I did new torture tests in the death charts movie, with image stabilization enabled, this time I formated the card before recording, I did it handheld with different kinds of shaking, iso 6400:

SH 720p60 Time Recorded: 8m18s File Size: 3.99 GB (4,288,151,616 bytes) Total Datarate: 68805kbps - APPROVED! (ended recording)

I also did tests in the green grass with the image stabilization enabled after formatting the card:

The camera did not stop recording in shutter 1/4000 iso 6400 auto exposure shutter priority for some minutes. But after some minutes I started to press the half shutter many times to force the camera to redo the focus, and after some half shutter press, the camera stop recording. But it did not freeze, it returned to stand by, and the file is not corrupted, it playback in computer. So my conclusion is: with a formatted card the performance of the new SH is better, less stop recording. Maybe a formated 95Mbps card will avoids the stop recording even more. And maybe a slightly reduction in datarate will help.

So if you prefer a slightly datarate reduction for improve the stability in the 45Mbps card the versions C and D uploaded here below can avoids the stop recording better and keeps almost same quality.

I did tests in the green grass with shutter 1/4000 and image stabilization enabled with a slightly datarate reduction version, with the same card without reformatting it, the auto exposure was in shutter priority and I did lots of half shutter press to redo focus, lots of camera shaking and zoom in/out. it did not stop recording!!!

So I think I found the highest possible datarate for the 720p60 with the 45Mbps card, it is in the patch below, the C version has a reduced 24H and the D version has the higher 24H

Considering that the 88000000 is the datarate limit for the 45Mbps card in 720p60 with the kit lens and image stabilization enabled (this is what is typed in the ini file, the datarate results in recorded video file is lower):

If you have the 95Mbps card you can do the following: increase the datarate for 720p in steps of 8 in each test, from 88000000 to 96000000, 104000000, 112000000, 120000000, 128000000 and then you will find the limit of your card. do tests as I did, using the death charts movie and in the green grass.

If you have a lower speed card, for example the 30Mbps you can do the opposite, to reduce the datarate for 720p in steps of 8: from 88000000 to 80000000, 72000000, 64000000, 56000000, 48000000, 40000000 to find the highest possible for the kit lens with image stabilization enabled. do tests as I did, using the death charts movie and in the green grass.

if you change the datarate you must change the numbers in the TOP LIMITS for the same values, they are near the end of the ini file: 720p top setting=, 1080i top setting=, other modes top setting= (the numbers in these settings have three zeros less than the datarate numbers)

Versions C and D have a conservative never stop datarate in the H slot. In the SH there is the top limit for the kit lens considering the 45Mbps card..

Version A is a special version for people who have the Sandisk 32GB SDHC 45Mbps card. In the H slot it has the highest possible datarate for 720p with the kit lens and image stabilization enabled. In the SH slot it has the higher datarate for 720p and manual lenses without eletronic contacts. The 24H and 24L datarate was calibrated to allow the similar time recording of the SH and H respectively. Highest possible aggressive settings for this card in 720p. Version B is a little lower datarate compared to A in all modes, for a little more reliability.

I think that now the people already know how to change the datarate for their cards. I did a special tweaking for the Sandisk I have. If you have different cards you can tweak as instructed above. If you keep all the settings the same and change just what I said above you will find the limits for your cards.

these are three versions which seems to be the last ones, 720p IP, 1080p IB, for 45Mbps cards and above (link for download mts files coming soon):

seta: highest possible datarate in all modes, 24L match same time recording of SH, SH for manual lenses, H for native lenses. 96/76 - 104/88

setb: 24H and 24L lowered a little, 24H match same time recording of SH, 24L match same time recording of H, SH for manual lenses, H for native lenses. 76/64 - 104/88

setc: a little more conservative settings, lowered for all modes. 72/56 - 96/80, 24H match time recording of SH, 24L match time recording of H.

observations:

in the 24H and 24L modes the recorded video datarate is similar to the numbers typed in the ini file. in the SH and H modes the recorded video datarate is lower than the numbers typed in the ini file.

when I say match it is an approximation, similar time using the same amount of memory card.

here is the link for the original mts files shoot with the highest settings for 720p sh. there are other modes also, hbr, 24h...

when you see similar videos in the folder, they are for compare different recording modes and/or different isos.

perceive the low light random noise and good textures even in the dark low contrast areas.

feel free to download and try your denoise, grading, filmgrain, upscaling...

if you upload video tests, please point a link to this topic discussion.

thanks!

https://www.dropbox.com/sh/w573stryjec4sjy/AABW1-I6Rbu1A1QBeVb_IDdHa?dl=0

@NikolayM thanks for your test and feedback, I am glad it can be useful for you.

@LPowell @driftwood @duartix please, I have a question for you about improving the 720p even more...

in the dropbox folder link where I uploaded the original mts files, in the files 00105 SH and 00111 24H there is an interesting difference in streamparser which can make sense:

the 00111 24H file shows the B frames very high size compared to the I frames, the size of B frames is high. I think this makes the noise more random and less mud in B frames. the noise is more equal among the frames. This makes the 24H excelent looking.

the 00105 SH file shows the P frames more lower in size compared to the I frames, the size of P frames is low and when I go step by step, frame by frame in the timeline, I can see that in the I frames the noise in the girl's leg is better than it is in the P frames.

the 60fps avoids the eye to perceive this difference among the frames because each frame keeps small time in the screen, but doing a 40% slow motion to fit the 24p output makes this difference more noticeable, because each frame keeps more time on the screen. So my idea is:

StreamParser shows that in the SH file the average size of I frames is 250.000, and the average size of P frames is 106.000 Considering the gop3 have 20 I frames inside one second and 40 P frames inside one second, doing some math calculations I perceived that if I can lower the I frames average size 11% to 225.000 and raise the P frames average size 15% to 122.000 the datarate balance will be better, and the final datarate will be the same, so the stability will be the same. this can make a significant difference to make the noise texture more equal among the frames and would make the slow motion better looking.

So my question is: how can we do this in PTool? to raise the P frames size 15% and decrease the I frames 11% in size for the 720p SH ?

Thanks in advance.

Edit post: uploaded streamparser screen shoots. the ration between the average I and B in the 24H video is 1.54 the ratio between the average I and P frames in the SH video is 2.36 doing the better balance (11% / 15%) that I showed above the ratio will be 1.85, much better distribution of the datarate without decreasing the I frames size so much, because if we decrease the I frames size so much it can hurt the overall quality. So I think the 11% / 15% amounts for a better balance are ok and datarate will be the same. For slow motion videos without denoise this can make significant difference and for denoising purposes in very high iso I believe this will help also, because neatvideo will do a similar job in all frames and the results in denoised textures will be better.

This is the difference between the I and P frames in 720p SH. The texture in the I frame is better because the P frames are much lower size. When the video is in playback at 60fps the eye cannot perceive, but for slow motion it can be improved. in the post above I got an idea to improve it reducing the I frames size a little and increasing the P frames size a little. see the analysis above in previous post. problem is: I do not know how to do.

please @driftwood @LPowell @Ralph_B @bkmcwd @qwerty123 @papasarts @cbrandin @balazer @stonebat @Brian202020 @mpgxsvcd @sohus @sage @duartix and others, help to improve it with your knowledge!

Noise test in slow motion:

workhorse flowmotion patch gop3 IP frames quantizer zero 76Mbps maximum datarate for 720p

720p60 from GH2 rendered to 1080p24 with slowmotion

neatvideo to do sharpening and remove chroma noise, luminance noise was untouched

no grading, the look is from the camera adjusts

no filmgrain, the noise is from the camera high iso

I did new tests with quantizer 0, 14 and 17: quantizer 0 for 720p is better for both I and P frames in low light, so the last patches are the the best possible, unless someone find a way to to a better balance between I and P frames.

I think the quality is pretty good. maybe I am too much pixel peeper to want more improvements...

When the video is in playback the eyes do not perceive that the P frames are lower quality compared to the I frames.

Most people will not guess that this was shoot in 720p

Download the original rendered video from vimeo to perceive the full quality (h264 300Mbps gop1)

@apefos

Also when it was IB frames in 720p the size of I frames was similar to the I frames in the IP frames configuration. using P frames improved the quality because the P frames are bigger than B frames in the 720p gop3.

P-frames are generally larger than B-frames not because they are higher quality, but rather because they are encoded much less efficiently. The upper limit on image quality, however, is set by the size of the I-frames.

Ultimately , reliability must be assessed with extensive testing, rather than technical debate. Testing 720p mode on the GH2 is complex because the many auto-exposure and auto-focus lens options available can degrade the camera processor's real-time performance. My conclusion was that bulletproof reliably could not be guaranteed under all shooting circumstances in 720p mode with bitrates higher than 60Mbps, and that's what I designed the 720p H-mode to do in Flow Motion v2. For virtually all manual lenses and camera settings, however, two years of testing experience has shown that the GH2 can record and playback reliably at 100Mbps in 720p SH mode.

@LPowell

So lets change the debate from words to images, to original mts video files:

here is the link for the same scene shoot at iso 400 and same camera adjusts, workhorse vs. flowmotion, original mts files:

https://www.dropbox.com/sh/3erhko4jp1kevxm/AAAYZb_8gm58ZioUmf1OAMqda?dl=0

in playback full screen, for my eyes, the workhorse has a finer grain in the dark areas, compare the green sheet, the back of the radio, and the dark left lower corner with the blue plastic and it's shadow.

also, in timeline, frame by frame, the frames of workhorse shows more frames with full texture, and less compression artifacts in the frames without full texture. I do not believe that a B frame can be better than a P frame, the original mts videos proves this.

the streamparser screenshoots (see below) shows higher datarate in the SH mode of workhorse than in the SH mode of flowmotion, and the I and P frames of workhorse are much bigger than the I, P and B frames of flowmotion. also the ration between I / P is better in the workhorse: 1.79 in workhorse vs 2.09 in flowmotion, a better balance in the datarate distribution among the frames. and this was achieved with great stability and reliability for recording up to 5 - 8 minutes continously.

when workhorse stops recording after 5m32s in the torture test the camera does not freeze and saves the file. I recorded 5m32s of the death charts movie continously with both manual lens and 14-42 kit lens with image stabilization enabled. the SH mode is for manual lenses, the H mode is for native lenses with eletronic contacts. in the real world it can record the full 4GB file because real world images does not stress the codec so much as the death charts does.

I think that the GH2 hack is not about a debate to prove which patch is better, I believe more in a colaborative work to find the best settings. Due to this I started with the flowmotion because I liked the matrices and tables result. my work was to improve it. the workhorse is not a new patch, it is a modified flowmotion patch. I think we can improve it even more if you do the calibration balance to increase the P frames size a little bit compared to the I frames size.

THE ORIGINAL FLOWMOTION HAS A WEAKNESS, the P frames size are small in the 720p modes. The P frames size cannot be at same size of B frames, they must be bigger to justify it's presence in the codec (see the streamparser screenshoots). The workhorse imroves this increasing the datarate and using gop3 to deliver a finer random noise/grain, because more I frames means more frames with full texture and in playback this avoids the eye to perceive the difference between I and P frames.

There is the need to increase the P frames a little in the 720p modes of flowmotion to make it better. This tweak can be used in the workhorse also to increase the P frames size in the 720p a little. It will benefits all the GH2 comunity.

I have two guesses about this issue in the original flowmotion: the matrices for 720p P frames are not so good in the scalling tables and/or in the deblocking tables. the gop related weight distribution is not doing a perfect datarate distribution among the I, P and B frames. maybe both things, maybe other things...

To my eyes, in playback, the workhorse is already excelent quality, but if we can improve it even more, why not? problem is: I do not know how to do, how to increase the size of P frames without change the size of I frames, or increase the size of P frames and lower a little the size of I frames (-11% / +15%)

So @LPowell please consider doing this.

My job is finished now!

I don't know what you did to produce that SH_original_flowmotion.jpg, but it looks like no Flow Motion v2 chart I've ever seen over the past two years. If you think you've found something that works better for your purposes, that's fine and I'm happy to have laid the foundation for your exploration of GH2 patches.

However, your choice of a 3-frame IPP format is something I could only describe as idiosyncratic. It flies against the conclusions of virtually every dsp engineer who has worked in detail with H.264 encoding, and is a notion I discarded long ago after conducting my own evaluations. While I greatly appreciated the efforts of the many Personal-View posters who helped test the reliability of Flow Motion v2, its development was not a collaborative project, it was a complex challenge in multi-dimensional optimization. It was not something that could be done by coming up with a bunch of different versions and seeing which one worked best, there would've been far too many combinations of patch settings to even attempt to test them all. What you're proposing is just one example of many potential paths through the maze.

As you've not doubt discovered, the GH2's 720p mode cannot be optimized in isolation without considering its side-effects on the other video modes whose settings interact with it. At the time I developed Flow Motion v2, Panasonic had just upgraded the GH2 with 1080p HBR mode and the primary concern of most PAL shooters was whether 1080p25 HBR mode could be tweaked to comparable levels of image quality as1080p24 24H mode. My own second highest priority was to preserve 4GB file-spanning in all video modes at as high a bitrate as possible, which I had no doubt would require 95MB/sec SD cards. Once I determined that 100Mbps was a practical bitrate target, I set out to determine what it would take to bring 720p modes up to that level as well. I was unwilling, however, to allow 720p optimization to take priority over the challenging requirements of 1080p HBR mode.

What you've done may well have improved 720p mode in your eyes, but the way it degrades 1080p HBR mode is in no way acceptable to me. Flow Motion v2.02 was a high wire balancing act performed at 100Mbps and your approach falls inherently short of that goal. I appreciate your doing the tests yourself and confirming my own conclusion - that there was no way to beat its 720p performance and reliability at 100Mbps, and yet still maintain the integrity of the GH2's 1080p modes at that bitrate. However, I'm also well aware of the technical compromises it took to meet that goal. That is why I put even more effort into making sure that FM2's 60Mbps 720p H mode works reliably in all types of shooting conditions.

The 4 main workhorse patches here, as of september-03-2014, txt file about the versions inside the zip file:

I DID IT !!!

And I did it exactly as I wanted to do. I can say that the noise is almost indistinguishable comparing the I and P frames in 720p60

I learned how to distribute the datarate among the frames, not everything, but enough to make it.

I think I will call it "Boson Patch - the Arri Alexa inside a GH2"

I would like to thank @mozes

I will not keep posting different patches, I will fine tune it first.

It seems that the collaborative work about patches was forgotten in the past...

First impressions in 720p60 SH:

the datarate increased in low light shoot. the codec is using all available datarate

the trembling which is comon in gh2 low light dark areas is lower because the noise is finer, because the frames are more equal, and because there is less compression artifacts.

very similar I and P frames, the ration is lower between them

noise is less digital looking, fine, random, looks like film

this thing deserves a DCP output for the bigscreen!

here it is some stream parser screen shots, WorkHorse vs BOSON

the first setj.ini file is experimental, but if you area anxious to try, it is here

also two mts files, workhorse and boson, for comparison... see in playback and see frame by frame,

after some seconds the codec gets a better balance in the frames, so see frame by frame after some seconds in the file. (this is the first patch, not tweaked yet)

https://www.dropbox.com/sh/yruwiggwkz6763q/AADQh1NtUJrFnQEnqmbra4wwa?dl=0

There are two gop tables for 720p60:

720p60 Opt1 GOP Table=2, 12, 0, 0, 36, 0

720p60 Opt2 GOP Table=2, 36, 0, 12, 0, 0

when I use the Opt1 and Opt2, the Boson behavior happens in the SH but not in H

when I use the Opt2 in both slots, the Boson behavior happens in the SH but not in H, same thing

when I use Opt1 in both slots, the Boson behavior does not happen...

I believe that the mistery is in these GOP Tables, I think there will be settings here which will make the Boson behavior to happens in both SH and H slots.

The datarate in SH is set to 104000000 and results in 76Mbps with the Boson or WorkHorse behavior

The datarate in H is set to 64000000 and results in 64Mbps with the WorkHorse behavior