|
DZgas Ж
|
2026-03-09 11:29:08
|
There's something interesting about blue noise being the perfect middle ground between a Bayer grid and white noise |
|
|
Dunda
|
2026-03-09 11:29:50
|
What does the frequency plot of bayer actually look like? |
|
|
|
DZgas Ж
|
2026-03-09 11:31:24
|
|
|
2026-03-09 11:34:01
|
Even the minimum attraction for the arrangement yields almost perfect noise, precisely because of the number of starting points. That is, this is the answer to the fact that if we only made the starting points, and all the other points were randomly selected. |
|
|
|
Dunda
|
2026-03-09 11:34:39
|
How strange that the ftm also resembles bayer |
|
2026-03-09 11:41:43
|
Sorry to have given you "advice" earlier, you clearly know very well what you are developing and I misinterpreted what you meant by "lines" as tiling borders |
|
|
|
DZgas Ж
|
2026-03-09 11:50:12
|
I don't really understand what I'm doing, since I'm doing it based on practice and non-theory. There are probably dozens of specific names and designations for all of this. I don't care about any of that; I create things based on testing, not by thinking about how they should actually be. I'm not a theoretician, and besides, I'm self-taught. BUT this field, the creation of blue noise <:galaxybrain:821831336372338729> , is so small that no one even knows what to tell me, and no one knows anything deeper than the basics themselves hehe |
|
|
|
Dunda
|
2026-03-09 11:51:38
|
I am strangely pulled in two directions, half the time I am reading theory, and the other half I am forgetting theory when I want to make a project |
|
|
|
DZgas Ж
|
2026-03-09 11:54:42
|
Well, I've been doing *vibecoding *for four years now. I have zero programming skills, but I've developed a strong ability to accurately describe the neural network task, so you could say I'm a professional pro(mpt)grammer. I've created all the programs I've created over the past four years exactly like this. All the programs for testing and generation, I often create everything from scratch in Python. No problem, it's fast and interesting. |
|
|
|
Dunda
|
2026-03-09 11:55:48
|
Ah, I'm not a fan of vibe coding, but at least you are going in an interesting direction |
|
|
|
DZgas Ж
|
2026-03-09 11:57:35
|
I can understand. Ultimately, it's associated with bad code, but it's absolutely true. Vibecoding is as difficult as programming, but programming specifically doesn't come to me. I look at code for 10 years now and I see nothing, just look and my head is empty. Unfortunately, it's something personal. |
|
|
Elias
|
2026-03-09 01:43:17
|
I know the correct answer to the question is probably "it depends on the image", but roughly speaking, what technique improves compression the most compared to JPEG? the variable DCT, predictions or ANS? Or something else entirely? Or is there an easy way to test it for a dataset? |
|
|
|
DZgas Ж
|
2026-03-09 01:44:01
|
Are you asking about jpeg xl? |
|
|
|
Elias
|
|
|
DZgas Ж
|
2026-03-09 01:46:18
|
Well, the basis is the smart definition of importance and division into varDCT blocks, better compression of the data itself within the DCT blocks. The second is the XYB space instead of YUV, but that's very brief, talking about the differences in JPEG. But I think it's better to read this: https://docs.google.com/presentation/d/1LlmUR0Uoh4dgT3DjanLjhlXrk_5W2nJBDqDAMbhe8v8/edit#slide=id.gae1d3c10a0_0_17 |
|
2026-03-09 01:52:00
|
|
|
|
Exorcist
|
2026-03-09 01:57:33
|
> improves compression the most
intra prediction |
|
|
_wb_
|
2026-03-09 02:56:27
|
Assuming you're talking about lossy JPEG and lossy JXL: in ballpark numbers, I'd say better entropy coding gives ~20% improvement (not just ANS but also context modeling and the numzero approach for encoding HF and the modular LF encoding), varblocks with adaptive quantization gives another 15% maybe, XYB and chroma from luma give a few percent each, say together 5-10%, and then gaborish/EPF also help a bit. |
|
|
|
Elias
|
2026-03-09 03:05:35
|
yes exactly, thank you that helps a lot! |
|
|
ziemek.z
|
2026-03-09 09:31:13
|
Thanks <@532010383041363969> for merging my (and couple of others') PRs so Knusperli can be built again with modern Bazel 🙂 |
|
2026-03-09 09:32:33
|
> MP3 Pro Max
Will somebody tell him that mp3PRO already exists? 😉
https://www.rarewares.org/rrw/ctmp3pro.php
https://web.archive.org/web/20080713034003/http://www.mp3prozone.com/basics.htm |
|
2026-03-10 12:47:22
|
https://github.com/libjxl/libjxl/issues/477 |
|
|
Jyrki Alakuijala
|
2026-03-10 01:00:05
|
((I wrote a relatively nice two pass dithering method around 1997 but didn't opensource, I proposed it to HP for their printers, but they didn't want it)) |
|
2026-03-10 01:01:43
|
makes me wonder what the JPEG XL noise generation (4-connected Laplacians) algorithm looks in FFT space |
|
|
|
DZgas Ж
|
2026-03-10 01:01:56
|
not related to blue noise ? |
|
2026-03-10 01:03:43
|
Well, it's most likely random. It's just that blue noise is the most non-random noise; it can't be generated from a single pixel; the algorithm is extremely complex. Therefore, it's used only as a ready-made set tile texture |
|
2026-03-10 01:07:00
|
Using my blue noise texture as a random generator, I created this complex algorithm that projects each image pixel onto a three-dimensional-rhombus inside an RGB cube and selects the closest color based on the blue noise number. It's an extremely complex algorithm, and I'm starting to doubt anyone has ever done anything like this. For some reason, I can't see it... have I really invented something else?https://discord.com/channels/794206087879852103/794206087879852106/1479871428471029961
Can you tell me if this already existed and where? |
|
2026-03-10 01:08:15
|
I found the simplest three-color algorithm, R G B 3-color dithering similar, but I didn't find anything like that for a palette, for a completely own custom color palette |
|
|
|
Dunda
|
2026-03-10 01:24:07
|
It's hard to tell what exactly you mean |
|
|
|
DZgas Ж
|
2026-03-10 01:28:09
|
I mean dithering algorithms that can do this quality https://discord.com/channels/794206087879852103/794206087879852106/1479871428471029961 using 16 colors |
|
|
|
Exorcist
|
2026-03-10 01:28:37
|
Where is the source image? |
|
|
|
DZgas Ж
|
2026-03-10 01:29:21
|
https://discord.com/channels/794206087879852103/794206087879852106/1479457050835685397 part at the bottom of the image |
|
2026-03-10 01:33:57
|
16 colors [0, 0, 0], [255, 0, 0], [0, 255, 0], [0, 0, 255], [255, 255, 0], [255, 0, 255], [0, 255, 255], [192, 192, 192], [128, 128, 128], [128, 0, 0], [0, 128, 0], [0, 0, 128], [128, 128, 0], [128, 0, 128], [0, 128, 128], [255, 255, 255] |
|
2026-03-10 01:34:45
|
full image dither |
|
2026-03-10 01:36:31
|
PBBN_256.png which I used for dithering here https://discord.com/channels/794206087879852103/794206087879852106/1479454816991182930 |
|
|
A homosapien
|
2026-03-10 10:15:29
|
https://github.com/libjxl/libjxl/pull/4669 |
|
|
monad
|
2026-03-11 03:22:12
|
> clamped
Is it actually clamped, or is this misstated? |
|
|
|
A homosapien
|
2026-03-11 03:45:54
|
Would "scaled" be a more appropriate term?
Only ±0.5 got changed to ±0.49219 |
|
|
|
monad
|
2026-03-11 03:58:39
|
Sounds like it is not misstated then, but clamping seems a blunt approach rather than scaling all values to accommodate the range limit. |
|
|
|
A homosapien
|
2026-03-11 04:24:23
|
I'll scale the values when I'm not afk |
|
|
spider-mario
|
2026-03-11 09:27:06
|
https://www.keithcirkel.co.uk/whats-my-jnd/ |
|
|
Tirr
|
2026-03-11 09:44:29
|
I got 0.0054 |
|
2026-03-11 09:45:47
|
the whole experience is something like "I can clearly see it" and "looked at it very long time and got it wrong" |
|
|
|
_wb_
|
2026-03-11 09:46:01
|
nice, I only got here |
|
2026-03-11 09:46:18
|
I mostly noticed my screen is dirty, I should clean it and retry 🙂 |
|
|
|
Tirr
|
2026-03-11 09:47:54
|
I'm on somewhat clean iPhone 17 |
|
|
|
veluca
|
2026-03-11 10:14:54
|
|
|
2026-03-11 10:15:17
|
(not sure if my display is properly calibrated though, using a AW3423DW) |
|
|
lonjil
|
2026-03-11 10:15:18
|
I wonder how different my results might be between my budget screen from 2010 vs my midrange screen from 2020. |
|
|
|
veluca
|
2026-03-11 10:16:48
|
wonder if this is what spending years staring at compressed images does to you 😄 |
|
|
|
Lucas Chollet
|
2026-03-11 10:18:16
|
This has to be dependent on your screen and ambient luminosity too |
|
|
|
lonjil
|
2026-03-11 10:18:26
|
it says that the colors will get closer and closer but that's clearly not true in a perceptual sense |
|
|
|
veluca
|
2026-03-11 10:18:58
|
yeah not all the color pairs with the same deltaE are visually equally apart |
|
2026-03-11 10:22:10
|
yeah got 0.0047 on my phone, I suspect the desktop screen oversaturates a bit (or possibly it's just randomness) |
|
|
|
lonjil
|
|
Fahim
|
2026-03-11 10:35:47
|
I struggled a lot with the magentas and bright reds (Gigabyte M32U, should've probably switched to the sRGB mode lol) |
|
2026-03-11 10:42:13
|
0.0042 on my phone, similar struggles but with touches registering quite far off alongside |
|
|
|
spider-mario
|
2026-03-11 10:44:51
|
managed to optimise the viewing conditions to |
|
2026-03-11 10:46:59
|
(link: https://www.keithcirkel.co.uk/whats-my-jnd/?r=AHokKP___O_3)
(dark room, maximum screen brightness, no calibration, assigned a P3 profile to a roughly P3 screen, occasionally wiggled the window to make sure I wasn’t mistaking a screen non-uniformity for the actual transition) |
|
|
|
DZgas Ж
|
2026-03-11 11:48:24
|
A strange but funny test, my 20-year-old ips monitor be well
Frosted glass, sun shining on the wall, low contrast 0.0115 |
|
2026-03-11 11:49:53
|
Oh yea |
|
2026-03-11 11:50:52
|
Well, either I just have good eyesight, I still don't understand what this test measures, my eye or my monitor... |
|
2026-03-11 11:58:31
|
I think the old description was more intuitive because the phrase "to prevent rounding" begs the question of why the values aren't equal to 0.49999 for this |
|
|
|
veluca
|
2026-03-11 12:22:08
|
What's My JND? 0.0021
Can you beat it? https://www.keithcirkel.co.uk/whats-my-jnd/
(This on the phone) |
|
|
|
_wb_
|
2026-03-11 01:05:11
|
What's My JND? 0.0033
Can you beat it? https://www.keithcirkel.co.uk/whats-my-jnd/?r=AUkgKP__8v4u
On my phone |
|
|
Kleis Auke
|
2026-03-11 01:58:35
|
What's My JND? 0.0037
Can you beat it? https://www.keithcirkel.co.uk/whats-my-jnd/?r=AXMgKP__8fTP
(with my old SDR (Dell UltraSharp U2414H) monitor, I should probably test this on my phone as well) |
|
|
jonnyawsom3
|
2026-03-11 02:17:53
|
I discovered 2 things. The first is that my desaturation testing basically trained me for this. The other is that my decade old phone just doesn't have a display capable of showing all the colors. A few times it had 'noise' because the range wasn't high enough
What's My JND? 0.0115
Can you beat it? <https://www.keithcirkel.co.uk/whats-my-jnd/?r=BHofKP_7_Dh_> |
|
|
Meow
|
2026-03-11 02:30:38
|
What's My JND? 0.0032
Can you beat it? https://www.keithcirkel.co.uk/whats-my-jnd/?r=AT0iKP__v8-P |
|
2026-03-11 02:30:52
|
At the exhausted night |
|
|
awxkee
|
2026-03-11 03:27:12
|
What's My JND? 0.0029
Can you beat it? https://www.keithcirkel.co.uk/whats-my-jnd/?r=ASEjKP__33-n |
|
2026-03-11 03:28:00
|
It's fairly complicated on my MacBook |
|
|
AccessViolation_
|
2026-03-11 04:22:20
|
([Q1572121](<https://www.wikidata.org/entity/Q1572121>)) image file format
([Q138645418](<https://www.wikidata.org/entity/Q138645418>)) general-purpose image coding system
`general-purpose image coding system` **is subclass of** `image file format` |
|
2026-03-11 04:22:58
|
It's unfortunate there's no distinction between *file* formats and formats in general |
|
2026-03-11 04:23:56
|
e.g. JPEG XL defines the file format and the bitstream, but they're not one and the same |
|
2026-03-11 04:36:48
|
it seems like generalizing data formats or coding systems to all be file formats means things now can't be classified properly. I'm not about to change up a bunch of entities but it irks me a little |
|
|
Kampidh
|
2026-03-11 07:55:49
|
https://www.keithcirkel.co.uk/whats-my-jnd/?r=AMUiKP___3Gf
LG 27GL850, clamped to sRGB @ 130 cd/m2 |
|
|
|
DZgas Ж
|
2026-03-11 08:13:15
|
0.0070 at night |
|
|
|
monad
|
2026-03-11 08:22:24
|
What's My JND? 0.0027
Can you beat it? <https://www.keithcirkel.co.uk/whats-my-jnd/?r=AQ0jKP__33f5>
XB273K GP in bright room |
|
2026-03-11 09:23:42
|
What's My JND? 0.00032
Can you beat it? <https://www.keithcirkel.co.uk/whats-my-jnd/?r=ACAkKP____o_>
Second attempt, tryharding in a dark room |
|
|
Adrian The Frog
|
2026-03-11 10:27:18
|
.0034 on my Oneplus 12, .0039 on cheap gaming IPS monitor |
|
|
|
A homosapien
|
2026-03-11 10:32:17
|
Done |
|
2026-03-11 10:32:44
|
I just tried it and a bunch of tests fail, 0.49999 still rounds to 0.5 which causes issues. I chose 0.49219 since it was also the limit of the older dithering implementation. It's passes all tests, which is good enough for me. |
|
|
|
lonjil
|
2026-03-11 11:00:17
|
I did my previous one when the sun was blasting this room. Tried again now with just regular indoor lighting and got 0.0021.
Both times on my cheap Lenovo monitor. |
|
|
|
A homosapien
|
2026-03-11 11:01:57
|
I initially got 0.0018. The more I retake the test, the worse my scores become, I think my cones are getting fatigued. |
|
|
|
jonnyawsom3
|
2026-03-11 11:06:05
|
Did better on my desktop |
|
|
|
AccessViolation_
|
2026-03-11 11:47:38
|
What's My JND? 0.0029
Can you beat it? https://www.keithcirkel.co.uk/whats-my-jnd/?r=ASEhKP__-181 |
|
2026-03-11 11:56:08
|
second time, I assume higher is better because I took more time to allow my eyes to adjust and look around carefully before assuming I couldn't see the difference |
|
2026-03-11 11:56:44
|
if lower is better then I don't have an explanation |
|
|
|
ignaloidas
|
2026-03-12 12:03:33
|
Mostly struggled with bright colors, the darker ones were all basically free (in a dark room, mid-tier LG monitor) |
|
|
|
Adrian The Frog
|
2026-03-12 02:54:31
|
lower is better. I had a decent amount of variance between attempts though |
|
2026-03-12 03:11:40
|
got a new best, .0025 on my ips monitor
after cleaning it |
|
|
|
AccessViolation_
|
2026-03-12 09:05:12
|
there's no way any of these results here are comparable btw xD |
|
2026-03-12 09:05:33
|
unless people here have the same display or viewing conditions |
|
|
|
DZgas Ж
|
2026-03-12 09:22:16
|
Literally everyone has normal monitors and normal vision, wow, that's a test <:PepeOK:805388754545934396> |
|
|
|
monad
|
2026-03-12 09:24:10
|
what if we're all the same person |
|
|
|
DZgas Ж
|
2026-03-12 09:24:55
|
<:monkaMega:809252622900789269> I get it, we're all just have pixels on we're monitor, so the tests are similar |
|
2026-03-12 09:27:37
|
I have the worst results, but they're so microscopic that it's not even worth it. My monitor is from 2009, it's 17 years old, IPS VGA, I bought it for literally $10 |
|
2026-03-12 09:28:07
|
Probably, to do poorly on the test you need to have a CRT monitor or just a terribly calibrated monitor |
|
|
|
AccessViolation_
|
2026-03-12 10:00:24
|
careful, you're going to be invited to a philosophy podcast if you keep this up <:KekDog:805390049033191445> |
|
|
|
Jyrki Alakuijala
|
2026-03-12 02:01:00
|
oklab color model doesn't have a gamma compression bias, i.e., no linear range for small values, it will be very confused there |
|
|
|
ignaloidas
|
2026-03-12 02:09:53
|
tbh it wasn't even that dark for the easy ones, I'd say anything around 60% or below (okalab) lightness was easy |
|
|
|
Jyrki Alakuijala
|
2026-03-12 03:07:02
|
in dark values oklab is totally off |
|
2026-03-12 03:07:28
|
in bright values it is a log (human eye) vs. cubic root (oklab) |
|
2026-03-12 03:07:47
|
in JPEG XL we also approximate the log as a cubic root, but at least it is biased for the low values |
|
2026-03-12 03:08:44
|
Butteraugli XYB color compression is log, JPEG XL XYB cbrt, SSIMULACRA 2 cbrt |
|
2026-03-12 03:11:04
|
I think SSIMULACRA cbrt is also biased, but never looked if it is really the case |
|
|
|
Adrian The Frog
|
2026-03-12 03:18:56
|
On the 3 very different screens I've tried in very different viewing conditions, I'm relatively consistently getting .0045-.0025 |
|
|
|
AccessViolation_
|
2026-03-12 03:19:03
|
this incidentally could also be a pretty good test of how well image encoders are able to represent different but similar colors at different quality settings |
|
|
|
Jyrki Alakuijala
|
2026-03-13 09:49:51
|
I got 0.0021 JND (usual office viewing environment) |
|
|
Froozi
|
2026-03-13 09:51:43
|
I've managed 0.0057 on an incredibly old monitor that has few dead pixels and two dead LEDs on the back panel; But I doubt I'd manage a better score on a proper modern display. |
|
|
|
Jyrki Alakuijala
|
2026-03-13 09:52:40
|
if you move around the window on the screen, you can escape some unevenness in the monitor that you cannot get with by just moving the eyes |
|
|
|
Dunda
|
2026-03-13 11:07:21
|
I got about 0.0038 which is a little disappointing |
|
|
|
Meow
|
2026-03-14 12:18:29
|
Maybe it meant that you should buy a better display |
|
|
|
DZgas Ж
|
|
HCrikki
|
2026-03-15 04:30:44
|
dng sdk received an update for security/other fixes and jxl library update (1.7.1 Build 2502, March 10, 2026) |
|
2026-03-15 04:31:24
|
anyone checked how meaningful change is? |
|
|
RaveSteel
|
2026-03-15 07:19:12
|
No, but apparently there is now a license file included. Seems the immich devs have contacted a lawyer to check for compatibilty with FOSS licenses |
|
|
|
Kleis Auke
|
2026-03-15 11:17:46
|
It looks like the "Adobe DNG SDK License Agreement" is not [OSI-approved](https://opensource.org/licenses), so it is unlikely to be included in Immich (though IANAL). |
|
|
|
spider-mario
|
2026-03-15 01:53:42
|
```
Example used in this integration:
- File: `camera_raw/libjxl/client_projects/mac/Configs/jxl.xcconfig`
- Added: `HWY_DISABLED_TARGETS=25952256`
- Purpose: disable problematic SVE targets in this environment.
```
curious way to write `(HWY_SVE|HWY_SVE2|HWY_SVE_256|HWY_SVE2_128)` |
|
2026-03-15 01:54:06
|
or even 0x18C0000 |
|
|
|
jonnyawsom3
|
2026-03-15 08:21:38
|
Library update will just be 0.11.2, so nothing new |
|
|
|
HCrikki
|
2026-03-15 08:59:25
|
really? i thought it shipped some old slow libjxl |
|
|
|
Froozi
|
2026-03-15 09:20:47
|
Noo, me want 0.12… <:FeelsSadMan:808221433243107338> |
|
|
Ryoumiya
|
2026-03-16 08:48:03
|
Hello everone, I was recommneded to share this here... as part of SIGBOVIK me and a few friend decided to yeet a bunch of model weights into image compression (lossy and lossless) algorithms, and found that JXL performed pretty good for small models (0.15 MiB).
Please focus on the top left part around 0.15 MiB that is lossless, with JXL, achiving a small 0.02MiB drop from the baseline, |
|
|
|
spider-mario
|
2026-03-17 01:09:10
|
there is now apparently a “hard mode” https://www.keithcirkel.co.uk/whats-my-jnd-hard/ |
|
|
|
lonjil
|
2026-03-17 02:27:22
|
0.0058
Lack of gradient made it harder, but having differences in two axes made it easier to not get fooled by monitor weirdness. |
|
|
Weaselslider
|
2026-03-17 06:33:08
|
a fun side effect of this is finding that 100% quality mode is consistently worse than lossless mode lol |
|
|
|
HCrikki
|
2026-03-17 06:39:04
|
jpg is vague. could be handy comparing mozjpeg to jpegli |
|
|
|
Weaselslider
|
2026-03-17 06:41:15
|
its whatever PIL has packaged in it |
|
|
|
Fahim
|
2026-03-17 06:41:31
|
Should be plain libjpeg-turbo at best |
|
|
|
jonnyawsom3
|
2026-03-17 07:21:11
|
One day we'll get into Blender... |
|
|
|
RaveSteel
|
2026-03-17 08:03:36
|
JXL payload in EXR when? /s |
|
|
|
_wb_
|
2026-03-17 08:13:09
|
TBH I would prefer it if JXL wouldn't be used as a payload codec inside application-specific formats, but specific applications would just use JXL files with their application-specific stuff in extra boxes. So instead of JXL in DICOM, or JXL in DNG, or JXL in GeoTIFF, etc, you would just have valid JXL files (even though you could give them a different filename extension if you like) with some extra box(es) in them with all the stuff needed for DICOM/DNG/GeoTIFF/etc.
That way any viewer including browsers could just render the image natively with zero effort. |
|
|
|
Exorcist
|
2026-03-17 08:14:23
|
> with their application-specific stuff in extra boxes
Same wrong as APNG🤮 |
|
2026-03-17 08:16:20
|
In today, many viewers still show only first frame (and this is conformant) |
|
|
|
spider-mario
|
2026-03-17 10:55:29
|
if the alternative is no support at all (like MNG), though? |
|
|
Demiurge
|
2026-03-18 12:06:17
|
People will do what they want. Not everyone is smart enough to ask for consultation... |
|
2026-03-18 12:17:43
|
Sometimes you have to reach out to someone and send an email to the right person and they're usually appreciate of the advice. |
|
|
Mine18
|
2026-03-20 04:34:44
|
0.0189 |
|
|
|
AccessViolation_
|
2026-03-20 07:48:25
|
|
|
2026-03-20 07:53:19
|
I bet I'll do much better when I do it on my phone which can be set to stretch the sRGB range into amoled wide gamut |
|
|
|
lonjil
|
2026-03-20 07:59:18
|
isn't that cheating? |
|
|
|
AccessViolation_
|
2026-03-20 08:19:25
|
certainly, I wouldn't submit it |
|
2026-03-20 08:28:35
|
|
|
|
|
DZgas Ж
|
|
|
A homosapien
|
|
|
DZgas Ж
|
2026-03-21 09:16:15
|
I just think it's a funny long delay, given that not a single line of executable code that could be checked has been changed and importance |
|
|
KonPet
|
2026-03-26 02:24:26
|
https://booth.pm/ja/items/5245660
lmao |
|
2026-03-26 02:24:55
|
it's almost like the jpegxl equivalent of llvm.moe |
|
|
hana
|
|
|
Meow
|
2026-03-26 10:12:12
|
We need JXLchan |
|
|
|
jonnyawsom3
|
2026-03-30 03:01:38
|
<@1156997134445461574> How's the mobile photo upload quality been going by the way? Last time you said you started measuring quality with PSNR and SSIM and saw room for improvement. Currently (I think) Discord lowers quality based on resolution, but I'd use a fixed quality and lower the resolution instead. Then modern 50MP phones aren't uploading worse photos than a 1MP desktop resize
Also if you hadn't seen, JXL support is now behind a flag in Chrome, with progressive loading implemented in Canary
At some point you might even be able to replace the blurhashes and WebP previews with progressive JXL thumbnails instead
(Quick comparison against an existing WebP preview from Discord's CDN. 31 KB of the WebP vs 4 KB, 11 KB and 31 KB of the same JXL) |
|