Messages

Since Danne's suggestion that I should use separate card readers for MLVFS using card spanning, I've been doing every video in at least 3000 pixels across, and having zero problems!

Lately I have gotten lazy (and impatient) so I have switched from "superscaling" in resolve (GPU intensive and slower, but with beautiful results) to simple upscaling using the SINC algorithm.  The issue with sinc is how it strongly emphasizes any aliasing in the source footage (very badly, actually to the point of unusability).  But it seems that this only happens when the source footage is below a certain resolution, for I have never seen it at all in these latest videos.

This video was shot at 3072*1536 (I really like 2:1 aspect ratio) and upscaled to 4k using simple SINC.  To me it looks like it was shot at 4k. 

speaking of 4k, make sure to view it at 4k to avoid youtube quality degradation...

https://youtu.be/vNrEmilbvdw

I can't praise MLVFS highly enough.  To pull the cards from the camera and have a completed edited graded video uploading minutes later is pretty amazing!

PS: the banner image for the video is a frame from the video, cropped 200%

the peach blossoms look astonishing, just like real life!

5d3 3.5k raw - mlvfs - davinci resolve - cuda hevc 10 bit 444

1 hour from start to finish.

https://youtu.be/YgpW2Z5Z-Tk

Since danne's suggestion to use 2 card readers, I've done several videos using card spanning with mlvfs for instant grading/editing in resolve, with absolutely no issues!

the latest example:
https://youtu.be/o9w99l9zFgc

workflow:  stick cards into readers, create 32MB ntfs ramdisk, run "mlvlinks" script, run mlvfs script, jump into resolve!

see this thread for details:
https://www.magiclantern.fm/forum/index.php?topic=25473.msg243387;topicseen#msg243387

maybe we will have a new card to test?

https://www.anandtech.com/show/18809/samsung-updates-pro-plus-uhsi-memory-cards-with-improved-speeds

please post the updates on the forum!
i'm sure i'm not the only one who enjoyed reading your first update...

for years i've been considering buying a 5d4 just to start developing it, but the ridiculous amount of hassles involved just to get set up (sorry guys, it's true) have been a complete showstopper for me.  knowing that someone is trying to make it less hostile is very encouraging!

i think it could be a fun project, but when all the fun is removed before step 1, no thanks...

race to the bottom!
today it seems to be the standard to cut every corner possible. 

i tried to update firmware on the reader and found nothing.

it looks like every product that matters needs its own "magic lantern" 

maybe gpt-5 will be capable of performing reverse engineering, but only after it has been jailbroken and trained by the mod community! 

i only have the onboard reader apart from the kingston device, and it is very slow, but i gave it a try, and it worked nicely.

great suggestion, danne!

i never even bothered to try the onboard reader, so i guess i should have been more optimistic

I shot another video in the forest, this time at the highest resolution i could get: 3664x1534

i was hoping that updating dokan solved the problems with mlvfs, but it failed even worse now, with crashes of mlvfs and/or dokan resulting in the virtual drive disappearing (as well as the command prompt window for mlvfs).

i tried updating to the very latest version of dokan, and renaming the resulting dll so it would work with mlvfs.  it worked, but didn't improve the situation.  in fact, it made it worse, by adding corrupted frames back into the mix...

in the end i got a perfect result by reverting to the last 1.x version of dokan, and then copying all the *.m000 files to the CF card

so it seems to be related to reading both cards simultaneously.  whether it is a driver issue, dokan, and/or mlvfs is unknown.  it could even be related to the actual card reader (kingston).  who knows...

hopefully others will explore this, particularly now that sd overclocking is so reliable.

again, in the end it worked perfectly by copying the footage files from the sd card onto the cf card (in the same directory)

it was simply amazing to shoot continuously at this resolution with no problems!

check out the exquisite detail of the footage.

watch in 4k!

https://youtu.be/RN4WaEOJonA

windows doesn't have fuse.  dokan is the alternative.  look at the first post on this thread for details.

the dokan link provided there is from 2014 and is a much older version.

To windows users:
I ran into a problem with very high resolution footage using card spanning, with frames being corrupted.
It seems to be solved by updating to dokan v 1.5.1.1000

for my latest 5d3 video i decided to go for the highest possible continuous recording resolution (over 3500 pixels across) for maximum detail.

i was surprised and dismayed to find that at this resolution mlvfs makes errors every time i render in resolve.  i've done over a hundred long form videos using mlvfs and resolve at 1920 superscaled to 4k and it has never dropped a frame ever until now.

there were problems in the past, but they disappeared as soon as i switched from windows 7 to 10.

it most likely is because of timing difficulties with card spanning.  the CF card is much faster than the SD, even though it is supposed to be fast.  (sandisk 170)

maybe it would be improved with using 2 card readers on separate buses.

but it should be more robust than this.  it seems to be always on the edge of failure...

the glitch happens in the opening shot when i begin panning right:

https://www.youtube.com/watch?v=jJLPWZG8-ww

on a happier note, check out how absolutely beautiful the video looks at 4k!

it's a shame it suffers from the frame corruption because otherwise this is dreamlike in the convenience of pulling the cards from the camera, throwing them into the card reader and going straight to grading and editing and rendering in resolve...

i came in from the forest, and had a completed rendered copy ready to upload in under an hour, including recording the dialogue.

EDIT: upgrading to the last 1.x version of dokan seems to have solved the issue... v 1.5.1.1000

i re-rendered the video twice and both times it was flawless.

i'll do more videos at this resolution and hopefully it keeps working perfectly.

thanks for pointing out my error.

that was the wrong timestamp!  that's the 3616 shot, so it should look amazing

the correct time for the ugly shot is 8:50...  i'll change the original post.

EDIT:
An additional note regarding the 5d3:  since it is now capable of such high speed data writing, 1920 probably wouldn't require data compression, freeing up CPU resources for resampling.

---------

I intended to shoot this video at 3616x1536, but there was a problem, even with card spanning and sd 240mhz overclock...

because of the bare branches on all the trees in the forest, the scene is just too complicated for that resolution, even when underexposed!

i had already shot the "beauty shot" for the middle of the video of the baby chestnut in the morning sunlight at that resolution, but when i got immediate sudden stops repeatedly in the forest i decided to shoot this talk at 1920 for reliability, using 3x3 mode since i was shooting at f/2.8 so there would be no problem with aliasing and moire on all the branches of the trees.

i wanted maximum detail on the closing shot of the cloudy sunset through the trees, so i shot that using 1x1 mode.

i was really surprised when i was editing the video to find that there was so much aliasing on the 1x1 shot!

to see the 1x1 footage, skip to 8:50 and pause it before the messages pop up for other videos.  the problem is very clear in the upper left corner of the screen.

i used 4x supersampling in resolve with gaussian sampling to minimize the problem, but it is still bad.

seeing this made me wonder if it is possible to do better resampling (even simple bilinear) of the full sensor down to 1920 instead of the horrible binning.  i know that it is beyond the capabilities of the cpu in most of these cameras, but has anyone done any experiments with this at all?  maybe certain dual cpu cameras could do it?  it makes me wonder if this could be done well in the very recent crop sensor models that have a newer (faster?) cpu.

it's just a thought...

on another note, if anyone is interested i could share a full sensor resolution still of the extremely challenging forest shot to use for a "stress test" of card overclocking.  if you print it out at full resolution, it could be useful as a "standard" torture test.  there was no wind at all in the forest and the camera was on a locked tripod, so a still should be just as difficult, especially with perfect exposure.

view the video at 4k, of course.
https://youtu.be/A8elVCzKfr4

here's those frames:

https://drive.google.com/file/d/1qnjV-fG1Z_hIuiOCEBSgyd663cx2GG-g/view?usp=share_link

the frame with the sunlit cut log showcases how well mlvapp handles highlight reconstruction when only one channel is clipped.  i intentionally clipped the red channel on this shot to get enough of the dark scene at the beginning of the shot.

all that is required to correct it in mlvapp is to enable highlight reconstruction and then pull back the highlights just enough to get red of the clipping on the red channel in the histogram, and suddenly you can see the nice detail in the sunlit surface of the cut log, with no pink highlights.

it was a lot harder to do this in resolve!

with the renewed debate about 3k, i decided to shoot my latest forest video in 3k, utilizing resolve's superscaling algorithm to bring it up to 4k using 4X upscaling:

must view at 4k to see it correctly!  check out the incredible smoothness of the razor sharp spider webs that are clipping in the sunlight...

https://youtu.be/UA_o5qI5iqQ

ps: i'll send some frames to ml700d

my internet connection is quite slow here in the forest so i can send you some frames if that is ok for you?
it sometimes takes a day to upload a compressed hevc to youtube...

I used resolve for grading.  i use MLVFS for instant grading.  put the card into the reader, and a few clicks later it's ready to edit and grade!  Outputting to NV-HEVC is ridiculously fast.  in this case, from camera to uploading graded video in a few minutes...

the lens used is sigma 105mm F/2.8 IS.
300mm focal distance for 1:1

focal length was 12 inches.

i actually like a bit of movement when doing a focus pull, plus the branches are moving in the wind...

most of all, i'm just too lazy to do software stabilization! 

IS was a lifesaver for a handheld focus pull at maximum closeup on the tiny redwood bud!

watch in 4k to bypass youtube quality degradation...

https://www.youtube.com/watch?v=N_2dI7dSzEc

Everything's beautiful when filmed with magic lantern!

Thanks to you wonderful people for all you do:  keeping these old cameras going...

The winter that wasn't...

Long point, Ontario - February 20, 2023
View in 4K for original quality.

https://www.youtube.com/watch?v=GtkoZhF2P5c

update (feb 2023) : Dokan version 1.0.3.1000 is recommended

mine is 1.0.3.1000

i hope that fixes it for you... fingers crossed!

@TG:
I am using win10x64 with mlvfs and it is rock solid.  i have it packaged in the file i linked in the opening post of this thread:
https://www.magiclantern.fm/forum/index.php?topic=25473.msg232499#msg232499

are we getting close to a new mlv app official version?
i'd like to try it again comparing to resolve, and any extra speed for the windows version would be most welcomed!