JPEG XL

Info

rules 58
github 38694
reddit 688

JPEG XL

tools 4270
website 1813
adoption 22069
nomacs 7
UI 11
chrome 34
image-compression-forum 0
glitch-art 1071

General chat

welcome 3957
introduce-yourself 294
color 1687
photography 3532
other-codecs 25116
on-topic 26652
off-topic 23987

Voice Channels

General 2578

Archived

bot-spam 4577

jxl

Anything JPEG XL related

Jyrki Alakuijala
2021-04-21 07:20:59 if you do a basic economic optimization you need to assign a value for a human life
fab
2021-04-21 07:21:05 aren't you at meeting
Jyrki Alakuijala
2021-04-21 07:21:31 no, Luca and Jon will deal with it very well
2021-04-21 07:21:48 if they need me, they will ask me 😄
fab
2021-04-21 07:22:00 can jpeg xl model verification
2021-04-21 07:22:00 do
2021-04-21 07:22:01 for %i in (C:\Users\User\Documents\jpg.png) do cjxl "%i" "%i.jxl" -I 3.9 -X 8 -P 8 -Q 86.6 -m -s 8 --num_threads=2 for %i in (C:\Users\User\Documents\jxl1.png) do cjxl "%i" "%i.jxl" -q 60.4 -s 4 --faster_decoding=3 --num_threads=2
2021-04-21 07:22:06 same quality as this
Jyrki Alakuijala
2021-04-21 07:22:13 it is not like this ISO meeting is going to make earth shattering new design decisions (or any other ISO meeting)
fab
2021-04-21 07:22:15 or better quality
2021-04-21 07:22:25 i did this with 0.3.5
monad
2021-04-21 07:23:07 Why do you still do I > 1?
fab
2021-04-21 07:23:31 i thought it was I until 5
veluca
2021-04-21 07:24:09 nah, above 1 is exactly 1
fab
2021-04-21 07:25:04 so the quality has improved or it needs tweaking
2021-04-21 07:25:23 or is only efficient for graphical
2021-04-21 07:29:30 in comparison to q 83.2 new heuristics is not bad
2021-04-21 07:29:41 d 1 and d2 are good
2021-04-21 07:30:03 but i want to know lower quality in the q 60.4 ranged
2021-04-21 07:30:20 i think with new builds it will appear more loss
Jyrki Alakuijala
2021-04-21 07:30:35 quality has improved a lot from 0.3.2 and earlier and even more so against 0.2.x and earlier
2021-04-21 07:31:12 I don't remember exactly when the last patch of quality improvements surfaced, possibly in 0.3.4 or 0.3.3
fab
2021-04-21 07:31:16 from 0.3.7
Jyrki Alakuijala
2021-04-21 07:31:29 oh, that shouldn't be the case
fab
2021-04-21 07:31:31 like it can do better quality than the command i use
2021-04-21 07:31:41 at same distance
Jyrki Alakuijala
2021-04-21 07:31:46 (I didn't even know we have a 0.3.7 already :-D)
2021-04-21 07:32:14 if you compare 0.3. against 0.2. -- yes, a huge improvement
2021-04-21 07:32:22 but 0.2. was not bad
2021-04-21 07:32:42 I'm now back to coding and hope to make further quality improvements
fab
2021-04-21 07:32:46 or more compression less quality
Jyrki Alakuijala
2021-04-21 07:32:57 my next patch of improvements should improve on higher bpp
fab
2021-04-21 07:33:07 i hope it works for all files
2021-04-21 07:33:09 or most
Jyrki Alakuijala
2021-04-21 07:33:15 some improvement on low bpp, too
fab
2021-04-21 07:33:19 even if you sacrifice quality for compression
2021-04-21 07:33:33 at lower bpp
Jyrki Alakuijala
2021-04-21 07:33:43 but make the higher bpp compression not only more effective but 'safer', less surprise anywhere in the image
2021-04-21 07:33:59 yes, it should become a lot safer
2021-04-21 07:34:19 I don't know yet how slow it will be, possibly in -s 8
fab
2021-04-21 07:34:26 i want x3 compression
2021-04-21 07:34:34 from 171 kb to 60 kb
Jyrki Alakuijala
2021-04-21 07:34:47 x3 from jpeg size?
fab
2021-04-21 07:35:02 i think with png is good already
2021-04-21 07:35:08 for screenshots
2021-04-21 07:35:25 as long itworks
2021-04-21 07:35:37 current i have like 87 kb from 171 kb
2021-04-21 07:35:53 and 0.3.5
2021-04-21 07:37:17 can you do better with that image
2021-04-21 07:37:44 that is a 7z link
2021-04-21 07:37:58 without using same command i used
2021-04-21 07:38:05 so without 2 generation loss
2021-04-21 07:39:41 this don't satisfy me in image quality
2021-04-21 07:39:55 particularly colors
2021-04-21 07:40:04 they look undersaturated
2021-04-21 07:40:37 nhw codec hasn't the same problem more
2021-04-21 07:41:30 but i don't like nhw codec it feels worse than modular to me
2021-04-21 07:41:38 i like block based codecs
Jyrki Alakuijala
2021-04-21 07:50:00 This image we will be able to compress in far future using a new encoder
2021-04-21 07:50:33 the context modeling that Jon has been abusing for procedural art can be tamed and it will be able to do massive compression on this kind of material
2021-04-21 07:50:50 I anticipate that we will be able to invest in it in about a year
2021-04-21 07:51:04 it is a very complex problem on how to do it and no one has done it before
fab
2021-04-21 07:51:12 but with lossy
2021-04-21 07:51:22 like why nhw wavelet codec
Jyrki Alakuijala
2021-04-21 07:51:26 yes, we can create loss
fab
2021-04-21 07:51:29 can even with -l7
2021-04-21 07:51:33 have same saturation as original
2021-04-21 07:51:36 and jpeg xl
2021-04-21 07:51:38 not
Jyrki Alakuijala
2021-04-21 07:51:46 without block based, but with similar density
fab
2021-04-21 07:51:56 because jpeg xl is block based
2021-04-21 07:52:02 what is the problem
2021-04-21 07:52:10 or the encoder that isn't optimized
2021-04-21 07:52:14 and undersaturate
2021-04-21 07:52:18 to prevent loss
Jyrki Alakuijala
2021-04-21 07:52:25 do you see undersaturation?
fab
2021-04-21 07:52:43 honestly this image even at lowest quality should look good
Jyrki Alakuijala
2021-04-21 07:52:46 I'm just increasing butteraugli's red-green saturation concern for large differences
fab
2021-04-21 07:53:03 i see undersaturation for that image at lossy varDCT
Jyrki Alakuijala
2021-04-21 07:53:03 (but loosening on very small saturation differences)
fab
2021-04-21 07:53:30 from lower quality to normal d1
2021-04-21 07:54:18 it is very perceived to normal eye
2021-04-21 07:54:53 captured from firefox png
Crixis
2021-04-21 07:57:46 can be a problem of incorrect visualisation of the source png?
fab
2021-04-21 07:59:12 there are many colours
2021-04-21 07:59:26 is not simple as you think
2021-04-21 07:59:47 the font is carrier.
2021-04-21 08:01:47
2021-04-21 08:02:08 is not good as hakuna sans from dafont
2021-04-21 08:02:40 from ttf vk
Crixis
2021-04-21 08:05:24 I don't see any desaturation
fab
2021-04-21 08:06:46 which settings are
Crixis
2021-04-21 08:07:16 standard -d 1
fab
2021-04-21 08:07:21 and speed 9
Jyrki Alakuijala
2021-04-21 08:07:29 ❤️
fab
2021-04-21 08:07:34 do you have 50° commit?
2021-04-21 08:07:44 crixi what you did
2021-04-21 08:07:54 i'm trying to download all thumbnails
2021-04-21 08:07:59 for me
Jyrki Alakuijala
2021-04-21 08:08:06 looks amazingly good at d1
Crixis
2021-04-21 08:08:22 and is only 257 kb
fab
2021-04-21 08:08:34 crixis answer
2021-04-21 08:08:43 i didnt' produced image as good as you
Jyrki Alakuijala
2021-04-21 08:08:49 (on the lowest pink icon I see some artefacts, going to work on problems like that)
veluca
2021-04-21 08:08:51 friendly meeting reminder 😛
Crixis
2021-04-21 08:08:58
fab
2021-04-21 08:09:02 s7 or s9
2021-04-21 08:09:04 what commit
2021-04-21 08:09:07 answer crixis
Crixis
2021-04-21 08:09:18 0.3.7 no flags
fab
2021-04-21 08:09:22 what commits
2021-04-21 08:09:28 what speed
Crixis
2021-04-21 08:09:37 no flags
fab
2021-04-21 08:09:54 but what commit
2021-04-21 08:09:58 do you have a1212
2021-04-21 08:10:04 the new one 50° commit
2021-04-21 08:10:41 <@!424295816929345538>
Crixis
2021-04-21 08:10:48
fab
2021-04-21 08:11:00 https://gitlab.com/wg1/jpeg-xl/-/commit/a124844519310785445d0b6efcd536c5398e6a20
2021-04-21 08:11:01 this
2021-04-21 08:11:29 do you have the one with all the two group fix
2021-04-21 08:11:33 or the older one
2021-04-21 08:11:35 which commit
2021-04-21 08:13:40 the output you did
2021-04-21 08:13:44 but better than what i did
Crixis
2021-04-21 08:14:06 where?
fab
2021-04-21 08:14:13 all image
2021-04-21 08:14:16 pink becomes grey
2021-04-21 08:14:21 as jyri said
Crixis
2021-04-21 08:19:42 color piker give me same color, but i can see some red approximation
2021-04-21 08:24:56 at d 0.4
2021-04-21 08:28:01 but lossless do better in this image
2021-04-21 08:28:11
2021-04-21 08:28:26 same size as d 1
fab
2021-04-21 09:01:58 yes but all qualities matter
2021-04-21 09:02:35 ok i will upload the thumbnail i downloaded in full resolution
2021-04-21 09:02:49 they are all 1280x720 jpg
2021-04-21 09:03:02 webp i didn't found and also there is overlay background
2021-04-21 09:03:29 i will use mega.nz al
2021-04-21 09:03:49 or d1.0
2021-04-21 09:08:52 100 thumbnails of youtube music
2021-04-21 09:11:34 is that windows 10 photo viewer?
2021-04-21 09:11:41 it's a program i don't know
2021-04-21 09:14:36 BUT the great is the nhw codec until -l 7 does even better job at lossy with wavelet algorithms
2021-04-21 09:17:30 and don't lose saturation
2021-04-21 09:22:25 also i have another test problem
Crixis
2021-04-21 09:22:47 is QuickLook
2021-04-21 09:25:44 from 9 MB to 7 MB
2021-04-21 09:32:38 without loss
Jyrki Alakuijala
2021-04-21 10:53:05 better is possible
fab
2021-04-21 12:52:44 i don't know if adaptively saturating some image colors at more than -d 2 -d 4 help with the perception
2021-04-21 12:52:48 that's not my job
2021-04-21 12:54:23 maybe i would use lossless
2021-04-21 12:54:40 i did with latest version on default and i had 336 kb
2021-04-21 12:54:55 crixis had 250 kb so it used something different
2021-04-21 12:55:02 or maybe i used random settings
2021-04-21 02:05:27 ok i get 331 kb with default q 100
2021-04-21 02:05:43 so maybe crixis did s 9 E 3
2021-04-21 02:05:53 with ORI.png
diskorduser
2021-04-21 04:11:24 Saturated colors? I don't see any color changes bw source and jxl even on a calibrated monitor.
fab
2021-04-21 04:11:46 undersaturated
diskorduser
2021-04-21 04:11:57 I think you have issues with color profile
fab
2021-04-21 04:11:58 i see way less colours than original
2021-04-21 04:12:03 not this
2021-04-21 04:12:07 but less saturation
2021-04-21 04:12:20 i didn't counted colours with imagemagick or xnview
2021-04-21 04:12:31 i just downloaded the file crixis created
2021-04-21 04:12:38 i don't have my encoding
diskorduser
2021-04-21 04:12:55 Could you give me the source image? I will test it at my end.
fab
2021-04-21 04:13:37 yes but anyway it can be improved
2021-04-21 04:13:50 the saturation
2021-04-21 04:14:15 even if cjxl derives from pik and pik mission also fuif is to not oversaturate lossy output
2021-04-21 04:15:06 source image:
2021-04-21 04:15:07 https://discord.com/channels/794206087879852103/794206170445119489/834332420513464391
2021-04-21 04:17:54
2021-04-21 04:18:26 the colours are pretty different
2021-04-21 04:19:54
2021-04-21 04:20:27 how you can not see the difference
2021-04-21 04:20:49 this is speed 7
2021-04-21 04:21:08 maybe he hasn't the encoder without the group fix
2021-04-21 04:22:15 trying with other viewers
2021-04-21 04:35:51 here's there is und
2021-04-21 04:37:20 this is xnview decoded png with capture tool of windows 7
2021-04-21 04:37:43 but the loss is only few pixels is good png
2021-04-21 04:42:33 nhw codec don't do that
2021-04-21 04:44:28 i didn't see less saturation, but less appealing colours.
2021-04-21 04:46:00 i will try with 2 generation loss
2021-04-21 04:47:18 this is maybe similar to the issue Lee is having
2021-04-21 04:47:23 or correlated
2021-04-21 05:10:56 first re encoding
2021-04-21 05:11:09 with
2021-04-21 05:11:29 no i have another command with 0.37 that i'm trying now
Crixis
2021-04-21 05:19:53 oh god what is this mess?
2021-04-21 05:20:20 this is so blocky, how?
2021-04-21 05:21:09 why use Modular lossy?
_wb_
2021-04-21 05:21:42 It is nearly always a bad idea to use Modular lossy
fab
2021-04-21 05:26:34
2021-04-21 05:27:10 this is smaller as i can get with 0.3.7
2021-04-21 05:27:33 but ariana grande face was smoothed
2021-04-21 05:28:32 this inflates file size
2021-04-21 05:28:39 indeed better use 0.3.7
2021-04-21 05:28:45 longer encoding time pay off
2021-04-21 05:29:26 i will test with more images to make eyes bleed
Crixis
2021-04-21 05:29:50 this is after 100 rencoding
2021-04-21 05:31:02 I litteraly rencoded 100 times
fab
2021-04-21 05:31:11 but with d 1.0
2021-04-21 05:31:16 i did modular and vardct lowest
2021-04-21 05:31:37 also faster decoding
2021-04-21 05:32:00 the one you watched was d 3.664
_wb_
2021-04-21 05:32:30 I think we just synced
monad
2021-04-21 05:33:15 finally!
Crixis
2021-04-21 05:38:49 no one will use modular lossy on foto
fab
2021-04-21 05:39:52 jpeg xl always use modular
2021-04-21 05:39:56 even at d 0.1
Crixis
2021-04-21 05:40:04 not lossy modular
fab
2021-04-21 05:40:21 ok
2021-04-21 05:40:30 but this has to be explained by developers
2021-04-21 05:41:17 https://ci.appveyor.com/project/EwoutH/jpeg-xl/builds/38812570
2021-04-21 05:41:21 build is cooking
Crixis
2021-04-21 05:41:28 in vardct mode encoder use modular lossless text and similiar
Scientia
2021-04-21 05:45:47 has it been tested how much patches can misdetect something like text?
2021-04-21 05:46:06 like I can envision a very large lossy image with a lot of text-like elements
_wb_
2021-04-21 05:46:32 It is safe wrt jbig-style scandals, if that's what you mean
Crixis
2021-04-21 05:46:41 you can test with djxl
2021-04-21 05:47:08 -s 6 if I remember
_wb_
2021-04-21 05:47:12 It works reasonably well but still misses many patch opportunities
Crixis
2021-04-21 05:47:31 work only on solid color
_wb_
2021-04-21 05:47:37 Yes
2021-04-21 05:47:58 That's one of the limitations of the current detection heuristics
Crixis
2021-04-21 05:50:12 -s 8
2021-04-21 05:50:32 blur is not detected
Scientia
2021-04-21 05:50:41 that's rather interesting
_wb_
2021-04-21 05:51:10 Basically you see the 1:8 image plus all patches when you do that
Scientia
2021-04-21 05:51:28 i wonder what is the distinction between the blurred text and the unblurred, since it's all on a single color
_wb_
2021-04-21 05:52:03 Well it only uses a patch if it can detect at least two copies of the same thing
Scientia
2021-04-21 05:52:12 ohh
2021-04-21 05:52:25 that makes more sense
2021-04-21 05:52:36 I understand it a little more now
_wb_
2021-04-21 05:52:41 To have a higher likelihood of actually being good for compression :)
Scientia
2021-04-21 05:53:27 so something like "G" wouldn't have a patch used, but something like "GGGGGGGG" would use patches?
2021-04-21 05:53:56 assuming this is the only thing with G in the image
Crixis
2021-04-21 05:54:09 2 g in different places also
Scientia
2021-04-21 05:54:25 just as long as the pixels are the same?
Crixis
2021-04-21 05:55:10 teoricaly also not pixel perfect match work but the ecoder now do only pixel perfect
Scientia
2021-04-21 05:55:18 hm
2021-04-21 05:55:44 that makes patching make more sense now, originally i thought it just made areas around text lossless
fab
2021-04-21 05:55:45 notice the thinner nose
2021-04-21 05:55:57 it even edit the nose
2021-04-21 05:56:01 the image i deleted
diskorduser
2021-04-21 05:56:16 Just now tested it with -d 2. there is no decrease in saturation.
Crixis
2021-04-21 05:56:23 this not sound like a bad idea also
Scientia
2021-04-21 05:56:55 also found it interesting that the patching was able to work with icons in your screenshot
_wb_
2021-04-21 05:57:15 The encoder in lossy mode does not require pixel perfect, but pretty close to it
Scientia
2021-04-21 05:57:20 maybe in the future most image with reused visual elements could use patching
Crixis
2021-04-21 05:57:27 i have find a drop of 1-2 point on red in -d 1 but not seam to be recursive
2021-04-21 05:57:47 from 200 to 198
Scientia
2021-04-21 05:57:53 like a discord screenshot could use patches for profile pictures maybe?
_wb_
2021-04-21 05:58:06 It maybe already does
Scientia
2021-04-21 05:58:09 hm
Crixis
2021-04-21 05:58:20 for status icon work
_wb_
2021-04-21 05:58:26 For things like small icons it does
2021-04-21 05:58:48 Not sure what the current max patch size is the encoder looks for
fab
2021-04-21 05:59:25 https://discord.com/channels/794206087879852103/794206170445119489/834467420609576961
2021-04-21 05:59:31 see these three screenshots
Scientia
2021-04-21 05:59:38 maybe it could be specified by a flag in cjxl in the future
fab
2021-04-21 05:59:40 i didn't see less saturation, but less appealing colours.
2021-04-21 05:59:57 lossy makes colors less appealing
2021-04-21 06:00:03 now i will try with newer build
2021-04-21 06:00:14 hope the undersaturation is less
Crixis
2021-04-21 06:00:21 lossy mean to be less appeling
fab
2021-04-21 06:00:34 i know
2021-04-21 06:01:00 but this has to be explained by developers
2021-04-21 06:01:19 i'm just copy pasting messages
Scientia
2021-04-21 06:01:19 hm
fab
2021-04-21 06:01:25 is not what i think
Scientia
2021-04-21 06:01:26 how is appealing measured
2021-04-21 06:01:30 i'm just interested
fab
2021-04-21 06:01:42 is just to find easier the messages
Scientia
2021-04-21 06:01:52 devs can't really do much unless you give a real measure of what appealing means
_wb_
2021-04-21 06:02:05 Are you talking about lossy modular or about VarDCT?
fab
2021-04-21 06:02:17 this one orim.png
2021-04-21 06:02:22 i encoded only with vardct
2021-04-21 06:02:26 but only crixis uploaded it
2021-04-21 06:02:34 the one from yt playlist
_wb_
2021-04-21 06:02:50 Also note that possibly your viewer is not displaying the original correctly
fab
2021-04-21 06:02:56 the other 2 files that looks blocky are different
diskorduser
2021-04-21 06:03:14 -d 2 ORI.png
fab
2021-04-21 06:03:15 i used png with both xnview
2021-04-21 06:03:21 xnview is the one it displays beter
Crixis
2021-04-21 06:03:23 cjxl use a custom metric with a name that start with B but i cannot remember
_wb_
2021-04-21 06:03:25 Viewers often do not do color management on images without explicit icc profile
fab
2021-04-21 06:03:38 butteraugli
_wb_
2021-04-21 06:03:52 Which means they show oversaturated colors if you have a P3 display
fab
2021-04-21 06:03:59 boott aragli i pronounce
2021-04-21 06:04:34 this is correct link
2021-04-21 06:05:05 my discord is not linking correct thumbnails
_wb_
2021-04-21 06:05:18 But the pngs that djxl produces always have colorspace info. So it can happen that original.png gets displayed oversaturated and decoded.png gets displayed correctly.
fab
2021-04-21 06:05:39 i know the difference with photoshop isn't really great
Scientia
2021-04-21 06:06:08 discord is having trouble on this one but it displayed if you clicked it
2021-04-21 06:06:22 you should have kept it up and not deleted it
2021-04-21 06:06:34 discord seems to be having some issues rn
Crixis
2021-04-21 06:06:48 in some there is a very small drop on red
fab
2021-04-21 06:06:52 i'm on windows 7
2021-04-21 06:07:12 with firefox nightly experimental
diskorduser
2021-04-21 06:07:17 Colors differences aren't that much visible. -_-
_wb_
2021-04-21 06:07:27 Firefox does not do color management properly
Crixis
2021-04-21 06:07:29 very small drop
Scientia
2021-04-21 06:07:41 fabian you don't need to keep deleting the image, discord is just having issues rn
diskorduser
2021-04-21 06:07:45 yeah very small.
fab
2021-04-21 06:07:57 i used the same viewer for both png
2021-04-21 06:08:01 encoded decoded
2021-04-21 06:08:15 and in 3 parts it looks underappealing
2021-04-21 06:08:25 i'm uploading in mega.nz
Crixis
2021-04-21 06:08:26 this not solve the issue
_wb_
2021-04-21 06:08:43 No, probably the original is displayed incorrectly
diskorduser
2021-04-21 06:08:44 That doesn't matter. check color profiles.
fab
2021-04-21 06:08:52 https://mega.nz/file/uB9kzSgQ#eWTymHjYMSvZ5hXv9XDJqjXkbl3MhNENPAxUArXyWM0
Crixis
2021-04-21 06:08:52 a lot of viewer do bad color png
fab
2021-04-21 06:09:04 no profile is jxl encoded
Scientia
2021-04-21 06:09:06 can you screenshot the unappealing parts and the equivalent appealing parts maybe?
fab
2021-04-21 06:09:07 to png
2021-04-21 06:09:09 with djxl
2021-04-21 06:09:19 https://mega.nz/file/uB9kzSgQ#eWTymHjYMSvZ5hXv9XDJqjXkbl3MhNENPAxUArXyWM0
_wb_
2021-04-21 06:09:23 Try this: https://cameratico.com/color-management/firefox/#:~:text=By%20default%2C%20Firefox%20only%20enables,specially%20on%20wide%20gamut%20displays.
fab
2021-04-21 06:09:36
2021-04-21 06:09:42 that's source
Crixis
2021-04-21 06:09:55 the last pink is not perfect
2021-04-21 06:10:17 👀
diskorduser
2021-04-21 06:10:25 Open this in gimp and check diffrences by turning on and off the layers <@!416586441058025472>
Crixis
2021-04-21 06:10:52 left is more grey
fab
2021-04-21 06:11:21
2021-04-21 06:11:42 this is the output crixis did with cjxl
2021-04-21 06:11:48 do you did lossy
2021-04-21 06:11:57 i have not checked it
2021-04-21 06:12:02 this i should do
2021-04-21 06:12:03 sorry
2021-04-21 06:12:13 i should re check
2021-04-21 06:13:07 i use windows 7 default viewer but this was all captured with xnview
2021-04-21 06:13:23 also the screenshot of ORI.png (source)
2021-04-21 06:14:10 anyway i send 300 messages today
2021-04-21 06:14:20 i gonna stop in hashtag jxl
2021-04-21 06:14:30 thanks
Scientia
2021-04-21 06:15:47 what does the E flag do in cjxl? i've seen it used in encode comparisons to get the best compression but it's not in the help or the verbose help.
BlueSwordM
2021-04-21 06:16:10 Do `cjxl -h -v -v -v` to look at it.
Scientia
2021-04-21 06:16:15 oh
2021-04-21 06:16:17 hahaha
2021-04-21 06:16:42 that's quite a way to be verbose haha
Crixis
2021-04-21 06:16:58 oof, this show that the color is perfect so is a visual effect
Scientia
2021-04-21 06:17:24 I mean this is expected
2021-04-21 06:17:33 lossy can't preserve everything by nature
2021-04-21 06:17:41 but if it can perceptually perserve it
2021-04-21 06:17:47 then it's good
Crixis
2021-04-21 06:17:47 E make color channels see each other, work only with s 9
veluca
2021-04-21 06:18:04 this makes me think, a probable improvement for patches is to also use non-reused patches that make the background significantly smoother
2021-04-21 06:18:17 should help for text...
Crixis
2021-04-21 06:18:43 yes but is slow, i use s 9 E 3 as an s10
Scientia
2021-04-21 06:19:06 so 3 is for the 3 color channels?
2021-04-21 06:19:10 or am I wrong
BlueSwordM
2021-04-21 06:19:13 Yes.
fab
2021-04-21 06:19:18
2021-04-21 06:19:45 my skin
2021-04-21 06:19:55 with new jpeg xl a bit better
Scientia
2021-04-21 06:20:24 instead of skin you might want to use some color test images
fab
2021-04-21 06:20:26 i did
2021-04-21 06:20:27 https://discord.com/channels/794206087879852103/794206170445119489/834480278080192600
2021-04-21 06:20:34 but without the modular lossy
2021-04-21 06:20:46 i will test new settings after
2021-04-21 06:20:58 i have the youtube screenshot
2021-04-21 06:21:01 in webp
2021-04-21 06:21:17 i can't see more colours than those
2021-04-21 06:21:44 scope probably know better images
Scientia
2021-04-21 06:23:19 I mean like this
2021-04-21 06:24:46 photographic might be good but I don't think testing an image from a phone camera or anything similar would yield that big of a difference
2021-04-21 06:24:57 phone pictures are pretty bad already
2021-04-21 06:26:13 also new commit to the jxl repo, going to try jxl from tree now ( :
fab
2021-04-21 06:26:18 gimp display better
Crixis
2021-04-21 06:27:07 gimp is a good piece of software
fab
2021-04-21 06:27:49 how you did
2021-04-21 06:27:57 do you used more recent build
2021-04-21 06:28:05 which commit and which distance speed
diskorduser
2021-04-21 06:29:24 -d 2
fab
2021-04-21 06:29:32 speed and commit?
diskorduser
2021-04-21 06:29:43 Stable builds
2021-04-21 06:29:53 Latest stable release
fab
2021-04-21 06:29:55 the one from today or the older
diskorduser
2021-04-21 06:32:30 0.3.7 release
fab
2021-04-21 06:32:33
2021-04-21 06:32:33
2021-04-21 06:33:01 not the one from today?
2021-04-21 06:33:13 do you used ab7c5e9b
2021-04-21 06:33:16 be honest
2021-04-21 06:33:28 or speed 9
2021-04-21 06:33:31 speed 8
diskorduser
2021-04-21 06:33:52
2021-04-21 06:34:36 Speed 7
fab
2021-04-21 06:34:47 ok i'll move on benchmarks
2021-04-21 06:42:30 there is no difference between 48 and 50 commit
2021-04-21 06:42:37 is the same in vardct
2021-04-21 06:42:43 same filesize
diskorduser
2021-04-21 06:42:45 Always check color profiles of source images before coming to conclusions. Most of color difference are likely due to icc profiles.
fab
2021-04-21 06:50:53 the 0.3.8 decoder improves maybe the result
2021-04-21 06:51:09 i don't know if png sizes with decoder are same
diskorduser
2021-04-21 06:52:43 Are you using git builds?
fab
2021-04-21 06:54:02 no the decoder is the same
2021-04-21 06:54:08 now i will check new image
2021-04-21 06:54:16 (on gimp)
2021-04-21 06:57:03 i see a 19,8% and 23,6% drop grey
2021-04-21 06:57:05 plus
2021-04-21 06:57:06 a
2021-04-21 06:58:31 0,68% on white
2021-04-21 06:58:39 2,33% on blue
2021-04-21 06:58:54 8,88% on orange/red
2021-04-21 06:59:06 0,19-0,33% on green
2021-04-21 06:59:19 compared to 0.3.8
diskorduser
2021-04-21 07:01:26 Are you trying to say that you get difference on color b/w the builds?
fab
2021-04-21 07:03:10
2021-04-21 07:03:27 on this is zero difference
2021-04-21 07:03:35 but on other parts i see but not evident
2021-04-21 07:03:43 at least now i can't see more
2021-04-21 07:03:48 but the file size are different
2021-04-21 07:03:58 don't know if they are really improvement of quality for d 1
2021-04-21 07:04:07 i think there aren't
2021-04-21 07:04:18 i didn't do a fair comparison
Scientia
2021-04-21 07:05:34 Isn't d 1 perceptually lossless?
fab
2021-04-21 07:06:16 i see no difference
2021-04-21 07:06:40
2021-04-21 07:06:46 i perceived something
diskorduser
2021-04-21 07:07:07 D1. It's visually lossless when viewed from certain distance (at least for me lol)
Scientia
2021-04-21 07:07:13 Oh
fab
2021-04-21 07:08:48 i don't perceiveD nothing
diskorduser
2021-04-21 07:10:38 >i perceived something >i don't perceiveD nothing I cry
fab
2021-04-21 07:11:28
2021-04-21 07:13:12
2021-04-21 07:13:15 look at it
2021-04-21 07:13:17 motorola song
2021-04-21 07:13:20 at bottom
2021-04-21 07:13:32 there is a big difference
Scientia
2021-04-21 07:13:55 If you don't show the original image we can't tell
diskorduser
2021-04-21 07:14:14 Hello moto
fab
2021-04-21 07:15:02 i did ss folder
2021-04-21 07:17:27 s 7 d 1
2021-04-21 07:17:40 source
2021-04-21 07:18:24 but he didn't said they weren't image quality improvement on the month
2021-04-21 07:20:05 now i will use a normal photo
2021-04-21 07:26:00 sorry i did a dssim
2021-04-21 07:26:19
2021-04-21 07:26:19
2021-04-21 07:26:52 does dssim can relevate if the image has exactly same pixels
2021-04-21 07:27:12 ANYWAY
2021-04-21 07:27:16 jpeg xl decoding
2021-04-21 07:27:18 11,3 mpx
2021-04-21 07:27:29 for 2 threads
2021-04-21 07:27:32 22,2 mpx
2021-04-21 07:27:52 on a 600 geekbench 5 3000 geekbench 4 toast
2021-04-21 07:27:59 very improved
2021-04-21 07:28:06 and you don't need to re encode
2021-04-21 07:30:27 jyri didn't disappoint
2021-04-21 07:49:27 the best is the encoder is at same speed
2021-04-21 08:05:27 how they did improvements to image quality of red colours
2021-04-21 08:05:33 in only a day
2021-04-21 08:05:39 sure there aren't errors in code
diskorduser
2021-04-22 03:28:02 On which build?
fab
2021-04-22 07:01:24 600mb ram for a 2000x3000
2021-04-22 07:05:37 it already finised for 30 images
2021-04-22 07:12:25 the other 30% 20% 15% bigger
2021-04-22 07:12:59 my jxl build hasn't jpg support
2021-04-22 07:13:14 if someone has a build with jpg support
2021-04-22 08:48:20
2021-04-22 08:49:59 modular is right
2021-04-22 05:06:42 what are those beep sound in cjxl
2021-04-22 05:06:51 i heard in 0.239 bpp compression
2021-04-22 05:07:08 when i compressed three images in a row at d1 at that bpp
2021-04-22 05:07:19 also when it failed to read image
2021-04-22 05:07:26 but it was louder
2021-04-22 05:07:50 when it don't encode image it don't make any sound
2021-04-22 05:08:03 even when there is image.cc crash
2021-04-22 05:08:06 and then cjxl crash
Crixis
2021-04-22 05:15:09 <:NotLikeThis:805132742819053610>
Scientia
2021-04-22 05:18:28 Hey <@416586441058025472> have you tried messing around with the MA trees with jxl_from_tree yet?
diskorduser
2021-04-22 05:22:41 it beeps because it says f**k. 🤣
fab
2021-04-22 05:29:04 No seriously veluca and jon Said it will be integrated
2021-04-22 05:29:15 For cmd
Scope
2021-04-22 05:31:42 It's time to make jxl-beep-music tools
Scientia
2021-04-22 05:33:03 I've never heard Jon say the CLI tools would use audio output
_wb_
2021-04-22 05:48:46 They don't
2021-04-22 05:49:29 If cjxl beeps, something very weird must be happening
monad
2021-04-22 06:02:56 As I recall, Scope made a joke about adding beeps, but nothing more.
2021-04-22 06:03:57 In this discussion https://discord.com/channels/794206087879852103/794206087879852106/821797971506888734
fab
2021-04-22 06:04:00 With s7 d1
2021-04-22 07:03:56 I can't use jxl from tree
Scientia
2021-04-22 07:07:33 why not?
fab
2021-04-22 07:08:12 do not know how to do commands
2021-04-22 07:08:28 i can't open a .cc file and compile it
2021-04-22 07:08:31 with windows 7
_wb_
2021-04-22 07:09:08 Is it not in the windows builds?
2021-04-22 07:09:17 I suppose you could wait for the web tool
Nova Aurora
2021-04-22 07:09:58 wasm jxl trees?
2021-04-22 07:10:13 new favorite timewasting website
_wb_
2021-04-22 07:12:38 https://jxl-art.surma.technology/
2021-04-22 07:13:14 But it still needs to be updated I suppose to get the new things like Squeeze and custom bit depth