Menu

Show posts

This section allows you to view all posts made by this member. Note that you can only see posts made in areas you currently have access to.

Show posts Menu

Messages - ilia3101

#1
I miss the old design :(

But profile pictures work for the first time since like 2018  8)

Thanks!
#2
It's only a small fully connected neural network, with 3 inputs and 3 outputs (RGB channels).

@masc4ii correction: I think it would be possible to convert them to LUTs with a little bit of code.

I think a 64x64x64 LUT would be accurate  enough.
#3
Much more effective than dithering would be log encoding. I've dreamed about adding it to the MLV format for years.

Of course it would have to be done in transcoding, not possible to record that way in camera.

10-bit would be indistingiushable from 14-bit linear (maybe even 9-bit would be enough with how noisy these cameras are). Would be great for storage combined with lossless compression.

Edit: actually, I think custom encoding tables are already possible in DNG. Might be possible to add it to MLV App.
#4
Quote from: aiyik on June 02, 2023, 02:03:08 PM
MLVApp is the culprit because it does have lens correction profiles (CA + distortion + vignetting), which are automatically applied, recognized and cannot be toggled off. If you want to deactivate those you'd have to grade using CinemaDNGs inside RawTherapee/Lightroom/CameraRAW/etc.

None of this is true. There's no correction that can't be toggled off in MLV App. And there's no distortion capability in the app at all.

Quote from: aiyik on June 02, 2023, 02:03:08 PM
Besides, any decent CCTV or industrial lens is on par if not superior to the EF-M kit lenses (uncorrected).

Now that you know, feel free to demonstrate this using MLV App :)
#5
There's loads of papers about neural network demosaicing if you search on google scholar.

It's a waste of time unless you're some kind of scientist imo.
#6
Quote from: yourboylloyd on April 18, 2023, 09:09:21 AM
Always. You get the exact same dynamic range. 10bit vs 12bit vs 14bit is all the same range on these sensors.

Nah 10 bit is definitely a lot worse. But I agree that 12 bit is indistinguishable from 14.

Quote from: vastunghia on April 18, 2023, 10:17:52 AM
Latest test I've stumbled upon in this forum:

Interesting find. Never saw that much difference between 12 and 14 bit when I've tested it (both on 5D2 and 5D3). Perhaps it depends on recording mode/camera. May also be the black level, raising it in MLV App might reveal identical results between 12/14.
#7
Quote from: vastunghia on April 20, 2023, 09:04:53 AM
But what does this mean in terms of the transformation being applied to the RGB space? LUTs are easy to understand. But what do filters do? For instance, do they apply a continuous function instead of interpolating between discrete points defined in some RGB lattice? Do you have any documentation on this?

Exactly this. They are a continous function (a tiny network trained on film convert before/afters), unlike LUTs... not that it makes any visual difference.

It's basically a LUT with extra steps. It was a silly experiment. But they are nice anyway.
#8
If you look past "wow the noise is gone" you'll notice it looks like smeary computer shit.

Don't fall for the youtube tech hype.
#9
General Chat / Re: Anyone see Zeek lately?
April 05, 2023, 12:24:56 PM
I last had a direct message from him in July, said he's busy doing other stuff, so maybe (I hope) he's doing alright.
#10
Quote from: Mattia on April 03, 2023, 11:13:14 AM
That's for sure the main advantage over the M, other than being much more responsive, in my opinion.

Oh you'll see about reponsiveness ;)
#11
I've had early access on my EOS M. It's 100% real. WYSIWYG.

But there's more...

This feels like magic lantern did in 2016, the menus are snappy, doesn't slow down the camera, and there's almost no bugs/crashes (this actually applies to the 5D3 as well ;))

This is the best Magic Lantern we've ever had.

Bilal is doing amazing work. I can't thank him enough.
#12
Wait... FPGA? Could it do any image processing? Like log encoding raw video?

Do other cameras have one?
#13
@dpjpandone

Vertical stripes are like a different ISO in each column (multiplicative), same pattern repeats every 8 columns.

Fixed pattern noise is an offset (additive) - specific to each pixel, not repetitive or preditable.

Algorithm for one will not really fix the other.
#14
General Development / Re: The MLV format
November 22, 2022, 12:51:14 PM
Just made the connection...

Header files, at least interface definitions and structures, are not copyrightable, according to US copyright law (Google vs Oracle: as it stands, Google won), and the FSF itself (authors of GPL license).

Found this great answer too: https://softwareengineering.stackexchange.com/a/216480

So any system, even closed source, is safe to implement MLV!



Not that it matters anymore, with BRAW and ProRes raw being so popular. Apple/Blackmagic taking responsibility for patent issues makes those formats very appealing to manufacturers (on top of the great lossy compression)
#15
I want to darken my H264 proxies by roughly -1.0 stops, as I'm doing ETTR, but it seems ML digital ISO is broken on all crop_rec_4k branches. Only positive gain works. Negative gain does nothing.

Any idea why this is happening?

EDIT:

I am almost certain it's commit - https://foss.heptapod.net/magic-lantern/magic-lantern/-/commit/5c4e07340c534817def631511418f902c5cbf25a#26a8097d81e418f83d05a9b16c85b054342237ba

Impossible to revert unfortunately.

QuoteImage fine-tuning -> ML Digital ISO now goes up to 7 stops :)

Ok Alex. But working negative gain was more useful.
#16
Quote from: Kharak on October 17, 2022, 05:34:37 PM
I cant speak for mlv app, but working with dng in resolve, any clip with log conversion applied with the sun directly in shot will most likely give you "black sun" (pink) artifact.

Ah yes the "black sun" can show up as pink, I forgot! I see that all the time. Thanks for reminding me.


Quote from: tupp on October 19, 2022, 12:12:38 AM
However, would the recently added Sobatka AgX trasnsform help in reducing this clipping/black-sun problem?

It could help hide it, if it's pink, thanks to it's stronger film-ish highlight desaturation. But it won't truly fix it.

Black/pink sun needs to be handled at the raw correction stage.


Quote from: tupp on October 19, 2022, 12:12:38 AM
Troy Sobatka also made a "Filmic" Blender module, which has been adapted to Darktable, and which seems to deal well with clipping and tonal shifting in the highlights:

Yes, filmic was created a few years ago. AgX is basically Troy's replacement for filmic, it's better. It might get added to Blender soon as well.

And while darktable filmic is "inspired" by Troy's filmic, it works totally differently, and the aesthetic is different.


Quote from: tupp on October 19, 2022, 12:12:38 AM
It would be great if MLV App featured a similarly versatile "filmic" module!

It has so many parameters that don't make any meaningful change and just cause weird tonality distortions that don't look good whichever way you pull the slider. I don't love it.

I implemented AgX in MLV App because it was super convenient. The source code is a spaghetti house of cards and can't take much.
#17
Oh interesting. Thanks for the FRSP idea.

But what about simple MLV videos? It seems pretty uncommon there, at least on 5D3.


QuoteAFAIK firmware will do black sun surpression on its own.

😢😢
#18
Just realised the black sun effect could be used as information to extend dynamic range in processing. Would need a fancy new algorithm tho.

Does anyone know how to increase likelihood of black sun effect without using dual ISO?
#19
QuoteI was also comparing it with the dngs imported on photoshop camera raw

Oh ok. I wouldn't even claim MLV App matches the exposure correctly. The 1.2 stops is just an absolute guess. Adobe camera raw probably knows better about this stuff (at least on CR2s)... although in the case of DNGs generated from MLV, the exposure offset metadata may be missing, so it's possible camera raw doesn't do the 'hidden exposure' with those. Just guessing.

If you have any interest in doing it, I'd love to see a comparison between:
1. CR2 photo through camera raw
2. MLV video through MLV App
3. A DNG created from that MLV, through camera raw

Where the MLV and the CR2 use the same aperture, ISO and shutter settings. ISO 100/200/400/800, not the inbetween steps. I'd do it myself if I had Adobe software.

My (likely wrong) prediction: 1 and 2 exposure will match, but 3 will be darker.
#20
MLV is noisier than JPEG mainly because it's lower resolution with the same noise level. Signal to noise ratio ends up worse at viewing scale. And yeah, also the fact MLV App doesn't do chroma denoising by default.

Do you also want to know the values of the hidden camera matrix? It distorts all the hues if you think about it, and it means the camera's RGB channels do not correspond to the RGB channels that are output. But it's there for colour management. Without it you'd get pale colours and wrong hues.

I think of this 'hidden exposure' gain as another (admittedly less critical) form of camera calibration, just like the camera matrix.

And I still don't believe 'no exposure' is meaningful, as you must first make a choice about how you anchor the white balanced channels, which in itself, is an exposure choice. MLV App makes this choice by matching the luminance of before and after (which will produce some pink highlighs at 0 exposure, as the lowest clipping channel will end up below 1.0), but you could also make it such that the earliest clipping channel = 1.0 (this way they all clip at or above 1.0, so 0 exposure will not produce pink highlights, probably the best option for this), or you could make it so the highest clipping channel = 1 to preserve the most data, but then you'd get very visible pink highlights at 0 exposure.

Everything is relative. And there is no image in the raw file, it must be created.
#21
Quotebut I do think that it woudl be better to have the exposure boost removes in the rec preset, at least to understand what's going on

Disagree. Still cameras like our Canons usually leave 1 or 2 stops of highlight headroom, so 18% on the sensor is not meant to be 18% (middle grey) on the final image. That would leave no room for any kind of smooth roll-off curve.

This headroom also ends up being different on each channel, smallest on green, highest on red/blue, depending on the white balance, so which channel would you use as a reference point? They all get boosted/reduced during white balance, you can't really have no exposure... So why not just place middle grey sorta where the manufacturer intends!

Cine cameras like the Arri Alexa (and even the sigma FP at some ISO settings) leave 7+ stops of headroom above middle grey. That requires a masive amount of 'hidden exposure' during processing to match what the operator saw on the screen, or... to realise this isn't 'hidden exposure'. It's how the camera is engineered.
#22
Thanks @masc.

So the exposure offset is 1.2

Quote
Well just to have a preset by default with a "clea" option, nothing fancy. I like the loo of the tone mapped preset though.

Then you could try setting the following:
- rec709 preset (it does nothing aesthetic, it simply encodes with the rec709 transfer function, which video players should decode with)
- exposure -1.2
- disable creative adjustments
- as a bonus: you can lock the white balance to 6504K to get the most absolute colorimetry

It would be:
- techincally the most 'colour accurate' output you can get (except where it clips), in a silly absolute way
- hideously ugly
- a very inefficient encoding in terms of preserving data (unlike log ;))
- you'd still lose data by clipping due to the following: 1. white balance 2. the limited rec709 primaries
#23
Yeah I thought MLV App has an exposure factor of +1 or 2 by default. But I couldn't find it in the code just now, so I don't know anymore, could it be in the interface??

Why do you feel a need to undo this 'hidden exposure'? You could undo this gain by setting exposure to -1 or whatever, but... there's also white balance, which in itself increases blue and red by about +1, is that not hidden exposure?... do you want to undo that? What's your definition of 'no hidden exposure'? There's also highlight reconstruction, which (unpredictably) extends the dynamic range of your data by a couple of stops (above the cameras max white level), so... are those considered a part of the dynamic range? Should you decrease exposure to fit those in? I think it's useful sometimes, so maybe. Plus the camera matrix distorts values, and the resulting RGB values are often much higher than the camera's white level, or even negative 😱

Transfer function is a better word for gamma, yes. But there's nothing stoppping you from adding an exposure multiplier in to the expression.
#24
Sorry I don't get notified.

Raw2mlv is giving you as many pixels as possible (without exposing the raw file's black borders), according to information provided by libraw... which might still be underreporting the actual size of the image area, but bettwr safe than sorry.

The official 5760x3840 size is slightly cropped in.

So yes, there are spare pixels.

You can't resample raw files.
#25
Share Your Videos / Re: About Life and Its Journey
August 27, 2022, 01:52:26 PM
Exposure is always allowed, that's not just a creative adjustment, it's like changing your ISO. Sharpness is also allowed, that's just signal processing, not colour related.

Contrast, yeah, that's locked. I'll explain why...

QuoteThanks! I thought if you tick "creative adjustments" it will break log curve, but..

Yes, it breaks the log curve, specifically the dark strength adjustment is the one responsible by default, but any colour related slider will break the log encoding - If you want to export a pure log encoding from MLV App to grade with other tools, you should disable creative adjustments.

However if you want to grade log within MLV App, or use a manufacturer's LUT to convert their log to rec709 in MLV App, and then grade on top of that, you should allow creative adjustments. You don't need to care about breaking the log curve in this case.


But I think MLV App's new AgX is going to do a better job than manufacturers' log to 709 LUTs anyway ;)

And I recommend combining that with MLV App's filters, especially the first three options: