|
2022-01-19 02:48:03
|
Thank you for your reply 🙂
It's mean not recommend apply two different algorithm on single image?,
and need a heuristic to choose suitable algorithm for non-photo content?
(like av1 have DCT method and palette prediction method)
I understand lossy compress unlike lossless compress, always have distortion,
for now I don't need mathematically lossless,
so I try to find a high quality lossy method(like near-lossless),
less artifacts and preserve grain,noise better. |
|
2022-01-19 02:53:47
|
jxl lossy palette also a great lossy method,
but I think still on very experiment status. |
|
2022-01-21 01:35:00
|
For some non-photo picture I need increase intensity target to reduce artifacts(on cjxl d0.5 e8).
I hope jxl can reduce high contrast sharp edge area artifacts and implement auto palette feature. |
|
2022-01-22 04:30:32
|
Look like disable gaborish can reduce some artifacts,
but some area will get worse.
Increase intensity target can reduce some artifacts,
but higher intensity target will increase much file size.
> -j -d 0.5 -e 8 --epf=3 --strip --intensity_target=1500 --gaborish=0
> source picture: jpeg q100 444 drawing content(non-photo). |
|
2022-01-23 04:51:54
|
If source picture high contrast sharp edge is clear,
Is possible keep clear on vardct mode, I should enable or disable some parameter? |
|
2022-01-23 04:52:09
|
Original jpeg q100 444 cut to png rgb 8bit |
|
2022-01-23 04:52:26
|
jxl d0.5-e8-epf3 |
|
2022-01-23 04:52:34
|
av1 q7-s3-d12 |
|
2022-01-24 05:06:07
|
I just thinking if jxl vardct can get some improvement for non-photo content(drawing,comic,picture)
and have some auto palette feature to avoid DCT worst case, that will be very useful. |
|
2022-01-25 06:16:40
|
I test some complex(detail,noise) drawing sample,
I totally agree jxl vardct is better than av1 for complex area,
but I still think jxl vardct have too much artifacts for smooth or high contrast sharp edge area. |
|
2022-01-26 02:45:25
|
Look like -c 1 and --intensity_target can reduce artifacts, but will create another issue.
> -d 0.5 -e 8 --intensity_target=4000
> -d 0.5 -e 8 -c 1 |
|
2022-01-26 02:45:55
|
jxl d0.5-e8-epf3-IT4000 |
|
2022-01-26 02:46:09
|
jxl d0.5-e8-epf3-c1 |
|
2022-01-27 03:22:43
|
Look like separate haven't work on this sample?🤔
> djxl -s 8 |
|
2022-01-27 04:41:02
|
I found this article before, maybe this method can let separate mode better?
An efficient lossy cartoon image compression method
> https://link.springer.com/article/10.1007/s11042-019-08126-7 |
|
2022-01-30 04:30:05
|
Look like `-c 1` can remove those artifacts,
but will happen some strange artifacts. |
|
2022-02-01 07:57:05
|
I have a idea for sharp edge and line area,
Implement some edge detection algorithm jxl,
If detect value large than threshold(Identify non-photo) apply higher epf,
probably a good idea? |
|
2022-02-01 08:03:36
|
In my test, cjxl -d 0.5 -e 8(XYB) + slight bilateral filter can remove tiny artifacts,
but I think apply higher jxl epf probably a better choose? |
