|
Jyrki Alakuijala
|
2021-07-27 02:49:26
|
d16 😕
|
|
2021-07-27 02:49:57
|
I don't think any real users should ever use this bad quality, but some people doing compression comparisons will use it 😄
|
|
|
Cool Doggo
|
2021-07-27 02:50:11
|
I thought d15 was lowest?
|
|
|
|
Jyrki Alakuijala
|
2021-07-27 02:50:27
|
that's great if that is the case
|
|
2021-07-27 02:50:46
|
in benchmark_xl you can specify practically any value and it works
|
|
2021-07-27 02:51:03
|
I have done d128 in some previous optimizations (it is of course complete non-sense)
|
|
2021-07-27 02:51:14
|
this is already at 0.07 bpp
|
|
|
|
Cool Doggo
|
2021-07-27 02:52:49
|
I believe if you do target size the lowest it will show is 16 but you can't do d16 on its own iirc
|
|
|
|
Jyrki Alakuijala
|
2021-07-27 02:54:33
|
yes, target size is something that no actual user would do
|
|
2021-07-27 02:54:51
|
it is only practical if you want to match some other slower codec
|
|
2021-07-27 02:55:04
|
even then it is questionable
|
|
2021-07-27 02:55:14
|
file size matching should always be done at corpus level
|
|
2021-07-27 02:55:58
|
different codecs have different amounts of complexity, and different strengths and weaknesses by design
|
|
2021-07-27 02:56:21
|
forcing a codec to adapt into strengths of another codec will make it behave erraticly
|
|
|
lithium
|
2021-07-27 03:24:16
|
<@!532010383041363969> Thank you for your work 🙂
jxl-v0.5-b97a929
cjxl -d 0.5 -s 9 red and blue area have some quality improve,
but color gradient area still some tiny issue,
maybe this situation still have some space for quality improve?
I hope -d 0.5 -s 9 can near transparency quality for anime content.
|
|
|
|
Jyrki Alakuijala
|
2021-07-27 03:41:21
|
I promise to fix all -d 0.5 -s 9 problems
|
|
|
spider-mario
|
2021-07-27 03:57:37
|
targeting a given size is what I do when uploading JPEGs to Discord or imgur as they have a size limit (8MB and 5MB respectively iirc, less on imgur if not logged in)
|
|
2021-07-27 03:57:49
|
so I compress at the highest quality that is still below the limit
|
|
2021-07-27 03:58:00
|
(I have a small perl script that does that by binary search)
|
|
|
|
Jyrki Alakuijala
|
2021-07-27 04:31:46
|
Targeting a 8 MB or 5 MB image size is better than tarketing 8 kB or 5 kB 😛
|
|
2021-07-27 04:32:30
|
if d1 images are 1.5 bpp, you can have 42 megapixel images at that for 8 MB 😄
|
|
|
|
lithium
|
2021-07-27 04:43:06
|
I thought more balanced overall quality is a good idea, I don't like compressed have a big error.
( hope -d 0.5 -s 9 like webp near-lossless 60 have transparency quality,
-d 1.0 -s7~-s9 like webp near-lossless 40, have some risk and lost some detail but still good,
I still love webp near-lossless, that is a great lossy mode 🙂 )
|
|
|
|
Jyrki Alakuijala
|
2021-07-27 05:56:17
|
I'm proud of it, but I have learned more tricks in the last 10 years
|
|
|
BlueSwordM
|
2021-07-27 06:33:20
|
So, I've just reencoded my test sets again, and I've noticed a particular image posing some problems...
https://cdn.discordapp.com/attachments/668978554586791957/869648279250563112/Variance_edge_issue.7z
Essentially, the JXL image is encoded quite well at speed 8(speed 7 has the same issue, except a bit worse), except for the fact that on the left, the darker area of the image where the pillars reside are completely destroyed in the JXL image, while tuned aomenc with AVIF still preserves rather well.
|
|
2021-07-27 06:33:54
|
cjxl: `cjxl input.png output.jxl -s 8 -d 1.0`
avifenc using aomenc: `avifenc -s 4 -j 16 --min 0 --max 63 -a end-usage=q -a cq-level=15 -a color:sharpness=2 -a tune=butteraugli -a color:qm-min=0 -a color:deltaq-mode=3 input.png output.avif`
|
|
|
|
Jyrki Alakuijala
|
2021-07-27 10:24:33
|
could you just post a small png demonstrating the effect here?
|
|
2021-07-27 10:24:42
|
(-: chromebook 🙂
|
|
|
|
BlueSwordM
|
|
Jyrki Alakuijala
could you just post a small png demonstrating the effect here?
|
|
2021-07-28 04:08:44
|
All images contained used for comparison are in PNG, as it makes viewing much easier in this regard.
However, I have just realized that I could just have posted a normal comparison on slow.pics instead.
|
|
|
Jyrki Alakuijala
(-: chromebook 🙂
|
|
2021-07-28 05:39:25
|
https://slow.pics/c/W3iPm4nn
|
|
|
frank cilantro
|
2021-07-28 07:02:56
|
ive tried looking through the codebase but im not quite sure which location in the code would let me play around with the vardct blocks that get chosen.. any pointers?
|
|
|
_wb_
|
2021-07-28 07:10:57
|
`enc_ac_strategy.cc` is where the block selection is done
|
|
|
|
frank cilantro
|
|
|
BlueSwordM
|
|
BlueSwordM
https://slow.pics/c/W3iPm4nn
|
|
2021-07-29 02:59:50
|
Anything new regarding this?
|
|
|
raysar
|
|
BlueSwordM
Anything new regarding this?
|
|
2021-07-29 03:21:29
|
jxl is visually better even in dark area (watch at 200% zoom). Always keep more grain and details.
|
|
|
|
BlueSwordM
|
|
raysar
jxl is visually better even in dark area (watch at 200% zoom). Always keep more grain and details.
|
|
2021-07-29 04:09:13
|
Look on the left...
|
|
2021-07-29 04:10:43
|
It is pretty obvious cjxl is spending on the bits everywhere in the picture, except for the bottom left spot.
|
|
2021-07-29 04:12:06
|
Average quality may be higher on the JXL image(to be fair, it is 18% bigger than the AVIF copy) but the bottom left spot makes it unusable.
|
|
2021-07-29 04:12:19
|
I'd prefer it if the image did not exhibit this behavior.
|
|
2021-07-29 04:12:50
|
Upping the brightness, bringing the quality up to -d 0.5, nothing seems to completely remove it unless I go lossless.
|
|
|
|
raysar
|
2021-07-29 04:13:34
|
Yes some low frequency detail on very low light is better with aom. We can see it on top left and bottom left in dark area.
|
|
|
|
BlueSwordM
|
|
raysar
Yes some low frequency detail on very low light is better with aom. We can see it on top left and bottom left in dark area.
|
|
2021-07-29 04:14:01
|
Yeah, and I would like to know why jxl is spending the bits everywhere very well in the rest of the image, but not that spot.
|
|
2021-07-29 04:14:05
|
It's very weird.
|
|
|
|
raysar
|
2021-07-29 04:17:25
|
Test the same picture with --intensity_target 400 (and a bit higher value to keep the same file size)
|
|
2021-07-29 04:18:27
|
we all have very bright display
|
|
|
|
BlueSwordM
|
|
raysar
Test the same picture with --intensity_target 400 (and a bit higher value to keep the same file size)
|
|
2021-07-29 04:21:21
|
Same thing.
|
|
2021-07-29 04:21:23
|
It doesn't fix it.
|
|
2021-07-29 04:21:31
|
It looks to be a bpp allocation issue.
|
|
|
Petr
|
2021-07-29 05:49:52
|
Does anyone know why there are no files available at https://ci.appveyor.com/project/EwoutH/jpeg-xl/build/job/a3qainag5pdepy8a/artifacts even though there should be 6 of them?
|
|
|
|
lithium
|
|
BlueSwordM
It looks to be a bpp allocation issue.
|
|
2021-07-29 05:55:51
|
I guess vardct heuristic probably still have some issue?
(for anime content, sometime will get some unexpected quality issue in -d 0.5 and 1.0)
|
|
|
|
Petr
|
|
Petr
Does anyone know why there are no files available at https://ci.appveyor.com/project/EwoutH/jpeg-xl/build/job/a3qainag5pdepy8a/artifacts even though there should be 6 of them?
|
|
2021-07-29 06:08:40
|
I see the messages "Artifacts older than 1 month are automatically deleted" and "Failed 3 months ago". But was the last build attempt really soooo long ago? 😮
|
|
|
|
BlueSwordM
|
|
Petr
Does anyone know why there are no files available at https://ci.appveyor.com/project/EwoutH/jpeg-xl/build/job/a3qainag5pdepy8a/artifacts even though there should be 6 of them?
|
|
2021-07-29 06:08:41
|
All builds seem to have failed.
|
|
|
|
lithium
|
2021-07-29 06:08:49
|
I guess jxl appveyor is broken?
|
|
|
|
BlueSwordM
|
2021-07-29 06:08:52
|
But yes, it has been more than 1 month.
|
|
|
|
lithium
|
2021-07-29 06:15:08
|
> jxl appveyor-master-ab7c5e9b
> libjxl github-main-1a932c6 (1a932c6c42380718b632ccecb7bb840c3e31c068)
> jpeg xl gitlab-main-1a932c6c (1a932c6c42380718b632ccecb7bb840c3e31c068)
🤔
|
|
|
|
_wb_
|
|
BlueSwordM
It is pretty obvious cjxl is spending on the bits everywhere in the picture, except for the bottom left spot.
|
|
2021-07-29 08:12:29
|
I have difficulties seeing where there is a problem. What spot specifically is problematic?
|
|
2021-07-29 08:16:02
|
the scarf on the left-most person does get some artifacts in the jxl image, is that the issue?
|
|
2021-07-29 08:18:50
|
the dark red coat of the woman next to him also gets artifacts
|
|
2021-07-29 08:19:50
|
I did need to zoom in to see that though
|
|
|
|
raysar
|
|
_wb_
I did need to zoom in to see that though
|
|
2021-07-29 11:03:14
|
look at the floor at the extreme bottom left, it's blur. But all the rest of the image is near visual lossless. And avif is better in this zone.
|
|
|
|
_wb_
|
2021-07-29 11:37:12
|
yes, I can see the blur there when I zoom in
|
|
2021-07-29 11:38:00
|
I have a really hard time seeing it when I don't zoom in though
|
|
2021-07-29 11:39:20
|
<@!532010383041363969> is the viewing distance / zoom factor something that could be adjusted in Butteraugli (like the intensity target), or is it fixed?
|
|
|
|
BlueSwordM
|
|
_wb_
I have difficulties seeing where there is a problem. What spot specifically is problematic?
|
|
2021-07-29 04:30:02
|
|
|
2021-07-29 04:30:05
|
|
|
|
Scope
|
2021-07-30 11:17:56
|
So, after adding a delta palette:
https://github.com/libjxl/libjxl/pull/385
|
|
2021-07-30 11:18:25
|
My previous request makes more sense, but `-P 0 --palette=0`, without `-I 0`:
https://github.com/libjxl/libjxl/issues/119
|
|
|
kb
|
2021-07-31 12:44:16
|
sorry if I'm asking something obvious, a search didn't turn anything up: is there some real web page using / showcasing jxl today? particularly interested in showing off different features (ie. original jxl pics, old transcoded jpegs, animated jxls, transparency, transparency + animation, multi-channel / multispectra images, etc)
|
|
|
fab
|
2021-07-31 09:37:54
|
is something like that possible with jxl?
|
|
2021-07-31 09:37:55
|
-s 8 -d 1.23 -E 2 after 5 builds
|
|
2021-07-31 09:38:25
|
i want to compress more without destroying visual quality
|
|
2021-07-31 09:38:44
|
i think jxl for compressing real faces is mediocre
|
|
2021-07-31 09:38:53
|
they weight many kb
|
|
2021-07-31 09:38:56
|
like 118 kb
|
|
2021-07-31 09:39:16
|
i know is because they are fatter
|
|
2021-07-31 09:39:30
|
but jxl seems to be optimized for anime,
|
|
2021-07-31 09:39:43
|
for cartoons, reductions of ringing and not for real faces
|
|
2021-07-31 09:41:32
|
probably wb won't agree, as wb has never show agreement on the what thing i said
|
|
2021-07-31 09:42:19
|
i think he knows how to market jxl, and if fb or windows 11 (with an extension or without) or even windows 10 will receive jxl
|
|
2021-07-31 09:42:42
|
or if nomacs can be signal lossless or modular and start faster
|
|
2021-07-31 09:43:00
|
wb knows everything and can design sites very well, jpeg xl info designed himself
|
|
2021-07-31 09:43:15
|
he just can't spoiler
|
|
2021-07-31 09:43:29
|
wants to build hype
|
|
2021-07-31 09:44:20
|
and it isn't his job
|
|
2021-07-31 09:44:24
|
. as he said
|
|
|
w
|
|
diskorduser
|
|
fab
i think jxl for compressing real faces is mediocre
|
|
2021-07-31 09:56:41
|
Are you encoding with lossless source like developed raw / DNG / png / tiff?
|
|
2021-07-31 09:57:24
|
Or using some webp compressed youtube thumbnails?
|
|
|
|
fab
|
2021-07-31 10:05:42
|
big question
|
|
2021-07-31 10:06:18
|
jpeg
|
|
2021-07-31 10:06:25
|
it was png
|
|
2021-07-31 10:06:43
|
but anyway is hdr almalence camera so bad quality
|
|
2021-07-31 10:06:51
|
no raw at all
|
|
2021-07-31 10:07:29
|
the thing is that precision is boring
|
|
|
|
_wb_
|
|
kb
sorry if I'm asking something obvious, a search didn't turn anything up: is there some real web page using / showcasing jxl today? particularly interested in showing off different features (ie. original jxl pics, old transcoded jpegs, animated jxls, transparency, transparency + animation, multi-channel / multispectra images, etc)
|
|
2021-07-31 10:37:08
|
It would be good to add some demo page(s) to jpegxl.info, besides the jxl art there (which is not very representative of what jxl is made for)
|
|
|
|
Scope
|
2021-07-31 10:44:12
|
Yep, I think it would be nice, like on the FLIF site <https://flif.info/example.html>
|
|
|
|
diskorduser
|
|
fab
no raw at all
|
|
2021-07-31 10:55:34
|
If you don't have lossless source, don't complain about quality then.
|
|
|
|
_wb_
|
2021-07-31 02:32:08
|
I don't know what this is supposed to mean but it sounds like it might be inappropriate language.
|
|
|
|
fab
|
2021-07-31 02:32:28
|
yes it is
|
|
2021-07-31 02:33:00
|
it's only a fake news
|
|
2021-07-31 02:33:09
|
twitter is damaging my mind
|
|
2021-07-31 02:33:27
|
anyway vardct does too perfect vardct block selection
|
|
2021-07-31 02:33:32
|
the scope is this
|
|
2021-07-31 02:33:43
|
to optimize for numbers, for butteruagli
|
|
2021-07-31 02:33:51
|
and to compress
|
|
2021-07-31 02:34:15
|
not to have quality at normal quality and for representing people that are strange
|
|
2021-07-31 02:34:26
|
probably in internet there aren't much of those faces
|
|
2021-07-31 02:34:45
|
i think less than 1% of the entire face
|
|
2021-07-31 02:35:15
|
since image developers don't share them/use as stock photo or in rare cares not being bad but is the reality
|
|
2021-07-31 02:35:31
|
and anyway jpeg xl is meant for screenshots/synthethic images
|
|
2021-07-31 02:35:44
|
but i'm interested to delta for the file sizes
|
|
2021-07-31 02:35:52
|
did this commit reduce file size?
|
|
2021-07-31 02:35:56
|
for that type of image?
|
|
2021-07-31 02:36:10
|
so i need this commit of iulia comsa
|
|
2021-07-31 02:36:16
|
i think colors looks great
|
|
2021-07-31 02:36:19
|
i'm no expert
|
|
|
_wb_
I don't know what this is supposed to mean but it sounds like it might be inappropriate language.
|
|
2021-07-31 02:36:52
|
i will delete the message
|
|
|
|
veluca
|
2021-07-31 02:37:12
|
We should avoid this kind of language, it creates a toxic environment - I don't particularly want to ban people, but at some point it becomes the best option...
|
|
|
|
fab
|
2021-07-31 02:37:58
|
please not
|
|
2021-07-31 02:38:14
|
in av1 server they never ban anyone, they kick and they re enter
|
|
2021-07-31 02:38:20
|
after a day with no moderation
|
|
|
|
veluca
|
2021-07-31 02:38:59
|
there are such things as temporary bans, but I don't think I'll have to resort to that if we keep the conversation civil and respectful 😉
|
|
|
|
Deleted User
|
2021-07-31 04:49:59
|
<:JXL:805850130203934781>
|
|
|
|
_wb_
|
2021-07-31 05:49:53
|
<:logo:829708783336816671>
|
|
|
monad
|
2021-07-31 06:33:47
|
Just to clarify, mentioning autists is not allowed? or was there another comment deleted?
|
|
|
|
_wb_
|
2021-07-31 06:41:09
|
Mentioning autists is fine, weird statements about the faces of autists that's where it gets more questionable
|
|
|
|
monad
|
2021-07-31 06:51:51
|
Well, if it was just the one comment, I must disagree with the interpretation and the ban threat.
|
|
|
|
fab
|
2021-07-31 08:39:22
|
yes but i hate delta, vardct, delta for 720p 300 kbps
|
|
2021-07-31 08:39:26
|
all those things
|
|
2021-07-31 08:39:30
|
i want simpler things
|
|
2021-07-31 08:39:35
|
maybe i ask too much
|
|
2021-07-31 08:39:58
|
dev does research and pay on amazon to find 30 euros per hour to do subjective evaluation tests
|
|
2021-07-31 08:40:04
|
i think lot of money is going on
|
|
2021-07-31 08:40:21
|
maybe i don't understand also what the effort being in jpeg xl
|
|
2021-07-31 08:40:30
|
the technology, the difficulty
|
|
2021-07-31 08:41:28
|
i also like s 8 d 1.23 -E 2 and i don't find any bad about it
|
|
2021-07-31 08:41:40
|
have not tried it
|
|
2021-07-31 08:41:46
|
jpeg xl evolved
|
|
2021-07-31 08:41:54
|
4 months of active development
|
|
2021-07-31 08:44:02
|
the metric, the butteraugli, it doesn't inflate if you have an image in jxl or at least it doesn't too much, it's optimized for generation loss, big resolutions without tiles, fast speed, the only thing missing is software adoption.
|
|
2021-07-31 08:44:52
|
the perfect block vardct selection optimized for ringing and for anime girls, the delta by iulia comsa and the amazing work she done on colour and codec.
|
|
2021-07-31 08:45:51
|
obviously wb gone crazy if someone still criticize ah it doesn't compress p... it doesn't compress aut.... i censored the word but other users didn't at all, they say chinese, pakistan all possible words how they want
|
|
2021-07-31 08:46:49
|
because wb has done work, testing the encoder optimizing maths for butteraugli and after doing the math after the butteraugli, veluca has done most of the math and the decoder
|
|
2021-07-31 08:47:15
|
so we have to not press too much the devs
|
|
2021-07-31 08:47:30
|
we don't know nothing about jpeg xl
|
|
2021-07-31 08:47:56
|
yes there are people compiling who knows like more
|
|
2021-07-31 08:48:51
|
but is normal
|
|
2021-07-31 08:48:57
|
people always complain in 2020
|
|
2021-07-31 08:49:06
|
about the battery of a car abourt everything
|
|
2021-07-31 08:50:13
|
also high quality at screenshots
|
|
|
|
Cool Doggo
|
2021-07-31 09:12:24
|
that is a lot of text 🤯
|
|
|
|
kb
|
|
_wb_
It would be good to add some demo page(s) to jpegxl.info, besides the jxl art there (which is not very representative of what jxl is made for)
|
|
2021-07-31 09:50:08
|
is there a channel in here for collecting samples? (and voting on them!)
|
|
|
|
monad
|
2021-07-31 09:51:12
|
No dedicated channel.
|
|
|
|
necros
|
|
_wb_
Haven't tried any of the viewers yet, except for that MacOS one and `eog`, and both don't support animation yet
|
|
2021-07-31 11:39:15
|
is it lossless?
|
|
|
|
w
|
2021-07-31 11:58:54
|
ah yes fabian wall of text
|
|
|
190n
|
|
necros
is it lossless?
|
|
2021-08-01 04:09:38
|
hm? eog supports lossless jxls i can confirm
|
|
|
|
fab
|
2021-08-01 06:18:04
|
the point is just respect jon sneyers and ignore av1 server way of behaviour
|
|
2021-08-01 06:18:19
|
and that jpeg xl has evolved in just 4 months
|
|
2021-08-01 06:18:38
|
the magic is that is many work done and it's yes simple
|
|
|
|
BlueSwordM
|
|
fab
the point is just respect jon sneyers and ignore av1 server way of behaviour
|
|
2021-08-01 06:18:39
|
*The point is just to respect everyone.
|
|
2021-08-01 06:18:42
|
🦧
|
|
|
|
fab
|
2021-08-01 06:19:05
|
latest encoder uses only 500 mb for 4000x5000 photos (20 mpx)
|
|
2021-08-01 06:19:29
|
the one eclipseo shared (user that for second created the comparison in github)
|
|
2021-08-01 06:19:41
|
many images can't be inserted, there is a size limit of github pages
|
|
|
|
_wb_
|
2021-08-01 06:24:17
|
Is there a size limit?
|
|
2021-08-01 06:25:56
|
Anyway, we don't need a huge gallery of example images, and they don't need to be huge images.
|
|
|
eddie.zato
|
2021-08-02 05:10:53
|
Will there be more improvements to the "jpeg lossless transcode"? I plan to convert ~100 gigabytes of jpeg to jxl. Since good image viewers already have support for jxl, there's nothing stopping me from saving some disk space.
|
|
|
|
raysar
|
2021-08-02 06:08:07
|
There is a good question on reddit, if there is a way to add jxl decoding solution for browser without this feature?
A greasemonkey script using the wasm or javascript jxl decoder could be easy to write for a web dev?
|
|
|
|
_wb_
|
2021-08-02 07:00:25
|
You could do that, but it would only make sense for below-the-fold images
|
|
2021-08-02 07:01:24
|
For above-the-fold images, picture srcset with avif/webp/jpeg fallback(s) will be faster
|
|
2021-08-02 10:05:05
|
this is a thread
|
|
|
|
fab
|
2021-08-02 10:16:24
|
for %i in (C:\Users\User\Documents\a*.png) do cjxl -q 71.7 -s 2 -I 3.45 %i %i.jxl eclipse build
for %i in (D:\Documenti\jpg*.jpg) do cjxl -j -s 7 -d 1 %i %i.jxl
|
|
2021-08-02 10:18:06
|
|
|
2021-08-02 11:54:19
|
i'm sending new parameters for jxl eclipse
|
|
2021-08-02 11:56:55
|
|
|
2021-08-02 12:19:11
|
eclipsebuild
1) for %i in (C:\Users\User\Documents\a.png) do cjxl -q 76.215 -s 9 %i %i.jxl
2) for %i in (C:\Users\User\Documents\a.png) do cjxl --photon_noise=ISO882 -d 1.781 -s 8 -I 0.431 --epf=2 --patches=0 --gaborish=1 %i %i.jxl
3) for %i in (C:\Users\User\Documents\a.png) do cjxl -d 7.476 --photon_noise=ISO400 --epf=2 --patches=0 --gaborish=1 %i %i.jxl
4) for %i in (C:\Users\User\Documents\a.png) do cjxl -q 74 -s 9 %i %i.jxl
5) for %i in (C:\Users\User\Documents\a.png) do cjxl -d 1.11 -s 9 -I 0.38 %i %i.jxl 6) for %i in (C:\Users\User\Documents\a.png) do cjxl -s 6 -q 75.152 --epf=1 --faster_decoding=2 %i %i.jxl
7) for %i in (C:\Users\User\Documents\a.png) do cjxl -s 7 -q 55.78 --epf=1 --faster_decoding=2 %i %i.jxl
8) for %i in (C:\Users\User\Documents\a.png) do cjxl -s 7 -d 0.768 --gaborish=1 --epf=2 -I 1 --use_new_heuristics %i %i.jxl
9) for %i in (C:\Users\User\Documents\a.png) do cjxl -s 2 -N 3 -q 100 -m %i %i.jxl
|
|
2021-08-02 12:19:39
|
Don't know if photon noise iso 882 works
|
|
2021-08-02 12:19:51
|
and if near lossless 3 still requires q 100
|
|
2021-08-02 12:20:31
|
choosing different values of photon noise probably will not destroy the result
|
|
2021-08-02 12:20:35
|
so stay calm
|
|
2021-08-02 12:21:50
|
for discord this parameter can be useful
|
|
2021-08-02 12:22:13
|
|
|
2021-08-02 12:24:57
|
Also the encoder is optimized at maximum it uses 500 mb ram at speed 7 for 4000x5000, the one i download i will re give the link
|
|
2021-08-02 12:25:34
|
eclipseo
—
Oggi alle 21:56
For those looking for a Windows build of the latest GIT tip: https://www.dropbox.com/s/bpog8d4ca2j02dk/jxl_20210728.zip?dl=0 <@!416586441058025472>
|
|
2021-08-02 01:07:21
|
edited the parameters from 1 to 5
|
|
2021-08-02 01:07:51
|
|
|
2021-08-02 01:09:36
|
i need to try those parameters with real photos
|
|
2021-08-02 01:13:02
|
|
|
2021-08-02 01:13:04
|
|
|
2021-08-02 01:14:23
|
I did the 5° command
|
|
2021-08-02 01:14:25
|
worked
|
|
2021-08-02 01:42:50
|
|
|
2021-08-02 01:42:53
|
AV1 CODEC SAMPLE IMAGE
|
|
2021-08-02 01:43:28
|
9 July 2021
|
|
|
|
_wb_
|
2021-08-02 01:51:53
|
test 2
|
|
|
|
fab
|
2021-08-02 04:42:43
|
new release
|
|
2021-08-02 04:42:44
|
hkk
|
|
|
Scientia
|
2021-08-02 07:29:17
|
is there any plan to allow recompression of existing jxls from cjxl?
|
|
2021-08-02 07:30:02
|
some compressed losslessly with older jxl versions can be losslessly converted to ones encoded with the newer builds and could have some size savings due to improvements.
|
|
|
|
monad
|
2021-08-02 07:39:28
|
A jxltran would make sense.
|
|
|
|
fab
|
2021-08-02 08:21:10
|
scientia we need more of your comment
|
|
2021-08-03 08:52:23
|
JPEG XL vs Guetzli Does libjxl 0.5.0 do well on screenshots? is the file size different from eclipse build? are there image cases or type of faces like old romans or neurdiv that need delta, is right to use near lossless for flowers and green lake at -s 2 -N 3? do you can do near lossless 3, how many are the near lossless options? Is only speed 8 possible like when i first tried libjxl and it wasn't even out? Or even speed 9 and speed 2? What is the cmd for delta today? Comparison of file sizes done by users (evn without posting images)?
|
|
2021-08-03 08:53:01
|
change the name
|
|
|
|
Cool Doggo
|
2021-08-03 02:22:06
|
does anyone know why on this image if i use ``-d 0 -e 3`` i get 102057 bytes, but if i use ``-d 0 -e 7`` i get 172610 bytes?
|
|
2021-08-03 02:23:00
|
it seems like -e 5-9 all do much worse than even -e 1 on this image
|
|
|
improver
|
2021-08-03 02:27:45
|
try `--patches=0`
|
|
|
|
Cool Doggo
|
2021-08-03 02:29:21
|
still the same
|
|
|
|
Scope
|
2021-08-03 02:30:50
|
Because lossless mode is not always optimal at the moment, adding something like this might help: `-E 3 -g 0` or `-g 0`
|
|
|
|
Cool Doggo
|
2021-08-03 02:37:08
|
still weird at -e 5 or -e 6 but that does help 👍
|
|
2021-08-03 02:37:42
|
its also not as much of a difference as without
|
|
|
|
_wb_
|
2021-08-03 03:00:25
|
Likely it does some transform (e.g. local palette) at slower speed settings that happens to be bad for compression on this image.
|
|
|
|
raysar
|
2021-08-03 05:03:56
|
I compile (static build) the last main build (a812d5c3) for windows user: https://1drv.ms/u/s!Aui4LBt66-MmnQhG4Ujd_pKnWjLU?e=74CSDn
|
|
2021-08-03 05:36:38
|
There is always an problem with apng in cjxl. I detect when the problem occurs:
Works with commit 8f81c4c 26/06/2021
And did not works with commit a64f576 09/07/2021 and always not works in 0.5
It's not problem with compilation, it's the same from docker for windows.
I see there is also an post about debian compilation, i don't know if it's linked https://github.com/libjxl/libjxl/issues/392
|
|
|
|
spider-mario
|
2021-08-03 05:53:03
|
is it a specific apng file that triggers the problem? what happens?
|
|
|
|
_wb_
|
2021-08-03 06:29:15
|
The apng reader fails on certain dispose modes, iirc
|
|
|
|
testerrrrr
|
2021-08-03 07:07:21
|
can I ask why there is an mtf transform before ans encoding?
|
|
2021-08-03 07:14:22
|
isnt ans already optimal
|
|
|
|
raysar
|
|
spider-mario
is it a specific apng file that triggers the problem? what happens?
|
|
2021-08-03 07:18:30
|
All apng fail to read for me.
|
|
|
|
spider-mario
|
2021-08-03 07:18:46
|
that indeed doesn’t sound normal, let me see if I can bisect this
|
|
|
raysar
There is always an problem with apng in cjxl. I detect when the problem occurs:
Works with commit 8f81c4c 26/06/2021
And did not works with commit a64f576 09/07/2021 and always not works in 0.5
It's not problem with compilation, it's the same from docker for windows.
I see there is also an post about debian compilation, i don't know if it's linked https://github.com/libjxl/libjxl/issues/392
|
|
2021-08-03 07:25:28
|
hm, I don’t seem to have a not-dispose-to-0 apng at hand, would you mind sending an apng file that works with 8f81c4c?
|
|
|
|
_wb_
|
|
testerrrrr
can I ask why there is an mtf transform before ans encoding?
|
|
2021-08-03 07:31:14
|
iirc that's about the context map, not ans itself, right <@179701849576833024> ?
|
|
|
|
raysar
|
|
spider-mario
hm, I don’t seem to have a not-dispose-to-0 apng at hand, would you mind sending an apng file that works with 8f81c4c?
|
|
2021-08-03 07:37:21
|
It works on 8f81c4c
|
|
2021-08-03 07:37:41
|
it fail on 8f81c4c
|
|
|
|
spider-mario
|
2021-08-03 07:40:49
|
thanks, apparently manga.apng uses dispose-to-0 which is why it doesn’t work, bisecting with wolf now
|
|
|
|
veluca
|
|
_wb_
iirc that's about the context map, not ans itself, right <@179701849576833024> ?
|
|
2021-08-03 07:41:44
|
it is about the context map, yes
|
|
2021-08-03 07:42:19
|
but ANS is only optimal for order-0 entropy coding, MTF allows exploiting some high-order correlations that you normally couldn't exploit 🙂
|
|
|
|
raysar
|
2021-08-03 07:42:32
|
And wolf.apng does not works on all recent build.
|
|
|
|
spider-mario
|
2021-08-03 07:43:33
|
curiously, the commit that broke it is “Get rid of fmemopen in codec_apng.” (131592184)
|
|
2021-08-03 07:44:44
|
https://github.com/libjxl/libjxl/pull/300
|
|
2021-08-03 07:49:39
|
I don’t see anything obviously wrong with it
|
|
2021-08-03 07:50:39
|
never mind, found it
|
|
2021-08-03 07:52:36
|
https://github.com/libjxl/libjxl/pull/404
|
|
|
|
raysar
|
2021-08-03 07:53:13
|
well done! 👏
|
|
|
|
_wb_
|
2021-08-03 07:59:17
|
https://c.tenor.com/VBisZgcKBhkAAAAM/homero-simpson.gif
https://c.tenor.com/muj4RR_sPl0AAAAM/homer-simpson-les-simpson.gif
https://c.tenor.com/0RremLvtrSIAAAAM/homer-simpson-doh.gif
https://c.tenor.com/_XC7t_3I-U4AAAAM/homer-simpson-doh.gif
|
|
|
|
fab
|
2021-08-03 08:06:54
|
the file size is different update
|
|
2021-08-03 08:07:05
|
less artificially fat you will look
|
|
2021-08-03 08:07:32
|
https://discord.com/channels/794206087879852103/794206087879852106/872196576087523358
|
|
2021-08-03 08:07:52
|
i don't think this setting is bad, it improves the legibility of twitter images
|
|
2021-08-03 08:08:01
|
but no more than 30% compression
|
|
2021-08-03 08:08:15
|
there are more images when compression can be great
|
|
2021-08-03 08:09:32
|
but for sony wx 60 photos that settings is bad
|
|
2021-08-03 08:10:07
|
|
|
2021-08-03 08:10:15
|
is a jxl not sure i can upload in benchmarks in info there isn't written you can upload jxls
|
|
2021-08-03 08:10:41
|
the speed is 5,6 mpx
|
|
2021-08-03 08:10:48
|
19,1 mpx reading jp
|
|
2021-08-03 08:11:02
|
54,2 mpx reading png
|
|
2021-08-03 08:11:04
|
|
|
2021-08-03 08:12:16
|
animated in thread when you switch the thread
|
|
2021-08-03 08:12:20
|
second from first
|
|
|
|
raysar
|
2021-08-03 08:26:39
|
i compile it and it works like 8f81c4c 😄
|
|
|
|
Jyrki Alakuijala
|
|
veluca
but ANS is only optimal for order-0 entropy coding, MTF allows exploiting some high-order correlations that you normally couldn't exploit 🙂
|
|
2021-08-03 11:37:31
|
in brotli I decided to use move-to-front for block ids -- I considered that there is a proximity effect there
|
|
2021-08-03 11:42:54
|
https://datatracker.ietf.org/doc/html/rfc7932#section-7.3 'encoding the context map'
|
|
2021-08-04 12:10:26
|
cessu's blog is pretty cool -- he also built SR3C -- http://cessu.blogspot.com/
|
|
2021-08-04 12:12:16
|
by coincidence (before either of us built any compression stuff) he is also the godfather of my daughter
|
|
2021-08-04 12:15:48
|
I like the blog post title "At 8am Helsinki is a 1.77-Dimensional"
|
|
|
Maiki3
|
2021-08-04 02:17:02
|
I notice that JXL produces banding artifacts in this image
|
|
2021-08-04 02:17:19
|
Original:
|
|
2021-08-04 02:18:29
|
JXL decoded to PNG - settings used: distance 1, speed 7
|
|
2021-08-04 02:19:07
|
the banding is subtle, but this is just a fraction of the entire image. when zoomed out to view the full image, the banding becomes very noticeable
|
|
2021-08-04 02:20:30
|
are there any improvements that can be made to JXL to minimize banding issues?
|
|
|
|
_wb_
|
2021-08-04 02:58:36
|
Decoding to 16-bit and dithering helps
|
|
2021-08-04 02:59:12
|
I think we should dithering by default when decoding to 8-bit buffers
|
|
|
|
fab
|
2021-08-04 03:03:03
|
why the cjxl crashes when there is "at" written in a image?
|
|
|
|
BlueSwordM
|
2021-08-04 03:03:07
|
I get less banding on my end encoding myself.
The banding is still preset, but quite a bit smoother.
|
|
|
|
fab
|
2021-08-04 03:03:22
|
is because i have cmd on windows 7?
|
|
2021-08-04 03:03:38
|
or is because the devs can't do their job?
|
|
2021-08-04 03:03:52
|
i got 523 images of 1375
|
|
2021-08-04 03:03:58
|
208 of 330
|
|
2021-08-04 03:04:08
|
and i don't know how to delete duplicate
|
|
2021-08-04 03:04:20
|
easy duplicate finder totally sucks as a program.
|
|
2021-08-04 03:04:24
|
don't know what to do
|
|
2021-08-04 03:04:39
|
i installed dupeguru don't work
|
|
|
BlueSwordM
I get less banding on my end encoding myself.
The banding is still preset, but quite a bit smoother.
|
|
2021-08-04 03:05:25
|
what an horrible image
|
|
2021-08-04 03:05:57
|
i won't compress p... with this
|
|
2021-08-04 03:06:05
|
if i don't know i can stay safe
|
|
2021-08-04 03:06:22
|
the command for -s 2 -d 1.238 was right
|
|
2021-08-04 03:07:05
|
ruined by jyrki criticism, jon that said there is a special heuristic that don't work because of this, github/gitlab openings, zeimek pessimism
|
|
2021-08-04 03:07:22
|
and average low bitrate 60 kb 3 mpx image not looking good
|
|
2021-08-04 03:07:34
|
the heuristic change in continuation
|
|
2021-08-04 03:08:08
|
windows 11 GPU AND MOTHERBOARD get licensed next month
|
|
2021-08-04 03:08:12
|
SEPTEMBER
|
|
2021-08-04 03:08:53
|
sorry
|
|
|
|
diskorduser
|
|
Maiki3
JXL decoded to PNG - settings used: distance 1, speed 7
|
|
2021-08-04 03:09:32
|
Could you check color profile of the image? I too had a image which had banding artifacts. Converting it to sRGB-v4 fixed it.
|
|
|
|
Jyrki Alakuijala
|
|
Maiki3
JXL decoded to PNG - settings used: distance 1, speed 7
|
|
2021-08-04 03:09:57
|
That is terrible. Try upping the --intensity_target and see if it goes away?
|
|
|
|
fab
|
|
Jyrki Alakuijala
That is terrible. Try upping the --intensity_target and see if it goes away?
|
|
2021-08-04 03:10:32
|
can you answer to my question? is a bug i have with windows 7
|
|
2021-08-04 03:10:33
|
https://discord.com/channels/794206087879852103/794206170445119489/872494854972772442
|
|
2021-08-04 03:10:38
|
sorry for disappointing you
|
|
|
|
diskorduser
|
|
fab
why the cjxl crashes when there is "at" written in a image?
|
|
2021-08-04 03:11:09
|
Send us that image
|
|
|
|
fab
|
2021-08-04 03:11:21
|
it makes a beep sound
|
|
2021-08-04 03:11:27
|
is expected on windows 7?
|
|
|
|
Jyrki Alakuijala
|
2021-08-04 03:11:32
|
I don't know how to write "at" into an image
|
|
|
|
fab
|
2021-08-04 03:11:39
|
cmd windows 7 is a problem?
|
|
|
|
diskorduser
|
2021-08-04 03:11:47
|
Yeah probably
|
|
|
|
fab
|
2021-08-04 03:12:03
|
so linux if you send an app and tells to do at
|
|
2021-08-04 03:12:07
|
it does
|
|
2021-08-04 03:12:14
|
is only win 7 problem
|
|
2021-08-04 03:12:17
|
and older
|
|
2021-08-04 03:12:25
|
i probably can send an image
|
|
|
|
Jyrki Alakuijala
|
2021-08-04 03:12:29
|
I haven't use Windows in the last 15 years, don't know what is "at"
|
|
|
|
fab
|
2021-08-04 03:12:32
|
but it will be waste of time
|
|
|
Jyrki Alakuijala
I haven't use Windows in the last 15 years, don't know what is "at"
|
|
2021-08-04 03:12:45
|
file that is named at screenshot a 12:30
|
|
|
|
Deleted User
|
2021-08-04 03:12:50
|
<@!532010383041363969> at = @
|
|
|
|
fab
|
|
Jyrki Alakuijala
I haven't use Windows in the last 15 years, don't know what is "at"
|
|
2021-08-04 03:13:18
|
i don't want to send an image especially here because i now is a waste of time for you
|
|
2021-08-04 03:13:23
|
you have many things to do
|
|
2021-08-04 03:13:24
|
in a month
|
|
|
|
Jyrki Alakuijala
|
2021-08-04 03:13:26
|
filename is @.jxl ?
|
|
|
|
fab
|
2021-08-04 03:13:31
|
no .png
|
|
|
|
diskorduser
|
2021-08-04 03:14:03
|
@.png ?
|
|
|
|
fab
|
2021-08-04 03:14:06
|
should i try batch
|
|
2021-08-04 03:14:13
|
ok i try
|
|
|
|
Jyrki Alakuijala
|
2021-08-04 03:14:19
|
if it crashes, file a bug
|
|
|
|
fab
|
|
|
Jyrki Alakuijala
|
2021-08-04 03:14:42
|
while we are fixing it, don't use "@"
|
|
|
|
fab
|
|
Jyrki Alakuijala
while we are fixing it, don't use "@"
|
|
2021-08-04 03:14:57
|
ah no i want older build
|
|
2021-08-04 03:15:06
|
but i'D FILE SAME
|
|
2021-08-04 03:19:53
|
i'll try
|
|
2021-08-04 03:19:59
|
let's restart computer
|
|
2021-08-04 03:37:23
|
In my opinion in 5 improvements after 0.5.0 will become good at q 70.08 s 9 epf 1 faster decoding 1
|
|
2021-08-04 03:38:23
|
Obviously jyrki will Not release that veesion
|
|
2021-08-04 03:38:36
|
The focus for jpeg xl is speed 7
|
|
2021-08-04 03:46:53
|
Why you don't make an interprediction jxl
|
|
2021-08-04 03:47:06
|
with this quality it will be good
|
|
|
|
Jyrki Alakuijala
|
2021-08-04 04:02:34
|
how could I make interprediction?
|
|
2021-08-04 04:35:10
|
you mean directional interprediction?
|
|
2021-08-04 04:36:04
|
we tried three different attempts of that with different engineers -- all were too slow to encode for, and they did not produce density gains in the quality target we were building jpeg xl for (d4 and better)
|
|
|
|
fab
|
2021-08-04 04:38:51
|
I have an idea
|
|
2021-08-04 04:39:44
|
|
|
2021-08-04 04:40:29
|
JXL still is rather weak on recompressing this
|
|
2021-08-04 04:40:39
|
Is a bug or a feature
|
|
2021-08-04 04:41:04
|
Is recompressing lossy especially webp ugly and discouragavle
|
|
2021-08-04 04:41:19
|
What progress can you make
|
|
2021-08-04 04:42:24
|
I will send this thumbnail in discord tomorrow i will look at Windows 7
|
|
|
|
Jyrki Alakuijala
|
2021-08-04 04:45:58
|
JXL should be better at that than JPEG, possibly the same or slightly better than WebP, and likely worse than AVIF
|
|
|
|
fab
|
2021-08-04 04:46:18
|
at this image by macaluso giampiero?
|
|
|
|
Jyrki Alakuijala
|
2021-08-04 04:46:20
|
I consider AVIF achieves that through palettes and prediction, but I don't know for sure
|
|
2021-08-04 04:46:36
|
we can replicate palette behaviour in additional layers
|
|
|
|
fab
|
2021-08-04 04:46:37
|
that type of image is a problem for jpeg xl
|
|
2021-08-04 04:46:49
|
because it tries butteraugli precision
|
|
|
|
Jyrki Alakuijala
|
2021-08-04 04:46:52
|
but this is not a huge priority
|
|
2021-08-04 04:47:06
|
the main focus is photographs
|
|
2021-08-04 04:47:18
|
jPeg, P is for photograph 😛
|
|
|
|
fab
|
2021-08-04 04:47:21
|
yes in my opinion from 0.5.0 make 5 improvements
|
|
2021-08-04 04:47:28
|
will become good at q 70.08 s 9 epf 1 faster decoding 1
|
|
2021-08-04 04:47:33
|
after you focus at this
|
|
2021-08-04 04:47:40
|
then the encoder will be good at speed 7
|
|
2021-08-04 04:47:51
|
then at marketing
|
|
|
|
veluca
|
2021-08-04 04:55:34
|
IIRC @ is not allowed in file names in Windows
|
|
|
|
Scope
|
2021-08-04 04:57:53
|
```cjxl.exe "@.png" "@.jxl" -s 8 -d 1 --num_threads 8 -v -v
JPEG XL encoder v0.5.0 [AVX2]
Read 2003x2003 image, 34.5 MP/s
Encoding [VarDCT, d1.000, kitten], 8 threads.
Compressed to 323364 bytes (0.645 bpp).
2003 x 2003, 0.64 MP/s [0.64, 0.64], 1 reps, 8 threads.
Average butteraugli iters: 3.00
Total layer bits headers 0.002745% 71
Total layer bits TOC 0.037967% 982
Total layer bits quantizer 0.000619% 16
Total layer bits quant tables 0.000077% 2
Total layer bits order 0.118540% 3066 [c/i: 5.00 | hst: 28 | ex: 52 | h+c+e: 392.183]
Total layer bits DC 17.128391% 443020
Total layer bits ControlFields 4.140321% 107088
Total layer bits AC 0.912829% 23610 [c/i: 35.00 | hst: 2923 | ex: 0 | h+c+e: 239976.704]
Total layer bits ACTokens 77.449385% 2003202
Total layer bits modularGlobal 0.169807% 4392 [c/i: 41.00 | hst: 548 | ex: 0 | h+c+e: 61076.625]
Total layer bits modularTree 0.039320% 1017 [c/i: 4.00 | hst: 19 | ex: 25 | h+c+e: 133.026]
Total image size 2586466 [c/i: 85.00 | hst: 3519 | ex: 20882 | h+c+e: 322382.913]
Allocations: 1729 (max bytes in use: 1.056327E+09)```
```djxl.exe "@.jxl" "@.jxl.png"
JPEG XL decoder v0.5.0 [AVX2]
Read 323364 compressed bytes.
Decoded to pixels.
2003 x 2003, 40.17 MP/s [40.17, 40.17], 1 reps, 4 threads.
Allocations: 349 (max bytes in use: 1.126603E+08)```
|
|
|
|
Deleted User
|
|
veluca
IIRC @ is not allowed in file names in Windows
|
|
2021-08-04 04:58:27
|
Only `\/:*?"<>|` are forbidden.
|
|
|
|
veluca
|
2021-08-04 05:12:14
|
Ah, I thought it did something weird
|
|
|
|
fab
|
2021-08-04 05:28:37
|
No at
|
|
2021-08-04 05:28:41
|
At.jxl
|
|
2021-08-04 05:28:59
|
screenshot at 18:49.jxl
|
|
2021-08-04 05:29:04
|
Cjxl Windows 7
|
|
2021-08-04 05:29:13
|
The problem is Windows 7 cmd
|
|
2021-08-04 05:30:02
|
The file is a png and is already named that
|
|
|
|
spider-mario
|
2021-08-04 05:34:53
|
yeah, there are a few more names one can’t (or shouldn’t) use on Windows, such as CON or NUL
|
|
2021-08-04 05:34:55
|
https://stackoverflow.com/q/17883481
|
|
|
|
fab
|
2021-08-04 06:36:18
|
-s 7 -d 1.24 --epf=0 --use_new_heuristics
|
|
2021-08-04 06:37:06
|
probably that gives a bit better quality
|
|
2021-08-04 06:37:21
|
with a newer build and new heuristic
|
|
2021-08-04 06:37:27
|
shame i can't compile
|
|
|
|
diskorduser
|
|
Maiki3
Original:
|
|
2021-08-05 02:55:44
|
There is banding in source image. Source is not lossless
|
|
2021-08-05 03:00:54
|
Are you the artist? The maximum zoomed out resolution from the artist at blender artist org is 3000x2000px. Also it's just a jpg.
|
|
|
|
fab
|
2021-08-05 01:40:13
|
why not add modular and vardct tag? can be done without increasing header size?
|
|
2021-08-05 01:40:19
|
i have looked your comments
|
|
2021-08-05 01:40:28
|
what JPEG has said on that?
|
|
2021-08-05 01:41:57
|
COMMENT ON 4CHAN
|
|
2021-08-05 01:41:58
|
i think the usefulness of picking between things like mp4/m4v/m4a, which all refer to the same container, is how they signify the basic purpose of the file
mp4 suggests strongly that it is multi-stream, especially video and audio
m4v suggests that it could be video-only
m4a suggests strongly that it's audio-only
you can separate music from video clips quite easily if they use these extensions
lossy vs. lossless extensions i don't think has much practical use, after all, most people don't care if a still image is jpeg or png, either way it's a still image
separating still from animated is more useful
|
|
2021-08-05 01:42:00
|
.....
|
|
2021-08-05 01:42:11
|
i think this could confuse users on windows
|
|
2021-08-05 01:42:31
|
i know more than 99 of users don't know if is lossless or lossy and this can't be known
|
|
2021-08-05 01:42:43
|
why not add modular and vardct tag? can be done without increasing header size?
i have looked your comments
what JPEG has said on that?
|
|
2021-08-05 01:43:17
|
Windows Explorer's "type" column is just a full type name based on the file extension. Windows as a platform in general tries to be dumb regarding file types, only acting on their names.
|
|
2021-08-05 01:43:48
|
also only 168 stars
|
|
2021-08-05 01:43:52
|
is jxl failing?
|
|
2021-08-05 01:44:58
|
....
|
|
2021-08-05 01:45:22
|
a 4chan user offended one of reddit and said it's autistic
|
|
|
|
_wb_
|
|
fab
why not add modular and vardct tag? can be done without increasing header size?
i have looked your comments
what JPEG has said on that?
|
|
2021-08-05 01:50:21
|
We don't standardize filename extension conventions, everyone can do whatever they want with that
|
|
|
|
fab
|
|
_wb_
We don't standardize filename extension conventions, everyone can do whatever they want with that
|
|
2021-08-05 01:50:41
|
no, in windows exploreer
|
|
2021-08-05 01:50:58
|
can microsoft windows 11 explorer recognize in proprieters that is modular or vardct and indicate that
|
|
2021-08-05 01:51:20
|
i know that propierters hasn't never been so explicit on information
|
|
|
_wb_
We don't standardize filename extension conventions, everyone can do whatever they want with that
|
|
2021-08-05 01:51:51
|
the answer is this? even to the second question?
|
|
|
|
_wb_
|
2021-08-05 01:52:30
|
The sensible thing to do is to ignore extensions and look at file headers to determine the file type, not believe extensions.
|
|
|
Fox Wizard
|
2021-08-05 01:53:23
|
Gotta love people who rename .jpg to .png :p
|
|
|
|
_wb_
|
2021-08-05 01:54:02
|
Modular or vardct is part of the frame header, but I think there are more relevant things to report about a jxl image, like its dimensions, bit depth, animated or not.
|
|
|
Fox Wizard
Gotta love people who rename .jpg to .png :p
|
|
2021-08-05 01:54:19
|
That's much better than converting jpg to png
|
|
|
|
Fox Wizard
|
2021-08-05 01:54:37
|
Basically half the "png" images on Discord <:sadge:855476349686513714>
|
|
2021-08-05 01:54:50
|
People tend to copy/paste images and then they get converted to png :/
|
|
|
|
fab
|
|
_wb_
Modular or vardct is part of the frame header, but I think there are more relevant things to report about a jxl image, like its dimensions, bit depth, animated or not.
|
|
2021-08-05 01:55:15
|
so who decide this?
|
|
2021-08-05 01:55:26
|
microsoft? wb? other people?
|
|
2021-08-05 01:55:38
|
i read somewhere that devs are strange to that
|
|
2021-08-05 01:55:47
|
they should only do the spec
|
|
2021-08-05 01:55:50
|
and stop
|
|
2021-08-05 01:55:56
|
is right?
|
|
2021-08-05 01:56:02
|
what is the work of the dev?
|
|
|
|
_wb_
|
2021-08-05 01:56:06
|
The author of a file navigator (plugin) decides what and how to present the file info
|
|
|
|
fab
|
|
_wb_
The author of a file navigator (plugin) decides what and how to present the file info
|
|
2021-08-05 01:56:26
|
yes i was disappointed by photo qt because of that
|
|
2021-08-05 01:56:32
|
only octect stream
|
|
2021-08-05 01:56:40
|
no vardct modular options
|
|
|
|
diskorduser
|
|
fab
can microsoft windows 11 explorer recognize in proprieters that is modular or vardct and indicate that
|
|
2021-08-05 02:03:01
|
Windows explorer doesn't even show file extensions by default
|
|
|
|
Maiki3
|
|
diskorduser
Are you the artist? The maximum zoomed out resolution from the artist at blender artist org is 3000x2000px. Also it's just a jpg.
|
|
2021-08-05 02:23:45
|
Indeed, source is not lossless. What I did is, I denoised the image, then I brought it into photoshop... set image to 16bit. applied a gaussian blur on the background. then set image to 8bit, resulting in dithering. then saved the image as a PNG and converted it to JpegXL.
|
|
2021-08-05 02:24:18
|
So basically, the photoshop dithering is about the best possible result you can achieve with an 8bit image and minimal banding
|
|
2021-08-05 02:24:25
|
But all of that is irrelevant
|
|
2021-08-05 02:24:51
|
the only thing that is relevant to me, is that I started with an image containing no obvious banding artifacts, converted to JXL using pretty much default settings, and the result has banding artifacts.
|
|
|
diskorduser
Are you the artist? The maximum zoomed out resolution from the artist at blender artist org is 3000x2000px. Also it's just a jpg.
|
|
2021-08-05 02:26:33
|
No. I am not original artist. Yes, I am aware original image is larger. I cropped it and saved it as a PNG so that I could upload them directly to discord because I do not pay for discord nitro or whatever they call it
|
|
2021-08-05 02:28:33
|
also, this is not the first image i've converted to JXL and noticed banding artifacts. it's apparent on certain images with skies. it would be nice to develop a solution to this.
|
|
|
diskorduser
Could you check color profile of the image? I too had a image which had banding artifacts. Converting it to sRGB-v4 fixed it.
|
|
2021-08-05 02:29:34
|
pretty sure color profile is just sRGB.
|
|
2021-08-05 02:29:49
|
I don't know what sRGB-v4 is
|
|
|
|
Deleted User
|
2021-08-05 02:32:11
|
Banding in JXL happens because it doesn't preserve dithering. Either add (photon) noise or decode to 10+ bit.
|
|
|
|
Maiki3
|
2021-08-05 02:32:51
|
photon noise? is that a feature in JXL that I can enable?
|
|
2021-08-05 02:33:38
|
i see it in your username now lol
|
|
|
|
Deleted User
|
2021-08-05 02:33:50
|
It's a parameter. Use -h -v -v -v for more details.
|
|
|
|
Maiki3
|
2021-08-05 02:34:28
|
off topic - any reason you prefer -q instead of -d?
|
|
|
|
Deleted User
|
2021-08-05 02:37:21
|
There is no difference between them if you convert the units correctly. But -q XX can mean -d Y.YYY now while older versions meant -d Z.ZZZ.
|
|
|
|
diskorduser
|
|
Maiki3
pretty sure color profile is just sRGB.
|
|
2021-08-05 02:42:27
|
I had similar experience when using sRGB 61966. converting source to sRGB v4 fixed banding.
|
|
2021-08-05 02:50:43
|
I can see banding in source image. So, the problem is not related to icc.
|
|
|
|
veluca
|
|
Maiki3
Indeed, source is not lossless. What I did is, I denoised the image, then I brought it into photoshop... set image to 16bit. applied a gaussian blur on the background. then set image to 8bit, resulting in dithering. then saved the image as a PNG and converted it to JpegXL.
|
|
2021-08-05 03:15:07
|
is the PNG a 16-bit one? what happens if you ask photoshop to convert it to 8-bit using dithering?
|
|
|
|
fab
|
2021-08-05 04:51:44
|
branch-main-a812d5c3
For Windows user, there is the last build 0.5 03/08/2021 with all encoder optimisation:
(download cjxl.exe) https://1drv.ms/u/s!Aui4LBt66-MmnQhG4Ujd_pKnWjLU
|
|
2021-08-05 04:52:01
|
stealed message
|
|
|
|
Maiki3
|
|
veluca
is the PNG a 16-bit one? what happens if you ask photoshop to convert it to 8-bit using dithering?
|
|
2021-08-05 08:03:36
|
That's what I did.
|
|
|
|
veluca
|
2021-08-05 08:04:10
|
I mean the PNG decoded by JXL 🙂
|
|
|
|
Maiki3
|
|
diskorduser
I can see banding in source image. So, the problem is not related to icc.
|
|
2021-08-05 08:04:51
|
The source is an 8bit image that I applied a gaussian blur to with 16bit, and then downconverted it to 8bit with dithering and then saved as an 8bit PNG. There should be no perceptible banding, and if there is, then it's microscopically tiny and can't be avoided.
|
|
|
|
_wb_
|
2021-08-05 08:05:17
|
`djxl --bits_per_sample 16` might give better results than default djxl
|
|
|
|
Maiki3
|
2021-08-05 08:05:30
|
I will give it a try. 🙂
|
|
|
|
veluca
|
2021-08-05 08:05:57
|
btw you might also want to encode the 16-bit png directly
|
|
|
|
Maiki3
|
2021-08-05 08:06:07
|
well, i definitely understand that
|
|
|
|
_wb_
|
2021-08-05 08:06:24
|
I think we really should do some simple dithering in libjxl whenever we're producing 8-bit output
|
|
|
|
Maiki3
|
2021-08-05 08:06:30
|
I think i've tried once before.... but without luck. I might have to do what _wb_ said and manually specify the bits per sample
|
|
|
_wb_
I think we really should do some simple dithering in libjxl whenever we're producing 8-bit output
|
|
2021-08-05 08:07:04
|
Yeah, I agree. Anything to reduce banding with default settings is good, imo.
|
|
2021-08-05 08:07:31
|
my goal here is to point out, that jpegxl with default settings can use some improvement in the banding department
|
|
2021-08-05 08:07:38
|
seems pretty good in most other places
|
|
2021-08-05 08:08:52
|
but also, regular jpeg and even mozjpeg sucks when it comes to banding artifacts as well, because with them, you just get blocking artifacts instead....
|
|
2021-08-05 08:09:09
|
so at least JXL improves there lol
|
|
|
|
veluca
|
2021-08-05 08:09:23
|
given the color I wouldn't be surprised if it's just one of those colors where 8-bit sRGB is just not enough
|
|
2021-08-05 08:09:47
|
I had a branch somewhere that adds ordered dithering at decode time, which might help, but we never did anything with it
|
|
|
|
Maiki3
|
2021-08-05 08:10:12
|
seems like it could be handy dandy
|
|
|
diskorduser
I had similar experience when using sRGB 61966. converting source to sRGB v4 fixed banding.
|
|
2021-08-05 08:11:45
|
Is this an option within photoshop? i've never heard of sRGB-v4
|
|
|
|
fab
|
|
veluca
given the color I wouldn't be surprised if it's just one of those colors where 8-bit sRGB is just not enough
|
|
2021-08-05 08:12:08
|
encode at 32 bit
|
|
2021-08-05 08:12:10
|
default
|
|
2021-08-05 08:12:20
|
maiki
|
|
|
|
Maiki3
|
|
|
fab
|
2021-08-05 08:12:44
|
you should specify s 7 d 1 and encode
|
|
2021-08-05 08:12:47
|
or s 9 d 1
|
|
2021-08-05 08:12:56
|
do not add bits
|
|
|
|
Maiki3
|
2021-08-05 08:13:16
|
thank you for your help
|
|
2021-08-05 08:13:23
|
I will try some of these ideas later tonight
|
|
|
|
fab
|
2021-08-05 08:13:58
|
i do not specify anything
|
|
2021-08-05 08:14:06
|
if you force color to 8 bit
|
|
2021-08-05 08:14:19
|
the floats will have less space to work with
|
|
2021-08-05 08:14:29
|
do not force colors
|
|
2021-08-05 08:15:06
|
jxl is designed to operate without option you specify a speed a butteraugli distance
|
|
2021-08-05 08:15:23
|
and the modular encodes at 32 bit and then uses VARDCT
|
|
2021-08-05 08:15:45
|
settings are s 7 d1 s 9 d 1
|
|
2021-08-05 08:15:51
|
q 90 same as d 1
|
|
|
|
diskorduser
|
|
Maiki3
Is this an option within photoshop? i've never heard of sRGB-v4
|
|
2021-08-06 04:13:24
|
I got it from color org website.
|
|
|
|
w
|
2021-08-06 04:27:56
|
does srgb v4 not just mean srgb in icc v4
|
|
|
|
diskorduser
|
2021-08-06 04:33:01
|
I don't know much about it. It just fixed my banding problem. That's all I know.
|
|
|
|
Maiki3
|
2021-08-06 05:12:13
|
|
|
2021-08-06 05:12:21
|
Well, my TIF doesn't wanna convert to JXL
|
|
2021-08-06 05:12:32
|
32 bit with LZA compression lol
|
|
2021-08-06 05:15:22
|
hmm jxl doesn't wanna convert this tif. i saved it as a PBM file
|
|
2021-08-06 05:15:36
|
pbm converts but it turns out way wrong lol
|
|
|
|
fab
|
2021-08-06 05:47:22
|
just convert to png 16 bit
|
|
2021-08-06 05:47:28
|
you don't need more than that
|
|
|
|
_wb_
|
2021-08-06 06:09:08
|
Pbm is 1-bit
|
|
|
|
Maiki3
|
2021-08-06 06:09:28
|
lol
|
|
2021-08-06 06:09:46
|
That definitely explains why it turned out so bad lol
|
|
|
|
_wb_
|
2021-08-06 06:09:46
|
Try pfm
|
|
|
|
Maiki3
|
2021-08-06 06:31:56
|
8bit version with dithering, with photon noise:
|
|
2021-08-06 06:32:03
|
That seems to be the best one so far
|
|
2021-08-06 06:32:25
|
16bit version with photon noise:
|
|
2021-08-06 06:33:40
|
Photon noise really seems to help, but it seems like cheating.
|
|
2021-08-06 06:35:06
|
i can definitely still see banding artifacts in both of those images, but the result is much more pleasant to look at compared to without photon noise
|
|
|
|
diskorduser
|
2021-08-06 06:39:00
|
If we had access to blend file or png, it would be easier check codec performance
|
|
|
|
Maiki3
|
2021-08-06 05:31:08
|
<@!263309374775230465> Here is a 16bit PNG of it: https://mega.nz/file/l6ZBRQha#_0Bk_jY51uNKgIvYqIwVDfQxpqhebp96GjmSusImcVk
|
|
2021-08-06 05:36:35
|
Here's an 8bit Dithered PNG of it: https://mega.nz/file/EixDSAyC#9Ysq7w67NYw-qt47JCwNfvjVH3xJ5VR_J6EO-whxKbo
|
|
2021-08-06 05:57:14
|
Here's a PSD file if anybody wants that too: https://mega.nz/file/0j53CIKb#HQGE8llSnt0NcLemtUEEmHXVx1wwifYrN8CYeV3Zi9w
|
|
|
|
diskorduser
|
2021-08-06 06:21:05
|
By png I mean rendered out png from blender.
|
|
|
|
Maiki3
|
2021-08-06 06:26:41
|
I don't have that.
|
|
|
|
Cool Doggo
|
|
diskorduser
By png I mean rendered out png from blender.
|
|
2021-08-06 07:17:59
|
this is the original image (jpg not png) https://www.oscurart.com/CG/OSCURART_01.jpg
|
|
|
|
fab
|
2021-08-06 07:26:25
|
ah 136,2 mb
|
|
2021-08-06 07:26:34
|
you are probably trying high compression
|
|
|
|
Scope
|
2021-08-06 11:19:00
|
<:Hypers:808826266060193874>
https://github.com/libjxl/libjxl/pull/419
|
|
|
|
Maiki3
|
|
|
BlueSwordM
|
2021-08-07 05:57:10
|
So, it looks like that tuned aomenc still outperforms cjxl in a non-negligible number of artificial image scenarios(IE games), which is rather surprising, even after the enhancements that it's been receiving:
https://slow.pics/c/wbbmMSJ8
|
|
|
|
Maiki3
|
2021-08-07 06:13:46
|
the jxl looks sharper
|
|
2021-08-07 06:15:43
|
aomenc:
|
|
2021-08-07 06:16:09
|
jxl:
|
|
2021-08-07 06:16:42
|
<@!321486891079696385> what metric are you using to determine how one outperforms?
|
|
|
|
BlueSwordM
|
|
Maiki3
<@!321486891079696385> what metric are you using to determine how one outperforms?
|
|
2021-08-07 06:17:08
|
No metric, just using my eyes and comparing against the original.
|
|
|
|
Maiki3
|
2021-08-07 06:17:50
|
ah, i see. well, jxl is going to make spurious artifacts around edges, because it's similar to how jpeg behaves, but it's going to preserve details in dark areas
|
|
2021-08-07 06:18:07
|
and aomenc is going to preserve sharp lines better
|
|
|
|
BlueSwordM
|
2021-08-07 06:18:12
|
Yeah, but there are clear areas where aomenc is preserving details *better*
|
|
2021-08-07 06:18:31
|
Look at the rock just above Sonic's head.
|
|
|
|
Maiki3
|
2021-08-07 06:19:17
|
can you pinpoint it with a screenshot?
|
|
|
|
BlueSwordM
|
2021-08-07 06:22:32
|
The texture might look sharper, but it's just full of microblocking(like what've you said about JPEG artifacts)
|
|
2021-08-07 06:22:58
|
Of course, if the main element on the image could have higher quality(the character itself), then it would be a wash as to wish one looks better overall.
|
|
|
|
Maiki3
|
2021-08-07 06:23:15
|
that triangular rock?
|
|
|
|
BlueSwordM
|
|
Maiki3
that triangular rock?
|
|
2021-08-07 06:23:45
|
Yes.
|
|
2021-08-07 06:24:06
|
This is where it gets interesting here.
|
|