Messages

Vertical, but they're basically gone unless you are digging in the shadows, but the noise there is just as strong.

For "easy" lighting conditions, this is fantastic now.  I just really need 24fps (or 23.976) at 2.35:1 working without problems...

For me, only the 3240 versions are stable.  Extremely bright areas get that interlacing failure on the 23.976 FPS versions in almost every instance.

But the 3240 versions are very nice.

The stripes are still there but now at the same level as the noise floor, only visible if pushed really hard.

Here's a sample MLV (1 frame) of a good test.  Raise the exposure by 2.5 stops and drop the shadow strength to 8 and you're in the right area.

That light bulb is very unique.  It is a point source with a cylindrical lens, so it's almost as bright as the sun but it puts out very little light.  The scene is actually quite dark.

https://drive.google.com/file/d/1F1n39758QmkMqapsJLWnogDTljBEGw51/view?usp=drivesdk

Here's a video with the same bulb in an even darker scene I shot using the ISO experiments.  The artifacting is thanks to YouTube.

https://youtu.be/Dcj16103JF8

I'm shooting a video on Sunday.  I'd love to try this dual ISO if it can be stable at 24 FPS at 2.35:1

I gave it a quick test and it looks pretty clean at 22 FPS.  Nasty unusable stripes at default FPS on my camera.

One thing that would be very helpful for me (apart from receipt slots!!!) is scale control for the video scopes.  It would be great to be able to crank it way up to see true cutoff of highlights and shadows when necessary.

At least, I think so

I just wanted to chime in to say I'm also very happy with how mlvapp is developing!

I had just finished the final grade of a video I filmed a few months ago in resolve, and I still felt it could be better, so I tried again in mlvapp, and I'm totally happy with the results.

If you can trust your eyes while grading, you can do almost anything with mlvapp.  There are definitely "better" tools out there but that doesn't mean you can't get just as good results from this great program!

Keep it going!

Thanks for all your efforts!

They still only work for me at 22fps, but they're the least stripey yet.

The dark areas of the paintings showcase how much you lose with only 10 bits...





Edit: added another ISO 109 shot.  Used the same 35mm prime so I had to move the camera to get approximately the same framing.  Sorry, but 109 still is best for me, especially considering it's 3072 pix across.

But I was reminded of something that might be a clue about the stripes...

In the amplifier hacking mode, any settings change, including 1x/5x/10x liveview mode, requires waiting about 30 seconds for the amplifiers to settle, and during that time, the screen is full of vertical stripes that slowly fade away.

Maybe this is somehow related?

I don't know causes, these are only my observations.

You're right, that fixed the aliasing.  I never pay attention to that function...

22 FPS

Fresh identical settings comparison of 10 and 14 bit versions...

The brightness difference is not my doing.
Interestingly, there's more aliasing in the 14 bit version.  Look closely at the modulation lever on the keyboard in the background.

Here's the two latest 10 bit versions stress



tested at 22fps override:

V2 definitely has less stripes.  Precisely the same settings and conditions, straight from mlvapp.

Good luck!
I hope you succeed, I'm starting to like this!

10 bit version:

I tried the new version and it's the same.

I hope you can solve this!

Simple test: expose for a light bulb in a dimly lit room just like the scene I did.  You'll know right away!

The new version fixed the interlacing issue, but the vertical stripes are very strong, especially if underexposing for highlights.

I did many exposures of this test scene, and even when I exposed exactly for midtones, the stripes are visible on the wall well into the mids...

As for dynamic range, it doesn't beat the ISO experiments yet.  Pretty much the same.

But, if you can eliminate the stripes, it could be a winner!

I still dream of both combined.

The following images are straight from mlvapp with no processing except exposure and shadow strength to match.  No noise reduction nor chroma smoothing applied.

The ISO 109 was shot at 3K and exported at 1080.  BTW, the 3K footage uses less data...

Different framing due to different crop sizes.

It was the shutter speed.  Seems like it was somehow set in your version.

Once I set it to 50 in the ml exposure menu, the problem disappeared!

For some reason, when I set FPS override, the camera crashed hard. Display stays on even with the battery door open.  Battery has to be pulled.

This version provides very nice results!  Maybe better than ISO experiments!

I'll do a comparison with this.

EDIT: I can't repeat this.  It must have been a stray neutrino.  But the hard crashing is definitely repeatable :p

Much better!

But something strange is happening now.  When I load a clip from the "less stripes" version into mlvapp, as soon as I switch to anything except bilinear demosaicing, the brights lose the dual ISO processing.

I checked by reverting to the prior version and it's fine.

I'm using the 14 bit version.

I tried your 14 bit version and it works very well.  Straightforward and easy!

Unfortunately (or fortunately, depending on your point of view) it doesn't seem to provide any benefits compared to the ISO experiments, for me at least...

I will keep watching and playing with any new developments you make, and hopefully you will work some magic!

I think that dual ISO combined with the ISO experiments is the way to get truly incredible results...

If only more people were playing with the ISO experiments, maybe this could happen.

I'll try some more tests with your 14 bit version and see if it beats the ISO experiments in some cases.  The shadows are just so clean in the ISO experiments...

I hope you don't think I'm discouraging you.  Quite the opposite!  Keep up the amazing work!

What an excellent idea!

This would be amazing for interesting effects!

If you already captured all this data, it's best to keep it, i.e. rather than exporting one frame out of 25, I believe it's better to average all of them in a way that minimizes temporal aliasing.

Somewhat like this: http://tessive.com/the-time-filter

I have yet to experiment with this, but any of these should give much better results than keeping just one frame out of 25:
- 180-degree averaging (i.e. average half of the frames in each group)
- use variable weights when averaging, such as a Gaussian bell curve

e.g.

- 180-degree averaging: weights = [  0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 ] / 13  # i.e. average 13 frames out of 25
- "Gaussian" weighting (just an example): [ 0.001 0.002 0.004 0.008 0.014 0.023 0.034 0.047 0.061 0.075 0.087 0.095 0.098 0.095 0.087 0.075 0.061 0.047 0.034 0.023 0.014 0.008 0.004 0.002 0.001 ]



To get the above "Gaussian" weights in octave:

Code Select Expand


w0 = [  0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 ] / 5;
w1 = conv(w0,w0)(25-12:25+12);
w2 = conv(w1,w1)(25-12:25+12);
w3 = conv(w2,w2)(25-12:25+12);
printf("%.3f ", w3);


Of course, the processing time will be much higher.

I hope to be able to perform this kind averaging on the camera's image processor at some point...

@70MM13

Cool!

It's nothing short of amazing, what we can do with these old cameras thanks to the efforts of the magic lantern community.

It just keeps getting better!

Exciting times.

Here's a sample frame from a test shoot for an upcoming video I'm working on.

It is straight from mlvapp, showing that you can indeed do excellent grading with mlvapp.

I'm using the ISO experiments with my 5D3, dropping ISO 200 to 109.  No dual ISO used.  The foreground is full shadow, and the background is all full sun.  Beautiful dynamic range.

Mlvapp rocks!

WOW!!!!
Thank you so much!
You rock.

I'm sure there are plenty of people like myself with great appreciation for mlvapp!

Yes, resolve can do more for very poor footage, but it's a different beast altogether.

Mlvapp is GREAT.  I have no problems getting fantastic results with it!

Keep up the amazing work!

Many thanks!

No worries on the jpeg 2000...  It was just hopeful thinking.

Multiple slots for receipts would be far more important!!!

Thanks!

I'm amazed at the results obtainable with mlvapp's elegant and uncluttered controls!

One thing that would come in really handy is being able to store and retrieve receipts quickly from ram with keyboard shortcuts, similar to resolve's "memories".  They use the number keys, plus alt to store or Ctrl to load.

Something like this would be excellent for powerful workflow on grading clips.

Also, any chance of high colour depth jpeg 2000 MOV export?  It is a very nice format in resolve, and maybe it would work well with mlvapp!

Mlvapp is an awesome program!

It makes sense...  I can't see any degradation on the 16 bit tiff, so perhaps faststone is just incorrect.

Maybe there's another way to get a clearer answer...  Perhaps gimp has a similar colour count function.

PS: gimp has a colour count function, and according to it, everything is fine!

Test image was approximately 94,000 colours in the 8 bit PNG export, and 101,000 in 16 bit tiff.

In case anyone is interested in the function, it's under colours/info/colourcube analysis

Thanks for the help!

I use faststone image viewer to count the colours.  Ctrl-shift-T

Yes, the example values I gave are fairly low, but the scene I'm working on is extremely low light!

Congratulations on the new version!  Keep up the amazing work!

I hope you can solve my mystery!

@70mm13: very interesting. How do you count the colours? The frames are exported in very different ways:
- 8bit: internally convert 16 to 8 bits, use Qt library to save 8bit PNG
- 16bit: directly stream the (processed) rawdata (internally 16bit) to ffmpeg which saves the 16bit TIFF
So if your values are right, there should be a problem with ffmpeg (or our ffmpeg call).
BTW: only 100.000 colours? (3x)8bit has already 16.000.000!