I am unable to dowmload. I am signed in :-(

Zipped apefos patch

Hey apefos, first congratulaitons and thanks for these settings and your time spent working on these! Looks really promising - would you recommend this over Sanity X? (same stability but better quality?)

Are these patches about 720p only or also about Full HD modes? If yes, what is modified in Full HD modes?

@apefos: is "The Apefos Patch 2" concerned the 1080p modes too (24H, 24L, HBR, etc.)?

Is the ISO improvement on 1080p, not only on 720p modes?

I am not working with video/audio anymore, but I am keeping the GH2 just for some fun.

I developed the version 2 because I found the version 1 shows some compression artifacts in iso from 160 to 2000 (very little, almost unoticeable). But it is better for very high iso from 3200 to 12800. The higher quantization in the luminance matrix in the version 1 helps to handle the heavy noise from isos from 3200 to 12800 but the downside is that it affects the texture in lower isos a little bit, mainly in light brown surfaces.

The version 2 has a lower quantization in the luminance matrix which improves the textures for isos from 160 to 2000, but can show some artifacts in very high iso from 3200 to 12800. This happens because the lower quantization cannot handle the heavy noise from the very high isos, but it improves the textures for lower isos.

People may ask: why don't you find the quantization settings in the luminance matrix which can do both things? Simple answer: because the datarate does not allow, and if I increase the datarate the auto electronic features will not work.

Both versions 1 and 2 works for 1080p and 720p.

Both versions works for PAL and NTSC.

Both versions can use the electronic auto features from camera body and lenses. (with these features enabled in NTSC the H is more stable compared to SH and FH is more stable than FSH, in PAL all are fine). (but if you use just one or two auto electronic features in NTSC SH and FSH it can work ok, and the SH and FSH are better for textures).

Both versions avoids the stop recording and camera freeze (tested with the 45Mbps card).

Both versions have the datarates in the maximum stable limit.

Both versions have all improvements from my previous patches, Workhorse, Boson, Gspot and DCT.

To perceive better the difference from versions 1 and 2, do some video in iso 160, 2000 and 12800 with both versions in 720p50 or 720p60 and then slow down to 24p or 25p in timeline to perceive how the slowmotion shows the textures in surfaces for both versions in iso 160, 2000 and 12800.

Enjoy!

@frullaccia I am happy with your words because you just understood it. You understood the apefos version2: to respect the GH2 hardware to make the auto electronic features to work properly and get the best possible image quality in this way.

What happens in the GH2 is: the camera hardware has a limit, no matter the memory card you use, I tested the 45Mbps card and the new 95Mbps card I bought (reliable from trusted seller and tested in usb 3.0). The limit happens this way: if you increase the datarate to get a better image, the native lenses, the image stabilization, and the idynamic will stop recording. What I did is to find the highest datarate to get best possible texture and keep using these auto features without stop recording. I did lots of test with matrix design and I found the lowest possible compression in the matrix that can keep recording without stop.

Also there is some things very important in the matrix: the matrix design must respect a math design which will not introduce problems in the image. This design is a mathematic calculation and XY design that can do it in a way that the codec reduces the data but does not shows a pattern in the image.

I got a very useful help from the "masters of hack" some time ago to understand the matrix desing and I myself perceived that the maths in the matrix should do the job without introduce patterns in the image. The patches from others that I tested did a good job but it introduces some pattern artifacts. I did analises in many patches to perceive the benefits and drawbacks... This way I did a setting which can do the compression without anoying problems.

the nebula is like this:

004 004 004 127

004 004 004 127

004 004 004 127

127 127 127 127

004 in the dct number is very little and shows a noise that is so small that it get into a flat look some times.

Also the high compression in the XY 127 lines shows high compression which introduces vertical and horizontal heavy compression lines in the image.

The Canis Night uses very low numbers which can be good for low light but crashes in light/shadow images.

The Flowmotion starts with number 007 in the dct number which I found is high for a fine noise, so the noise is a little gross. And it uses different settings for the fallback and interframe matrix which makes the camera processor to stress and sotp recording. The AV2 uses same careful tweaked matrix for all, intra inter and fallback so the camera processor does not stress.

In few words the solution I found is: to find the best main dct first number to get best fine noise, to find the best increase in next numbers XY first lines to get best image, to find the best numbers in the fill matrix to avoid a pattern, to make intraframe, interframe and fallback to be the same to not stress camera processor, and get the good compression. The luminance matrix is:

06 09 12 18

09 15 21 33

12 21 39 45

18 33 45 63

You can perceive that all numbers in the luminance matrix are multiple of 6, or half 6.

The first number being 6 means the best noise without get into a flat texture or gross noise. 4 in the dct number makes the noise to be very fine and some surfaces starts to be flat so the noise becomes different in image areas, 7 makes the noise to start to be gross in dark areas. 6 is the sweet spot, better than 5. 6 makes the gradients to be better also.

The other numbers in apefos2 means increase in quantization without get into artifacts in iso from 160 to 2000 and without stop recording in the datarate limit.

If you repeat this matrix side by side and down, you will perceive that there will be not high numbers so close to each other and this means no pattern in image.

the dct matrix, as you can see in wikipedia, has a frequency which increases to the right and to the bottom. From the dct number to the right and to the bottom I tested 1x, 1,5x, 2x, 3x and 1x, 2x, 3x, 4x I liked more the first option, less increase in quantization. In the diagonal the frequence increases more, so the numbers in the midle increases a little. in the bottom line and in the right line I used an increase of half value between the two options above: 1x, 1,75x, 2,5x, 3,5x but keeping multiples of 6 or half 6. this is the lowest possible frequency increase in these lines which avoid stop recording in this luminance matrix design.

the chrominance matrix is:

006 009 012 018

009 015 021 114

012 021 120 126

018 114 126 156

I believe that same matrix for luma and color can make the noise and texture to be similar in luma and color channels.

the higher numbers in the end are for minimize the image flickering. instead of using 127 in all positions I used different numbers to try to make the image texture better, all multiple of 6 and with similar quantization of 127 in all positions.

In chroma matrix, the Flowmotion does a good job using 127 in all the six end numers to avoid chroma flickering. I perceived that using 114, 120, 126 and 156 could keep same quantization and differ numbers to get a little difference to try to avoid a flat chroma look.

this is very fine differences but in the end... it is what makes you to say your words about the Apefos V2 to be "the best of the bunch".

@Apefos ...!

some matrices for comparison,

nebula 1080 (same luma/chroma)

004 004 004 127

004 004 004 127

004 004 004 127

127 127 127 127

nebula 720 (chroma is like above)

004 004 004 016

004 004 004 032

004 004 004 048

016 032 048 050

Flowmotion:

intra L

007 011 015 021

011 015 026 026

015 026 028 028

021 026 028 028

intra C

005 007 010 015

007 008 013 127

010 013 127 127

015 127 127 127

inter L

009 014 019 027

014 019 033 033

019 033 036 036

027 033 036 036

inter C

007 009 013 019

009 010 016 127

013 016 127 127

019 127 127 127

fallback:

in L it changes last 3 digits to 255, this is high quantization

in C it changes all 127 to 255, this is high quantization

increases quantization a lot, and makes gradients not so good as apefos v2

using same numbers near to each other is not good idea, they can implement same quantization and make pixels to be same color and luma or very near results, turning textures into a flat look. nebula does this in lines and rolls using 127. flowmotion does this also using 026, 028 and 127

the luma quantization in Flowmotion is 1.2 in Apefos v2 it is 1.27, a little higher but better design and better dct number for noise, this little higher helps to avoid fallback

the chroma quantization in apefos v2 is 1.91 in flowmotion it is 1.59, due to this flowmotion needs to use fallback to increase quantization when there is motion, apefos v2 does not need because the main matrix is already better calibrated in quantization, and better noise using the 6 as first digit and better design with different numbers multiples of 6 or half six and increasing frequency.

a little more quantization in color was good helps not need fallback in color and also in luma

(fallback is when the main matrix has low quantization and in some situations the camera uses fallback which is another matrix with higher quantization to avoid artifacts, apefos v2 does not need)

interframe matrix are matrices to calculate the information which will be used from one frame to another, the temporal compression, if the intraframe matrix have a low quantization, the interframe matrix can be used with a high quantization to save data, but the temporal information will decrease in quality from one frame to another. if the interframe matrix is diferent it will make the processor to work a lot. using same matrix for the frame and for the temporal compression makes processor be easy and better balance in overall quality, more constant noise and texture over time and over image areas.

camera processor works less in apefos v2 and image is ok.

apefos version 1 Luma (perceive that the begining is the same in version 1 and 2. version 1 has higher quantization in last digits to handle the heavy noise from very high isos)

006 009 012 018

009 015 021 063

012 021 072 084

018 063 084 126

apefos version 2 Luma (less quantization in last digits allows better texture in lower isos, dct frequancy is very well respected, 006 as first digit is best for noise fine/gross calibration, there is no very high quantization in any number, no similar numbers so close to each other avoids patterns

006 009 012 018

009 015 021 033

012 021 039 045

018 033 045 063

apefos 1 and 2 chroma: the higher quantization in last digits does not introduce problems because the luma channel keeps the texture ok. these higher numbers helps two things: to avoid chroma flickering and increase quantization to allow the Luma matrix to have lower quantization. even using higher numbers it respects the frequancy increase and there is no similar numbers close to each other

006 009 012 018

009 015 021 114

012 021 120 126

018 114 126 156

The Canis Night crossfire Luma:

002 003 004 005

003 004 006 007

004 006 008 010

005 007 010 020

Chroma:

002 003 004 006

003 007 008 012

004 008 016 018

006 012 018 026

this is so low quantization and will need very high datarate and will need long gop to save data, also the second quantizer will increase automaticaly , so lots of drawbacks happens

now the Moon T7: (same luma chroma) it does not respect the dct frequency and creates diagonal lines, quantization is very high in first digits and can turn some things like dark areas or constant colors in flat look.

012 026 040 056

026 040 056 213

040 056 048 021

056 048 021 213

now the Moon T8 Luma (lower numbers in begining makes the need of highr numbers in the and, noise very fine can becomes flat in some surfaces, also heavy compression in last 3 pixels

004 008 013 016

008 013 016 032

014 017 034 072

018 036 096 127

and chroma: (same as T7)

012 026 040 056

026 040 056 213

040 056 048 021

056 048 021 213

interest thing is when you start to understand matrices you can predict what will happens without test it on camera.

I could keep doing analysis of other patches but I think this is enough.

@apefos Sure.... !

where is the like button ?

this is the version 3, it can be unstable, I will keep myself on version 2.

I think the auto electronic features will not work on 720p or HBR

Also native lenses can make it more unstable than legacy lenses with adapters.

the version 3 increases datarate and also frame limit and frame buffer accordingly.

version2 datarates are: 96, 76, 76, 60

version 3 datarates are: 100, 84, 84, 68

the version 3 Luma Matrix reduces quantization from 1,27 to 1,17 in this design:

006 009 012 018

009 015 021 027

012 021 030 036

018 027 036 054

the difference from version 2 is little, but maybe can increase texture and noise quality.

I did not test it on camera, but the maths are correct on everything

remember: it can be unstable!

if you want to try: