-Screen-1.png (57291841 bytes) chunk IHDR at offset 0x0000c, length 13 7680 x 4320 image, 32-bit RGB+alpha, non-interlaced chunk IDAT at offset 0x00025, length 64778 zlib: deflated, 32K window, fast compression```

this took so long to load on my phone [av1_kekw](https://cdn.discordapp.com/emojis/758892021191934033.webp?size=48&quality=lossless&name=av1_kekw)

Please Sean, Alpha too? I can take your updates but not your memory usage

I need an ide ssd for my crappy netbook so i can try and load this page in it [av1_kekw](https://cdn.discordapp.com/emojis/758892021191934033.webp?size=48&quality=lossless&name=av1_kekw)

its having issues loading on my bloody desktop wtf is this

I always had issues loading the big updates because of all the videos loading and autoplaying at once causing my GPU to hit it's HW decode limit

Now they're going after the CPU with 8K images at the same time, god help us

I might *actually* @ him or find a contact or something, because jesus christ they need to learn the basics of optimization

15MB off and half the decode time with e1, can halve the filesize but add half the decode time with e7 Naturally lossy blows it's socks off with 10x smaller size and half the decode time

Hrmm... That's not right :P

rip all the images on the page, transcode them to jxl and see how much space is saved [av1_dogelol](https://cdn.discordapp.com/emojis/867794291652558888.webp?size=48&quality=lossless&name=av1_dogelol)

`Original 136MB, ECT 95.4MB, JXL Lossless 71.7MB, JXL Lossy 15.7MB`

nice lol

Scratch that https://github.com/libjxl/libjxl/issues/3823

<@206628065147748352> in JXL oxide how do you store just a large array in the heap? Like how in C you can do my_struct* array = malloc(sizeof(my_struct)*10000)

I’m trying to learn rust in the scope of data compression

```rust let mut floats = vec![0.0f32; size]; ``` <https://doc.rust-lang.org/std/vec/struct.Vec.html>

is there a way to make a fixed-length Vec?

what do you mean by fixed length vec?

like how malloc() gives you a fixed length array and you get a thin pointer to it?

can you have something like a `Box<[f32; 100000]>` ?

yes

how do you init one?

the easiest way to make an array on the heap is `Box::new([1, 2, 3])`

just, making an array and putting it on the heap

If you want something more like malloc, where you provide a size and then get a pointer back, you want a boxed slice, and the easiest way to do that is `Vec::with_capacity(n).into_boxed_slice()` giving you a `Box<[T>`

will this be optimised by the compiler?

optmized how?

probably

to just a malloc()

Since the actual `Vec` struct only exists locally, easyish to optimize

rust seems complex

you could alloc directly, though that gives you a raw pointer so requires a bit of unsafe

can i just put my entire code in an unsafe?

or have i just re-invented C

I tried it on Godbolt

didn't compile to just a malloc

because there's some error handling nonsense as well 😄

compiler explorer?

yeah

wait, please ignore everything I just said 😂

I'm sleepy af, the above code does not work

your knowledge is a bit.. rusty?

no im just tired

and I accidentally read what I believe is an incorrect but upvoted answer on stackoverflow

<@386612331288723469> here's something that looks correct ```rust pub fn alloc_i32_slice(n: usize) -> Box<[i32]> { vec![0; n].into_boxed_slice() } ```

is there something like `.into_boxed_array()` ? whenever using the slice doesn't it do bounds checking?

array access also bounds checks

but you can just use `unsafe { x.get_unchecked(idx) }` to bypass bounds checking 😉

behold, my definitely correct and not bad Rust code ```rust use std::{ alloc::{self, handle_alloc_error, Layout}, mem::transmute, }; fn alloc_array<const N: usize>() -> Box<[i32; N]> { unsafe { let layout = Layout::new::<[i32; N]>(); let ptr = alloc::alloc_zeroed(layout); if ptr.is_null() { handle_alloc_error(layout) } let ptr = transmute::<*mut u8, *mut [i32; N]>(ptr); Box::from_raw(ptr) } } fn main() { let blah = alloc_array::<45>(); println!("{:?}", blah); } ```

https://www.hq.eso.org/public/images/eso1242a/ The full 3.9GB TIFF converted to JXL at e7 weighs in at 2.4GB and takes a whopping 27 seconds to load using nomacs, a Qt based image viewer. At e1, 2.6GB, it still takes around 5 seconds. vipsdisp took 35 seconds at e1 and had not finished loading even after 4 minutes at e7

does nomacs use chunked decoding?

No idea

But I was expecting vipsdisp to beat nomacs

because it often has in the past

Funny thing is my 200MP image takes around the same time to decode. I should really replace my CPU...

does anybody have that link of the manga page that has an entire manga in just pure images?

nvm found it

Can you link iT please?

yeah sure when my PC starts responding lol

https://m.grass.moe/chapter/lacking_54

I got jxl-oxide working with servo using the image integration and uh, it crashed my PC

[av1_kekw](https://cdn.discordapp.com/emojis/758892021191934033.webp?size=48&quality=lossless&name=av1_kekw)

Thanks

so yeah, using jxl with servo is... risky

it uses tons of system resources lol

using gb's corpus page hurts a little bit, but it's more or less fine, using that gross.moe page on the otherhand is bloody rough

well this isn't right

anyone know of any readers for android that can decode jxl in a .cbz container (or any other cb format)

tachiyomi can In a container, I can't remember which one tho

mihon*

I think cb was added

i tried in mihon (successor of tachiyomi) before and it worked. iinm the issue of reading from within cbz was since fixed

<@604964375924834314> could that ICC tool you made help compress dark images? Or would the fact the file is flagged as 24bit mean there will always be banding unless dithered

loading depends on Qt plugin (usually from kimageformats). I think the chunked decoding was not implemented there. Most probably RaveSteel is using patched sources to avoid limit preventing to load too big images. Opening speed could be even higher when future libjxl API allows requesting data in various pixel formats directly. Now it is possible to get RGB(A) and then the app converts pixels to BGRA.

Loading the DC until zooming, and then doing so with cropped decoding would also speedup

Maybe it should become a rule for images above a certain size to have some kind of dc frame

Either that or Scrooge

Squeeze*

Having the option of a non-upsampled DC would help lower memory usage a ton for previewing images too

Since all images have the 1:8 DC anyway

All VarDCT images*

I am not using any patched Software, but I am setting a Qt environment variable to allow loading images over 256MP

MAN, I converted all my gifs to jxl but forgot to some how lable them, I probably should have changed the file extension to .jxla or something, how do you go about "standardizing" more file extensions anyways?

wait these numbers are weird one sec

Earlier I named my jxls as jxl (native lossy jxl), jxll (native lossless jxl), jxlj (jpeg-in-jxl), but later I modified my de-jxl script to run jxlinfo first and get the idea what is inside jxl to determine correct output format

for me, I dont particularly care if something is lossless or not, but animated is fairly important to me

It was mostly done because there are too many stupid apps, that look only to extension, and it is impossible to tell them that these extensions are jxl too

I think for now ill just actually keep the gif since I don't have all that many of them anyways

It would be cool to have something other than gif but nobody wants to support it

the largest issue is sorting between "animated" and "not animated"

discord supports hdr for emotes and profile pictures now, just need to upload a hdr png e.g. generated by `djxl`

(presumably they just stopped stripping icc profiles from those, so the specific requirement for hdr would be that the icc profile has a `cicp` tag)

This cannot end well knowing the average Discord user

that sounds horrid

<@604964375924834314> giving you another poke since you just posted, hope you don't mind

ah, right, sorry

in its current state, I doubt that it could, but I have been made aware of a theoretical possibility which, however, wouldn’t work in Chrome

turns out that some ICC profiles have a white point with Y<1

Huh.... Interesting

but Chrome renders them differently from other software (I don’t remember for sure in which direction, I’ll try to remember to check tomorrow)

I did it this way "xxxx.gif.jxl" (animated jxl) "xxxx.jpg.jxl" (jpeg-in-jxl) "xxxx.png.jxl" (lossless jxl)

yeah that's what im doing for now

ah, Chrome/skcms renders them dark, lcms2 and Adobe don’t <@238552565619359744>

and the ICC spec mentions some cases where the luminance of the white point might be greater than 1, but it’s not clear to me whether it’s legal for it to be less

https://crbug.com/skia/40044611

2 years later the bug has a lead

I made a plot

Not sure if I should keep that bump for cameras, it's based on the 102-megapixel Fujifilm GFX100 which was released in 2019, while today in 2024 pretty much the highest resolution camera around is the Sony α7R V at 61 megapixels. My conclusion is they went ahead and broke that 100 megapixel barrier and then decided ~50 megapixels is actually enough.

the GFX 100 line is still going

https://www.dpreview.com/products/fujifilm/slrs/fujifilm_gfx100sii 2024

and it’s more affordable than it used to be

certainly more affordable than the Phase One IQ4 with its 150 megapixels

right but that one is also from 2019

point is: there doesn't seem to be any camera released after 2019 that has a higher resolution than those ones

my hypothesis is that for both phones and cameras, we've reached the point of "enough pixels", and any further improvement in fidelity is not going to come from higher resolution but from more color precision / dynamic range / etc

the other option is that these curves will keep going up like they have for 3 decades now

higher, no, but as high, yes (GFX 100S II), so should it really go back down?

Sony never went that high, Fujifilm continued to

the Fujifilm GFX 100 cameras have larger sensors, so it makes sense that they have more megapixels.

And fidelity isn't necessarily the only reason to increase sensor resolution; another is ability to crop.

I believe the GFX 100 cameras use Sony sensors, with the exact same pixel design as in the a7R V

Designing sensor pixels is complicated work, so ideally you want to design as few different pixel designs as possible with each new generation. Smaller sensors need small pixels to have a high enough resolution, so larger sensors get higher resolutions almost for free.

https://x.com/CDisillusion/status/1848581815331635322

Sony does not produce MF cameras, even the smallest MF, as Fuji does. However, Sony produces all these MF sensors now, for Fuji, P1, Hassel, whatever

If I am running the modular mode (say for a single channel image, such as grey-scale), in which stage does the block-size is determined, I mean the block sizes that the image is going to be divided? -- for JXL

Block size for modular?

Or is it just for the VarDCT , I mean lossy ?