CryoPNG short introduction

**caveman** · 4th April 2011, 18:04

CryoPNG is OptiPNG with a dirty hack at the libpng level to actually perform a kind of advanced "Dirty Transparency" cleaning.
As explained by Sergey Chikuyonok (in this Smashing Magazine article) fully transparent pixels could have any RGB component since they are actually invisible, he suggested to turn them to black (0,0,0) to improve compression. A few tools have since implemented this trick (PunyPNG, PngOptimizer, ImageOptim...), but if you look closely at the PNG specifications you'll notice that images are filtered before being compressed. CryoPNG turns fully transparent pixels into something that will be nullified once filtered, and this can positively affect compression.

For instance the following mouse sample image used by PunyPNG (sizes in bytes):
228,066 Original
111,945 PunyPNG
105,035 Cryogen-x (CryoPNG script, also uses PNGOUT and DeflOpt)

For illustration purposes:

Top left: picture with alpha channel activated.
Top center: filter 0 (None) is nullified, fully transparent pixels are all black (classic dirty transparency)
Top right: filter 1 (Sub) is nullified, a fully transparent pixel has the same colour as its left neighbour
Bottom left: filter 2 (Up) is nullified, a fully transparent pixel has the same colour as its upper neighbour
Bottom center: filter 3 (Average) is nullified, a fully transparent pixel has the same colour as the average of its left and upper neighbours
Bottom right: filter 4 (Paeth) is nullified, a fully transparent pixel has the same colour as whether its left, upper or upper left neighbour (Paeth predicator is a powerful mess)

**Intrinsic** · 4th April 2011, 21:17

I've been following this for about the last 5 months and it can do some great stuff but have also found images which it removes visible(ie no alpha/trans) pixels from so stopped messing around with it.

And you're Frédéric Kayser? (yes i frequent the jonof forums ;p) or are you pretending to be him? as on those forums he replies to some of your questions...and you reply to those answers etc...smells fishy to me! at least confusing if you are talking to yourself on those forums lol.

**caveman** · 4th April 2011, 22:19

Originally Posted by Intrinsic

I've been following this for about the last 5 months and it can do some great stuff but have also found images which it removes visible(ie no alpha/trans) pixels from so stopped messing around with it.

And you're Frédéric Kayser? (yes i frequent the jonof forums ;p) or are you pretending to be him? as on those forums he replies to some of your questions...and you reply to those answers etc...smells fishy to me! at least confusing if you are talking to yourself on those forums lol.

Since JonoF forums closed a few weeks ago I'm moving over to the encode.ru forums.

Do you still have the images you mentioned? I did not encounter such behaviour, but I have stopped checking perfect match between optimized and non optimized images, visual inspection seemed fine enough. On which operating system did you try CryoPNG?

And yes, you're right I used two identities (I forgot Caveman's password once and could not access my private email from work so I created "Frederic Kayser" which is indeed my real name and it flattered my ego somewhat, since JonoF's Forum are now read only I just updated my nickname

).

**Intrinsic** · 4th April 2011, 22:30

This one, at least i think it was CryoPNG, could have been x128's TruePNG though which also had some odd transparency bugs but those were fixed in the most recent version.
http://uk.riftgame.com/img/classes/c...ing-cleric.png

And yeah i noticed the forums were closed down which is a shame.

Edit: And it was the dec 2010 windows version.

**m^2** · 4th April 2011, 22:57

Sounds interesting. Any chance that we get sources or binaries?

**caveman** · 4th April 2011, 23:21

Originally Posted by Intrinsic

This one, at least i think it was CryoPNG, could have been x128's TruePNG though which also had some odd transparency bugs but those were fixed in the most recent version.
http://uk.riftgame.com/img/classes/c...ing-cleric.png

And yeah i noticed the forums were closed down which is a shame.

Edit: And it was the dec 2010 windows version.

I will check this on a Linux box, I do not have a Windows system at hand right now.
By the way, I'm not sure but during my tests I've found that pngslim 1.0 had some problems calling pngoptimizercl when run in Windows Vista or 7 (XP was fine). I think I've found an easy fix for this, line 58 should be modified this way (simply add quotation marks around %1):

Code:

pngoptimizercl -file:"%1" >nul

I have improved the Unix cryogen scripts lately (not released yet), CryoPNG binaries have not changed.

Originally Posted by m^2

Sounds interesting. Any chance that we get sources or binaries?

I will probably (re)release everything by the end of the week including improved scripts, if you follow the old discussion in JonoF's forum from here, you'll find Linux x86, Mac OS X and Windows binaries.

**m^2** · 4th April 2011, 23:53

Originally Posted by caveman

I will probably (re)release everything by the end of the week including improved scripts, if you follow the old discussion in JonoF's forum from here, you'll find Linux x86, Mac OS X and Windows binaries.

OK, thanks.

**caveman** · 5th April 2011, 03:40

Originally Posted by Intrinsic

This one, at least i think it was CryoPNG, could have been x128's TruePNG though which also had some odd transparency bugs but those were fixed in the most recent version.
http://uk.riftgame.com/img/classes/c...ing-cleric.png

I did not find anything suspicious, I used all filters independently and converted the result using ImageMagick: convert source.png -alpha background target.png this produced exactly the same file each time, I will check on Windows 7 tomorrow.

calling-cleric.png
260,770 Original
242,294 Cryogen-x

I have enclosed the file produced by cryogen-x.

**Intrinsic** · 5th April 2011, 11:11

Yeah it seems fine so i did some checking and retesting and it turned out to be PNGKT killing off some pixels ;p

Plus there is another new tool just out, PNGWolf.

http://bjoern.hoehrmann.de/pngwolf/

If anyone can knockup a win32 binary that'd be cool ;p

**caveman** · 5th April 2011, 12:36

Originally Posted by Intrinsic

Plus there is another new tool just out, PNGWolf.

http://bjoern.hoehrmann.de/pngwolf/

If anyone can knockup a win32 binary that'd be cool ;p

It uses the Deflate implementation from 7-Zip and a genetic algorithm to select filters, really interesting!
OptiPNG uses the heuristic defined in RFC2083:
"The following simple heuristic has performed well in early tests: compute the output scanline using all five filters, and select the filter that gives the smallest sum of absolute values of outputs. (Consider the output bytes as signed differences for this test.) This method usually outperforms any single fixed filter choice. However, it is likely that much better heuristics will be found as more experience is gained with PNG."

I believe that PNGOUT uses different heuristics, pngcheck can be used to display the filter used on each line:

Code:

pngcheck -vv twit.png

File: twit.png (2013 bytes)
  chunk IHDR at offset 0x0000c, length 13
    32 x 32 image, 32-bit RGB+alpha, non-interlaced
  chunk IDAT at offset 0x00025, length 1956
    zlib: deflated, 32K window, fast compression
    row filters (0 none, 1 sub, 2 up, 3 avg, 4 paeth):
      1 4 4 4 1 1 1 1 4 4 4 4 4 4 4 4 4 4 4 4 4 2 4 4 1
      1 1 1 1 3 1 1 (32 out of 32)
  chunk IEND at offset 0x007d5, length 0

**Intrinsic** · 5th April 2011, 17:28

PNGOUTs is a lot simpler, basically 0 for indexed and 5 for anything else ;p his tests are a bit iffy though. He uses brute force using his own tool but for example uses pngout without any other options, and only z4 for advpng, where other levels can give better results. But still it has a lot of promise.

He has a bunch more info here: http://thread.gmane.org/gmane.comp.graphics.png.devel/

And also plenty of discussion on it.

**caveman** · 5th April 2011, 18:06

Originally Posted by Intrinsic

PNGOUTs is a lot simpler, basically 0 for indexed and 5 for anything else ;p

5 is the "mixed" mode where each line is filtered independently using filter 0, 1, 2, 3 or 4. What Bjoern is trying to do is to find an effective way to choose which filter is the best for each line and the overall image compression.

**Bjoern Hoehrmann** · 5th April 2011, 19:11

Originally Posted by Intrinsic

his tests are a bit iffy though. He uses brute force using his own tool but for example uses pngout without any other options, and only z4 for advpng, where other levels can give better results.

It is true for all the tools that certain settings beyond the ones I've chosen produce better results on occasion, 'pngwolf' is no exception. However, an exhaustive search for all the options is infeasible, so I went with the maximum suggested settings as best as I know them. I do very much encourage others to conduct tests they find more suitable and publish their findings.

**Intrinsic** · 5th April 2011, 19:22

Welcome Bjoern! I only just found your new tool recently and it's very exciting and can't wait to see what else you have in store for it. It's certainly something i'd work into my PNG compression tool once i am in a position to restart work on it and when combined with deflopt and defluff could produce some great results no doubt.

Plenty of very wise heads here to bounce ideas off of as i don't think many watch the gmain list.

**lorents17** · 5th April 2011, 23:10

Intrinsic
If it is possible an example where PNGKT or TruePng spoils a picture?

I Now will hand over all information X128, I just with it spoke today

**Intrinsic** · 5th April 2011, 23:24

This one Lorents is an example:
http://uk.riftgame.com/img/classes/c...ing-cleric.png

Only PNGKT had problems, it removes valid pixels or changes the colour value of them incorrectly. I noticed this as IrfanView i can easily displays colour totals and i saw the processed image had less colours. So i put it into a paint program and did a difference(+plus boosted levels to clearly show the problem) process on the images and saw the changes when there shouldn't be any. See the attached image.

**lorents17** · 5th April 2011, 23:40

Intrinsic
Now I will hand over this information x128, wait updating PNGKT in the near future.

**lorents17** · 6th April 2011, 09:04

Intrinsic
Look at the attached picture, here it is accurate advantages PNGKT are visible.
As you could notice, in your picture of loss insignificant.
It is possible to use TruePng for clearing, completely without quality loss.

**caveman** · 19th April 2011, 05:53

I did a quick test today, I tried to compress an RGBA image as two separate images (RGB and Alpha) and the overall size of the two images was smaller compared to the RGBA image.

RGBA : 105,035
RGB : 89,562
Alpha : 7,066

RGB + Alpha = 96,628

I've always thought that compressing images as RGBA was a bad idea, RGB and Alpha should be considered as two different data sources.

**caveman** · 19th April 2011, 06:40

Two more samples:

bigmac.png
RGBA : 33,906
RGB : 28,721
Alpha : 2,233
RGB + Alpha = 30,954

getadrink.png
RGBA : 65,317
RGB : 53,734
Alpha : 6,162
RGB + Alpha = 59,896

ImageMagick "convert" can easily be used to split an RGBA image:
convert RGBA.png -alpha off RGB.png
convert RGBA.png -alpha extract Alpha.png

**lorents17** · 19th April 2011, 08:52

Split by transparency

**Karhunen** · 25th February 2012, 01:24

Would you use ImageMagick's composite to "rejoin them" as follows:
RGBalphamask.png RGB.png -compose CopyOpacity -alpha on RGBAresult.png
Or do yo have a different solution ? Also, any ideas on setting BackGround color and select Alpha Transparencies on the command line ?
I can only use TweakPNG to set individual alpha indicies.

**caveman** · 5th April 2012, 03:23

Originally Posted by Karhunen

Also, any ideas on setting BackGround color and select Alpha Transparencies on the command line ?
I can only use TweakPNG to set individual alpha indicies.

As far as I know there's no command line tool that lets you directly modify tRNS or bKGD chunks, have you asked on ImageMagick's support forums/mailing lists?

**caveman** · 19th April 2012, 17:27

Originally Posted by Karhunen

Would you use ImageMagick's composite to "rejoin them" as follows:
RGBalphamask.png RGB.png -compose CopyOpacity -alpha on RGBAresult.png
Or do yo have a different solution ?

Had to do it today (I needed to move the alpha channel from one image to another), this command worked for me:
convert RGB.png Alpha.png -compose Copy_Opacity -composite RGBAresult.png

**Karhunen** · 4th September 2012, 04:28

Originally Posted by caveman

Had to do it today (I needed to move the alpha channel from one image to another), this command worked for me:
convert RGB.png Alpha.png -compose Copy_Opacity -composite RGBAresult.png

I was looking for the answer for this since I first posted it, and now I notice that Caveman answered it awhile back. As you noticed, transparency masks can compress quite differently from the RGB image. My motivation was testing FreeArc with two bitmaps, a PNM file and a windows BMP paletted file.

Btw, my transparency mask (BMP) seems to have 11 unique colors. I notice that they are greyscales of uniform distribution (see below). Is this redundancy i.e. aren't all Alpha Channels bi-level images ?
The command line that worked for me was convert RGB.pnm Alpha.bmp -compose Copy_Opacity -composite RGBAresult.png
Thanks!

PAL
16
0 0 0
17 17 17
34 34 34
51 51 51
68 68 68
85 85 85
102 102 102
119 119 119
136 136 136
153 153 153
170 170 170
187 187 187
204 204 204
221 221 221
238 238 238
255 255 255

**Karhunen** · 5th September 2012, 20:21

Originally Posted by caveman

As far as I know there's no command line tool that lets you directly modify tRNS or bKGD chunks, have you asked on ImageMagick's support forums/mailing lists?

GBMTOOLS allows command line chunks bkGD and tRNS, as long as the alpha is unassociated with the RGB:

back_index=#
transcol_rgb=#_#_#
back_rgb=#_#_#
Ignore_back

Thread: CryoPNG short introduction

Thread Tools

Display

CryoPNG short introduction

Similar Threads

Most efficient/practical compression method for short strings?

Introduction.

Posting Permissions