jpg, png, webp encoder and butteraugli distance

[Lithium Flower]

sorry, my english is not good.

Hello, I compress a lot of non-photographic images, image type is Japanese Anime and Japanese Manga,
png rgb24 using mozjpeg jpeg lossy q95~q99, png rgba32 using cwebp webp near-lossless 60,80,
and pingo png lossy pngfilter=100, get some problem in cwebp webp lossy.

I using butteraugli check compressed image quality, but had some question with butteraugli distance,
i need some hint or suggest for those question, thanks you very much.

my image set like this image set, Tab Anime, AW, Manga, Pixiv,

http://docs.google.com/spreadsheets/...gid=2135584682
Thanks Scope provide this image set.
https://encode.su/threads/2274-ECT-a...ll=1#post64829

1. butteraugli and butteraugli jpeg xl assessment difference butteraugli distance

I using butteraugli and butteraugli xl to check image, in *Reference 01,

butteraugli's xyb is likely more accurate,

but some image butteraugli assessment good distance(1.3), butteraugli xl assessment bad distance(near 2.0),
and some image butteraugli reject butteraugli xl good distance(1.3),
how to correct understand butteraugli distance and butteraugli xl 3-norm?

2. butteraugli safe area or great area

Compress png rgba32 image, my process is using near-lossless 60 and pngfilter=100 to first compress,
if compressed image not below safe butteraugli distance, using near-lossless 80 to second compress.

I collect Jyrki Alakuijala Comment and create a table, in *Reference 02

'1.0 ~ 1.3 definitely works as designed', '1.0 ~ 1.6 A value below 1.6 is great',

if i want my compressed image have a great quality, i should choose area 1.0 ~ 1.3 or area 1.0 ~ 1.6?, if i made a mistake please let me know.

pngfilter=100 butteraugli distance => [0.6 ~ 1.0, 1.3 ~ 1.6], [1.7]
webp near-lossless 60 butteraugli distance => [0.4 ~ 1.0, 1.3 ~ 1.6], [2.1]

,near-lossless 60 [2.1] => near-lossless 80 [1.3]

cwebp & pingo_rc3 command:
pingo_rc3.exe -pngfilter=100 -noconversion -nosrgb -nodate -sa "%%A"
cwebp.exe -mt -m 6 -af -near_lossless 60 -alpha_filter best -progress "%%A" -o "%%~nA.webp"

3. non-photographic image and jpeg encoder quality suggest

Compress png rgb24 image, my process is using quality 95 to first compress,
if compressed image not below safe butteraugli distance, increase quality to second compress.

In my png rgb24 image set, butteraugli assessment jpeg quality 95 doesn't get good butteraugli distance,

jpeg quality 95 butteraugli distance => [1.5 ~ 1.6], [1.7 ~ 2.5]
jpeg quality 95 butteraugli xl distance => [1.2 ~ 1.6], [1.7 ~ 2.2]

but in cjpeg usage.txt :

'specifying a quality value above about 95 will increase the size of the compressed
file dramatically, and while the quality gain from these higher quality values is measurable'
https://github.com/mozilla/mozjpeg/b...usage.txt#L115

If i compress non-photographic image to jpeg and want near psychovisual lossless, it is necessary using
above quality 95 to compress those image?
or in *Reference 03, possibly butteraugli is too sensitive in some non-photographic image?

mozjpeg command:
cjpeg.exe -optimize -progressive -quality 95 -quant-table 3 -sample 1x1 -outfile "mozjpeg\%%~nA.jpg" "%%A"

Update 20200929:

I using jpeg xl sjpeg features, sjpeg can get great butteraugli distance in quality 96,
but look like sjpeg features doesn't using jpeg xl vardct or modular mode?

​Size:
png - 12mb
mozjpeg -q 97 + jpegtran progressive 4.46mb
jpegxl sjpeg -q 96 + jpegtran progressive 4.0mb

cjpegxl command:
cjpegxl.exe" "%%A" "xl\%%A" --jpeg1 --jpeg_quality=96 --progressive

4. webp lossy q100 and butteraugli distance

I test another non-photographic image set in webp lossy q100, but some image get larger butteraugli distance,
it possibly webp lossy 420 subsampling and fancy upsampling will make some larger errors in some area?

and i test webp lossy alpha(alpha_q) features, this features will increase butteraugli distance,
but i don't understand, why lossy alpha will effect butteraugli distance?

webp lossy q100 and lossy_alpha:

q100.png 2.013666
q100_lossy_alpha 80.png 2.035022
q100_lossy_alpha 50.png 2.099735

webp lossy q100 butteraugli distance => [1.2 ~ 1.6], [1.8 ~ 2.3], [3.1, 4.4, 5.5, 10.8]
dssim => [0.000150 ~ 0.000749]

cwebp command:
​cwebp.exe -mt -m 6 -q 100 -sharp_yuv -pre 4 -af -alpha_filter best -progress "%%A" -o "%%~nA.webp"

I creating some table and quality test data, i will upload later, thanks you very much.

================================================== ===============================================
Reference Area (From Jyrki Alakuijala Comment)

*Reference 01
From Jyrki Alakuijala Comment:
Butteraugli vs Butteraugli(jpeg xl)
butteraugli's xyb is likely more accurate,
because of asymptotic log behaviour for high intensity values (instead of raising to a power),

jpeg xl's xyb modeling is going to be substantially faster to computer,
because gamma is exactly three there.

*Reference 02
From Jyrki Alakuijala Comment:
0.6 ~ 0.7 // most critical use

1.0 // normal use

1.0 ~ 1.3 // definitely works as designed

1.0 ~ 1.6 // A value below 1.6 is great

1.6 ~ 2.1 // a value below 2.1 okayish

2.1+ // Above 2.1 there is likely a noticeable artefact in an inplace flip test.

2.5+ // not necessarily psychovisually relevant and fair

4.0+ /* The non-linearities near the just-noticeable-boundary in scale At larger errors (4.0+)
butteraugli becomes less useful. Current versions of butteraugli only extrapolate these values
as multiples of just-noticeable-difference, but the human visual system is highly non-linear and
large extrapolationdoesn't bring much value. */

https://github.com/google/butteraugli/issues/22

*Reference 03
From Jyrki Alakuijala Comment:
Butteraugli is a lot more sensitive for lines (displaced, emerging or removed) than any other visual measure,

[Jyrki Alakuijala]

1. butteraugli and butteraugli jpeg xl assessment difference butteraugli distance

I using butteraugli and butteraugli xl to check image, in *Reference 01,

but some image butteraugli assessment good distance(1.3), butteraugli xl assessment bad distance(near 2.0),
and some image butteraugli reject butteraugli xl good distance(1.3),
how to correct understand butteraugli distance and butteraugli xl 3-norm?

In the very highest quality area (q95 to q99 jpeg), you need to use a higher norm than 3-norm. You could use the maximum butteraugli score or something like the 12th norm.

If you compress around q60-q75, then the 3rd norm may be appropriate. Even there I have some signals that 6th norm is more appropriate.

PSNR as a visual error metric is MSE-based and is homologous to using 2nd norm, and that is too low of a norm for aggregating visual errors. Consider 3rd norm as a compromise between tradition and what actually works. Lower norms are also easier for machine learning approaches to optimize for. Also note that butteraugli creates a mix of X, 2*X and 4*X p-norms for the given X, i.e., the 3-norm for butteraugli is actually an equal mix of 3-norm, 6-norm and 12-norm. In your use you'd need likely 12-norm, 24-norm and 48-norm and you can get them by setting the --error_pnorm=12

2. butteraugli safe area or great area

Compress png rgba32 image, my process is using near-lossless 60 and pngfilter=100 to first compress,
if compressed image not below safe butteraugli distance, using near-lossless 80 to second compress.

I collect Jyrki Alakuijala Comment and create a table, in *Reference 02

if i want my compressed image have a great quality, i should choose area 1.0 ~ 1.3 or area 1.0 ~ 1.6?, if i made a mistake please let me know.

pngfilter=100 butteraugli distance => [0.6 ~ 1.0, 1.3 ~ 1.6], [1.7]
webp near-lossless 60 butteraugli distance => [0.4 ~ 1.0, 1.3 ~ 1.6], [2.1]

,near-lossless 60 [2.1] => near-lossless 80 [1.3]

cwebp & pingo_rc3 command:
pingo_rc3.exe -pngfilter=100 -noconversion -nosrgb -nodate -sa "%%A"
cwebp.exe -mt -m 6 -af -near_lossless 60 -alpha_filter best -progress "%%A" -o "%%~nA.webp"

My understanding is that near_lossless 60 will provide sufficient quality for all uses where 8-bit per channel RGBA is ok and does not need an additional butteraugli analysis.

If someone has generated an image with webp near_lossless 60 that has a visually observable degradation, I'd like to learn about it.

3. non-photographic image and jpeg encoder quality suggest

Compress png rgb24 image, my process is using quality 95 to first compress,
if compressed image not below safe butteraugli distance, increase quality to second compress.

In my png rgb24 image set, butteraugli assessment jpeg quality 95 doesn't get good butteraugli distance,

jpeg quality 95 butteraugli distance => [1.5 ~ 1.6], [1.7 ~ 2.5]
jpeg quality 95 butteraugli xl distance => [1.2 ~ 1.6], [1.7 ~ 2.2]

but in cjpeg usage.txt :

If i compress non-photographic image to jpeg and want near psychovisual lossless, it is necessary using
above quality 95 to compress those image?

My experience is that libjpeg quality 95 with yuv444 is often enough. About 1/1000 images it is necessary to go to 97-98.

or in *Reference 03, possibly butteraugli is too sensitive in some non-photographic image?

It has been argued that butteraugli is not sufficiently sensitive for DCT8x8 artefacts.

Butteraugli does not make an attempt to model the pupil size, making it sometimes too sensitive and other times not sensitive enough.

mozjpeg command:
cjpeg.exe -optimize -progressive -quality 95 -quant-table 3 -sample 1x1 -outfile "mozjpeg\%%~nA.jpg" "%%A"

Update 20200929:

I using jpeg xl sjpeg features, sjpeg can get great butteraugli distance in quality 96,
but look like sjpeg features doesn't using jpeg xl vardct or modular mode?

sjpeg is a new jpeg encoder created by the WebP library maintainer Pascal Massimino.

I don't know which jpeg encoder (after guetzli) gets the best butteraugli scores or has the best psychovisual properties otherwise. I didn't follow the development of that field after the launch of guetzli.

​Size:
png - 12mb
mozjpeg -q 97 + jpegtran progressive 4.46mb
jpegxl sjpeg -q 96 + jpegtran progressive 4.0mb

cjpegxl command:
cjpegxl.exe" "%%A" "xl\%%A" --jpeg1 --jpeg_quality=96 --progressive

4. webp lossy q100 and butteraugli distance

I test another non-photographic image set in webp lossy q100, but some image get larger butteraugli distance,
it possibly webp lossy 420 subsampling and fancy upsampling will make some larger errors in some area?

I have seen lossy WebP have problems on the borders of the images. YUV420 can be a problem.

You can dump the butteraugli heatmap and see by yourself the area where butteraugli thinks the problem is.

and i test webp lossy alpha(alpha_q) features, this features will increase butteraugli distance,
but i don't understand, why lossy alpha will effect butteraugli distance?

Butteraugli models human vision, and humans don't see translucency. So, it should not affect the valuation.

The butteraugli executable, when seeing a translucent image, places the compared pair on white background and measures the scores.

Then it places the same pair on a black background and measures the scores again.

It reports the worse score of the two measurements. Losses in alpha are noticed by this methodology.

This is not a perfect approach, but it may be better than nothing.

[cssignet]

​

my process is using near-lossless 60 and pngfilter=100 to first compress

​

​pingo_rc3.exe -pngfilter=100 -noconversion -nosrgb -nodate -sa "%%A"
cwebp.exe -mt -m 6 -af -near_lossless 60 -alpha_filter best -progress "%%A" -o "%%~nA.webp"

if you actually combine those, you would do loss over loss — so the result of near_lossless would be biaised and non-representative. if you want to compare, i would suggest instead to experiment the latest pingo with WebP, since all transformations (-webp-lossless, -webp-lossy, -webp-near, -webp-color, etc.) have heuristics/pre-processing which would not be done by the reference encoder atm; those would affect size/metrics, but could not be always better

[Jyrki Alakuijala]

​if you actually combine those, you would do loss over loss — so the result of near_lossless would be biaised and non-representative. if you want to compare, i would suggest instead to experiment the latest pingo with WebP, since all transformations (-webp-lossless, -webp-lossy, -webp-near, -webp-color, etc.) have heuristics/pre-processing which would not be done by the reference encoder atm; those would affect size/metrics, but could not be always better

Well spotted!

Simply:

cwebp --near_lossless=60 -q 100 -m 6 xyzzy.png -o xyzzy.webp

(in cwebp near-lossless and lossless there are too parameters controlling the effort of computation, -q and -m)

I haven't tried pingo myself, but the published results suggest that it can be a better option than cwebp.

[cssignet]

​

I haven't tried pingo myself, but the published results suggest that it can be a better option than cwebp

​i consider my tools more like experimental stuff, implementations/demos of ideas: they do not pretend (and imo should not) replace well tested reference, but results could be compared sometimes

I have seen lossy WebP have problems on the borders of the images

not sure if it is the same issue, but among several solutions, some simple pre-processing (RGB->255 if a=0) could be done and would impact quality/size

original.png

cwebp -q 100 -sharp_yuv (56.78 KB)
butteraugli: 5.670024
butteraugli xl: 8.0223999023 (3-norm: 2.246375)
ssimulacra xl: 0.01917124

pingo -webp-lossy=100 (51.29 KB) <-- this use sharp_yuv too
butteraugli: 1.928053
butteraugli xl: 1.6332571507 (3-norm: 0.712645)
ssimulacra xl: 0.00517459

some simple pre-processing (RGB->255 if a=0) could be done

edit: better solution could be done

[cssignet]

about the near_lossless in cwebp, perhaps i am wrong but it seems that the current implementation would change an alpha value from 1 to 0, even with low quantization value. i am not sure this would be the expected behavior, since it would impact things significantly (R-G-B should be altered by the alpha optimization). reusing the sample from my precedent post:

cwebp -near_lossless 80 (112.46 KB)
butteraugli: 1.606584
butteraugli xl: 50.7915344238 (3-norm: 15.189873)
ssimulacra xl: 0.03084179

pingo -webp-near=80 (113.97 KB) <-- avoid quantization on alpha
butteraugli: 0.844221
butteraugli xl: 1.0970934629 (3-norm: 0.365217)
ssimulacra xl: 0.00218288

-near_lossless 80 (112.70 KB) <-- avoid quantization if alpha < 2
butteraugli: 1.609372
butteraugli xl: 1.1307358742 (3-norm: 0.364205)
ssimulacra xl: 0.00218043

[Lithium Flower]

rework some question in 20201221 new post

[skal]

​

original.png

cwebp -q 100 -sharp_yuv (56.78 KB)
butteraugli: 5.670024
butteraugli xl: 8.0223999023 (3-norm: 2.246375)
ssimulacra xl: 0.01917124

pingo -webp-lossy=100 (51.29 KB) <-- this use sharp_yuv too
butteraugli: 1.928053
butteraugli xl: 1.6332571507 (3-norm: 0.712645)
ssimulacra xl: 0.00517459

edit: better solution could be done

Interesting... Is this happening only for qualities around q=100? Or at lower ones too? (q=60-80 for instance)...

skal/
​

[Lithium Flower]

rework some question in 20201221

@ Jyrki Alakuijala

sorry, my english is not good.

I'm sorry for reply so lately.
Thank you for your reply, I'm really grateful.

1. use butteraugli metric compare different encoder and maxButteraugliSocre

I develop a python multithread program use butteraugli metric compare different encoder,

and choose below maxButteraugliSocre and smaller image file.

maxButteraugliSocre setting on 1.6 is a great value? or 1.6 is too risk?

2. libjpeg q95 and Butteraugli Socre

in libjpeg q95 most image can get Butteraugli Socre below 1.6,

but some image can't reach 1.6, need use libjpeg q96, q97.

libjpeg not like Guetzli, pik, jpegxl have vardct,

if i want my image can get great quality, use libjpeg q96, q97 is necessary?

3. Butteraugli Socre and Butteraugli_jxl p-norm reference

In my first post *Reference 02 Butteraugli Socre reference and guetzli quality.cc,

those score reference list still available in Butteraugli and Butteraugli_jxl?

https://github.com/google/guetzli/bl...quality.cc#L26

Butteraugli_jxl p-norm have quality reference?

I collect some comment, could you teach me how to read p-norm?

below 1.0 is fine, above 1.0 is bad?

max Butteraugli // q95+(q95~q99)

12-norm // q95+(q95~q99)

?p-norm // q80-q92

3-norm(6-norm) // q60-q75

// Butteraugli_jxl --error_pnorm=12 (3-norm, 6-norm, 12-norm, 24-norm, 48-norm)

4. Butteraugli and Butteraugli_jxl return different butteraugli score

if Butteraugli and Butteraugli_jxl return different butteraugli score,
i should choose Butteraugli score or Butteraugli_jxl score?

exsample:

butteraugl 1.748285

butteraugl_jxl 1.3875815868,0.553373

butteraugl 1.650636

butteraugl_jxl 1.9129729271,0.648280

5. lossy alpha(translucency) feature

humans can't see translucency, webp lossy -alpha_q have compression factor,

if i setting -alpha_q compression factor to 1 or 0, image translucency will get more lossy compression,

maybe lossy alpha this feature will make some unvisible bad effect?

I find some lossy alpha information in AV1 reddit discuss,

i very curious this question, could you teach me about this feature?

https://www.reddit.com/r/AV1/comment...a_compression/

When the alpha channel is opaque or uniform (all one value), lossless AV1 alpha encodes should be very small.

However, for any kind of "interesting" alpha (masks, gradients, etc), I think it is a mistake to do lossy encoding as

the kinds of artifacts you'll get will not just be a slightly blurrier image, but your masks will stop lining up or

you'll get holes in places you didn't intend.

https://developers.google.com/speed/webp/docs/cwebp

-alpha_q int

Specify the compression factor for alpha compression between 0 and 100. Lossless compression of alpha is achieved using

a value of 100, while the lower values result in a lossy compression. The default is 100.

​

@ cssignet
Thank you for your reply, I'm really grateful.

I'm sorry, i did't make myself clear,
I using butteraugli metric compare pingo near-lossless and pingo pngfilter,
and choose below maxButteraugliSocre and smaller image file.

Thank your suggest, pingo near-lossless is working very well,
and Thank you develop pingo.

[Jyrki Alakuijala]

sorry, my english is not good.

General advice on communicating on forums: ask one thing in one post. Don't be worried to post many posts.

Your communication is clear and structured -- no reason to apologize for it.

1. use butteraugli metric compare different encoder and maxButteraugliSocre

I develop a python multithread program use butteraugli metric compare different encoder,

and choose below maxButteraugliSocre and smaller image file.

maxButteraugliSocre setting on 1.6 is a great value? or 1.6 is too risk?

Depends on viewing distance. At a viewing distance of 900 pixels, 1.6 is on the high side. I recommend values between 0.6 (perfect) to 1.2 (with some issues).

If your viewing distance is much longer than 900 pixels, the current versions of butteraugli become less efficient and less accurate in modeling the viewing. I'll probably improve on that in the next two years.

2. libjpeg q95 and Butteraugli Socre

in libjpeg q95 most image can get Butteraugli Socre below 1.6,

but some image can't reach 1.6, need use libjpeg q96, q97.

libjpeg not like Guetzli, pik, jpegxl have vardct,

if i want my image can get great quality, use libjpeg q96, q97 is necessary?

My personal viewing experience is that about 0.1-0.3 % of images at yuv444 q97 jpegs still have visible faults. Mostly problematic areas are deep reds containing gray details. Somehow velvet-like red cloth is often very difficult.

For practical use I consider jpeg yuv444 q93 or q94 great compromises and q95 good quality. Yuv420 is irritating for me.

3. Butteraugli Socre and Butteraugli_jxl p-norm reference

In my first post *Reference 02 Butteraugli Socre reference and guetzli quality.cc,

those score reference list still available in Butteraugli and Butteraugli_jxl?

Butteraugli_jxl p-norm have quality reference?

I collect some comment, could you teach me how to read p-norm?

below 1.0 is fine, above 1.0 is bad?

Low p-norm values depend on how large are the areas that are easy to encode. This means that anything else than the maximum aggregation (infinite-p-norm) is more or less indicative only and application dependant.

In some use case (like kodak test images) a p-norm of 14 was found to work well.

Butteraugli's p-norm is not a simple p-norm but a mixture of three p-norms, N, 2*N and 4*N. This is to have a mixture of emphasis of not making big errors in one area, but also giving some pressure on making improvements in large areas.

Likely the p-norm value of 6 to 12 could be a good compromise. I don't have a guideline how to interpret those values.

With p-norm of 3 I can observe that good compression produces a butteraugli p-norm value * BPP is roughly 1.0 for complex detail-rich images.

4. Butteraugli and Butteraugli_jxl return different butteraugli score

if Butteraugli and Butteraugli_jxl return different butteraugli score,
i should choose Butteraugli score or Butteraugli_jxl score?

exsample:

butteraugl 1.748285

butteraugl_jxl 1.3875815868,0.553373

butteraugl 1.650636

butteraugl_jxl 1.9129729271,0.648280

Shouldn't matter in the end. All butteraugli's have been calibrated to the same reference corpus with pretty good results. The reference corpus contains 2500 image pairs in the area of 0.6--1.3 max butteraugli values. Earlier butterauglis may be slightly more accurate, later butterauglis more practical and compatible with compression. Later butterauglis can be slightly better at detecting faint larger scale degradations as they do recursive 2x multiscaling of top of the Laplacian pyramid, not just a single run of four levels of Laplacian pyramid.

I, as the author of butteraugli, use the latest butteraugli in jpeg xl for my use.

5. lossy alpha(translucency) feature

humans can't see translucency, webp lossy -alpha_q have compression factor,

if i setting -alpha_q compression factor to 1 or 0, image translucency will get more lossy compression,

maybe lossy alpha this feature will make some unvisible bad effect?

I find some lossy alpha information in AV1 reddit discuss,

i very curious this question, could you teach me about this feature?

One early version of alpha_q just tried to minimize sum of squared error. I haven't followed up on what is going on there, so I don't know the current state. If one considers maximum error (like in butteraugli) minimizing sum of square error is not the best strategy. Considering PSNR, it is great -- a near optimal strategy.

[Lithium Flower]

​@ Jyrki Alakuijala
Thank you for your reply

here is my program structure,
First i split image type by color space, rgb24, rgba32, pal8,

in present version maxButteraugliSocre set 1.66, minButteraugliSocre set 1.55.
but 1.66 is too risk, i researching a new maxButteraugliSocre,

rgb24 using libjpeg [q92, q93, q94] [q95] [q96, q97, q98] (yuv444),
compress q95 and use Butteraugli to quality assessment,
if q95 below maxButteraugliSocre and above minButteraugliSocre use q95,
if q95 below maxButteraugliSocre use q92, q93, q94,
if q95 above maxButteraugliSocre use q96, q97, q98,

rgba32 using pingo pngfilter 80,100 and near-lossless 40,60(convert to png),
use Butteraugli to quality assessment and find below maxButteraugliSocre and smaller file.

pal8 using png lossless or webp lossless compress.

in multithread processing, 1 image need 2~3s(average case).
[CPU:r5 1600 6c12t(all core 3.65oc), RAM:16GB]


but in ButteraugliSocre i get some problem,
Butteraugli_jxl assessment some image in q95 still get maxButteraugli 1.6~2.0,
in q97 still some image get high maxButteraugli(1.6~2.0),

I think in non-photographic images use over q95 is not worth and image will very huge,
in next version libjpeg probably can set in [q91, q92], [q93], [q94, q95],
pingo still keep setting [pngfilter 80, 100, near-lossless 40, 60],
and add check p-norm to quality assessment,

if my target quality is libjpeg q92~q95,
i want use 6-norm and 12-norm check this image quality area, maxButteraugli look like working in q95+ quality area,
could you give me some suggest, how to use 6-norm and 12-norm guide my program to find good quality(can allow not easy to see error)
and smaller file?
thank you very much for your help

​=============================================== ========================================
Please forgive me if I’ve offended you,
in Butteraugli_jpegxl p-norm flag is change, --error_pnorm=12 is unavailable,
new flag is, --pnorm <int> [3, 6, 12, 24, 48]
https://gitlab.com/wg1/jpeg-xl/-/blo...i_main.cc#L127

sample:

01_02_09_libjpeg_q95.png
maxButteraugli: 2.0917696953
3-norm: 0.708287
6-norm: 0.958020
12-norm: 1.274732
24-norm: 1.579831
48-norm: 1.802553
Butteraugli_old: 1.975622

CloudySky1_pf100.png
maxButteraugli: 2.3613216877
3-norm: 1.706211
6-norm: 1.984321
12-norm: 2.156714
24-norm: 2.253438
Butteraugli_old: 1.234330

CloudySky2_pf100.png
maxButteraugli: 1.8451137543
3-norm: 0.671635
6-norm: 0.902970
12-norm: 1.199747
24-norm: 1.455440
Butteraugli_old: 1.003504

CloudySky1_pf80.png
maxButteraugli: 1.4213246107
3-norm: 0.745485
6-norm: 0.904969
12-norm: 1.072658
24-norm: 1.203832
Butteraugli_old: 1.440932

CloudySky2_pf80.png:
maxButteraugli: 1.7285125256
3-norm: 0.758040
6-norm: 0.920547
12-norm: 1.136067
24-norm: 1.347241
Butteraugli_old: 1.626053

[Lithium Flower]

​update 20201225:
I think i find a answer, in this issue,

https://github.com/google/butteraugli/issues/15
take an average of 3, 6 and 12-norms to support decision making and mostly use the maximum based butteraugli score for reference only.

use average pnorm to assessment, and average pnorm need below 1.0.

but this method need too many process time, i need find another method to handle this.

[Lithium Flower]

​I using three value to assessment image quality,average of 6-norm and 12-norm, 12-norm, and maxButteraugli,


// q91 too risk, q97 not worth
libjpeg [q92, q93, q94 ,q95 ,q96]


use two conditions to check file, linear searsh,
if first condition fail, use second condition,
if two condition fail, use libjpeg q96.


first condition:
averageNorm(6-norm + 12-norm) < 1.0
& 12-norm < 1.2
& maxButteraugli < 1.55


second condition:
averageNorm(6-norm + 12-norm) < 1.0
& 12-norm < 1.2
& maxButteraugli < 2.1


I'm not sure this good idea, maybe i made a mistake?

[Jyrki Alakuijala]

​I using three value to assessment image quality,average of 6-norm and 12-norm, 12-norm, and maxButteraugli,

You don't need to do anything, you can just call --pnorm=3 and it automatically uses 3, 6, and 12 internally and mixes them in equal portions. That argument only specifies the lowest value and the two higher are generated internally by 2 * N and 4 * N.

(For high quality work pnorm=12 may be more appropriate. It would compute internally 12th, 24th and 48th norm.)

[Lithium Flower]

​@ Jyrki Alakuijala
Thank you for your reply

Actually i want my program imitate jpegxl and guetzli use Butteraugli to assessment image,
could you teach me, if in cjpegxl target distance is 1.0 or 1.5, cjpegxl use which pnorm and maxButteraugli to find best image?

My compress program is prepare use on legacy program, but look like use cjpegxl compress png and jpg to jxl,
and when legacy program need use image, use djpegxl convert to original format is best method,
but i still want finish this compress program.

use 12-norm to assessment image, and use maxButteraugli make sure compressed image don't have big error,
I think condition probable like this?

condition:
12-norm < 1.2
maxButteraugli < 1.6

Fail: use q96 and near-lossless 60

libjpeg [q92, q93, q94 ,q95 ,q96]
pingo [pngfilter 80, 100, near-lossless 40, 60]

[Jyrki Alakuijala]

​Actually i want my program imitate jpegxl and guetzli use Butteraugli to assessment image,
could you teach me, if in cjpegxl target distance is 1.0 or 1.5, cjpegxl use which pnorm and maxButteraugli to find best image?

Cool. The basic answer is that no pnorm value is used there. You should get the same .jxl file regardless of the pnorm goals.

In butteraugli-iterative version of cjxl:

Adaptive quantization field is searched using gradient-like search where the gradient is approximated from a simple multiplicative relation of bpp * butteraugli score.

There, we use the butteraugli scores directly, no reason to aggregate. There are some tricks that are used, like if by default we get a really good butteraugli score, I don't make it worse by more than 40 %.

In guetzli we used a less principled and more heuristic search that worked better for max-butteraugli score.

All local correction that I used use a local maximum butteraugli score, i.e., the optimization algorithms do not aggregate using a p-norm.

In cjxl, the optimization over the max-butteraugli correctness is done using a quadratic objective, but it is a quadratic (L2) optimization over the local max butteraugli scores.

For coming up with heuristics and default quantization matrices etc. I used pnorm=3 a lot for finging balanced heuristics. It is much easier to optimize for than more appropriate values like pnorm=6..12. The optimization space is flatter because all pixels participate more evenly.

[Lithium Flower]

@ Jyrki Alakuijala
Thank you for your reply

I think my word is too conceited in previous post, sorry,
My program is more like Jpeg Archive(SSIM), cjpeg-dssim(dssim), squoosh(butteraugli),
use single metrics to assessment image, jpegxl and guetzli doing more advanced thing.

I have some question in butteraugli_jxl, i compress some image in jpegxl_v0.2 d1.0 s8,
butteraugli_jxl assessment some image get big maxButteraugli(2.29328537),
https://docs.google.com/spreadsheets...tX7dDBCObf0zCQ

The -d distance target is only a target, not a guarantee.

About this answer, cjxl use heuristics and target distance(1.0) to keep quality and compess image,

In old butteraugli, if image get big butteraugli distance, this image will assessment bad quality image,
or have big error in this image.
butteraugli_jxl maxButteraugli metrics still like old butteraugli is a error metrics?

if cjxl d1.0 s8 image in butteraugli_jxl get big maxButteraugli, butteraugli_jxl assessment this image have big error,
but cjxl assessment this image is conform 1.0 target distance, this image should assessment good quality or bad quality?

If i using butteraugli_jxl to compare different image encoder, like libjpeg, cwebp, cjxl,
only use maxButteraugli to assessment image quality(big error), this method can correct reflect image quality?
or i should use other metrics, like butteraugli_jxl pnorm, help to assessment image and get more accurate result?

In my non-photographic image set test, 6-norm and 12-norm can reflect image quailty in high area,
​but i don't know, how to use pnorm to assessment image in correct method.

[Jyrki Alakuijala]

butteraugli_jxl maxButteraugli metrics still like old butteraugli is a error metrics?

Both work relatively well. Should not matter much to the utility of your system which one you use. For individual images there can be differences, but likely that neither is more accurate for a large group of images.

if cjxl d1.0 s8 image in butteraugli_jxl get big maxButteraugli, butteraugli_jxl assessment this image have big error,
but cjxl assessment this image is conform 1.0 target distance, this image should assessment good quality or bad quality?

You can consider looking at the butteraugli heat map and seeing by yourself how bad the situation is.

If i using butteraugli_jxl to compare different image encoder, like libjpeg, cwebp, cjxl,
only use maxButteraugli to assessment image quality(big error), this method can correct reflect image quality?
or i should use other metrics, like butteraugli_jxl pnorm, help to assessment image and get more accurate result?

pnorm may be more appropriate for -d1.5 and worse or -q 90 and worse.

So, use maximum butteraugli for decent quality photos (-d1.0 or -q 94 yuv444), the worse the quality the lower the norm should be (likely useful range is 3-24).