SIGMA FP with ProRes RAW and BRAW !

[Llaasseerr]

2 hours ago, OleB said:

Ok so I went back to the drawing board.

If you monitor regular ProRes 4:2:2 8 Bit (non RAW) with the Ninja V, the ISO values and meter tools seem to work absolutely correct. Clipping occurs at 100 IRE and false colors. 

The change is happening with the switch to external RAW. If you monitor "native" the clipping points are as follows (approximately, have noted them down from the Ninja V waveform). Picture is very flat and log like. Black point for all ISO is at approx. 10 IRE.

ISO 100 - 65 IRE

ISO 200 - 75 IRE

ISO 400 - 80 IRE

ISO 800 - 90 IRE

ISO 1600 - 90 IRE

ISO 3200 - 85 IRE

ISO 6400 - 90 IRE

ISO 12800 - 90 IRE

ISO25600 - 90 IRE

False colors will show the ranges accordingly to those max values, but not scaled. So ISO 100 leaves you with white being in grey (59-77 range).

Rec709 mode and PQ modes meter as from ISO 800 and upwards as like I have described earlier.

Have read a little more about ProRes RAW. What I found out is that RAW is always linear HDR, no matter how you monitor it. So basically when you bring it in a Rec709 timeline you have to bring back the overshooting dynamic range. At least FCPX is not applying any conversation, it is importing it as linear RAW. You can however transform the very same into LOG (different manufacturer options) and then from that LOG to a certain LUT you like to use.

Maybe that is the explanation why it is best to monitor that linear RAW best in an HDR mode, as this will show you as much information as the monitor can show will reviewing. 

What I do not get however is why the ISO range is acting strangely in RAW mode when monitored PQ. That could be a bug no body liked to address at Sigma/Atomos, or it is because they had to do it like this for their internal RAW recording. I do not know. 

 

You are correct, ProRes raw is linear raw the same as DNG. It is no great mystery. In ACES or a similar linear gamma project, the linearity is maintained but the exposure is shifted to account for where middle grey is so that there are superbright values in floating point above 1.0.

In editing software besides Resolve, the ProRes raw file needs to be interpreted as log but it should still set it to the correct exposure. And then as you say, you apply a film print LUT to Rec709 or similar.

You are meant to display raw images in log because it shows the entire range, before adding the look LUT in either HDR or SDR space. This is the big problem with this camera - people who say it doesn't need log are forgetting about monitoring. There is always an intermediate transform of linear raw to a log format before applying the display LUT since LUTs only work in 0-1 space.

I spoke to Atomos, and they could not even say what the colour space is of the Sigma fp when monitoring native on the Ninja V, but I suspect it's somewhat like the OFF profile. This goes back to Sigma not documenting this. So it might be a kind of Rec709 space, but it seems that it's more flat than that. I also asked if there was a way to transform the native image into a known log format and they said no.

So I don't believe there's any value in writing down those IRE values, since they are only ever within the context of the gamma encoded space of the image display (even though the underlying image is linear) and no-one can say how the image is encoded.

That is why I asked, can you check false colour on the camera itself? If as others have said, it takes the values directly from the sensor. That might also give you some clues as to how to set it on the Ninja V. You can also create your own false colour LUT for the Ninja V.

 

 

 

 

 

 

[Llaasseerr]

13 minutes ago, Llaasseerr said:

You are correct, ProRes raw is linear raw the same as DNG. It is no great mystery. In ACES or a similar linear gamma project, the linearity is maintained but the exposure is shifted to account for where middle grey is so that there are superbright values in floating point above 1.0.

In editing software besides Resolve, the ProRes raw file needs to be interpreted as log but it should still set it to the correct exposure. And then as you say, you apply a film print LUT to Rec709 or similar.

You are meant to display raw images in log because it shows the entire range, before adding the look LUT in either HDR or SDR space. This is the big problem with this camera - people who say it doesn't need log are forgetting about monitoring. There is always an intermediate transform of linear raw to a log format before applying the display LUT since LUTs only work in 0-1 space.

I spoke to Atomos, and they could not even say what the colour space is of the Sigma fp when monitoring native on the Ninja V, but I suspect it's somewhat like the OFF profile. This goes back to Sigma not documenting this. So it might be a kind of Rec709 space, but it seems that it's more flat than that. I also asked if there was a way to transform the native image into a known log format and they said no.

So I don't believe there's any value in writing down those IRE values, since they are only ever within the context of the gamma encoded space of the image display (even though the underlying image is linear) and no-one can say how the image is encoded.

That is why I asked, can you check false colour on the camera itself? If as others have said, it takes the values directly from the sensor. That might also give you some clues as to how to set it on the Ninja V. You can also create your own false colour LUT for the Ninja V.

 

 

 

 

 

 

I should add that with the Ninja V it would be possible to do a transform to a log-ish image with a LUT, but I'm not sure if the highlights would be preserved. And then an additional LUT would be required to transform to a Rec709 image. These two transforms can be collapsed into a single LUT at the expense of highlight values above 1. It's possible the entire range could be preserved by underexposing the image and using the LUT to lift the exposure.

In summary, if there isn't a built-in way, there probably is a way of forcing the Ninja V or anything else to display a usable monitoring image with some hacks and workarounds.

[Ryan Earl]

23 hours ago, Llaasseerr said:

I'm thinking of making this as simple as possible, just an ISO change. Just so that there's a bright enough source that will clip the sensor at all those ISO settings, and also that there's enough shadow detail to challenge the noise level a bit.

I exposed the following frames starting at ISO 100 using false color.  I just kept the wall behind the light without clipping and the mask with some clipping, the gray card at ISO 100 is underexposed. 

Not sure how helpful this will be. . .

https://we.tl/t-82kFKys7LH

[Llaasseerr]

2 hours ago, Ryan Earl said:

I exposed the following frames starting at ISO 100 using false color.  I just kept the wall behind the light without clipping and the mask with some clipping, the gray card at ISO 100 is underexposed. 

Not sure how helpful this will be. . .

https://we.tl/t-82kFKys7LH

Thanks! I'm taking a look right now. It's helpful that the light clips and that the underside of the cabinet is still relatively in shadow at the highest ISO.

[Llaasseerr]

10 hours ago, Ryan Earl said:

I exposed the following frames starting at ISO 100 using false color.  I just kept the wall behind the light without clipping and the mask with some clipping, the gray card at ISO 100 is underexposed. 

Not sure how helpful this will be. . .

https://we.tl/t-82kFKys7LH

Here's the highlight clipping  without/with highlight recovery based on the just the red channel:

ISO 100: 1.0 / 2.09

ISO 200: 2.0 / 4.13

ISO 400: 4.0 / 8.25

ISO 800: 8.0 / 16.1 

ISO 1600: 8.0 / 16.1

ISO 3200: 5.0 / 9.92

ISO 6400: 8.0 / 16.48

ISO 12800: 8.0 / 16.73

So the clipping point in the DNG is reached at ISO 800, but it's just due to amplifying the signal by +3 stops from base ISO 100. ISO 800 seems to be the starting point for the Ninja V implementation, which clips at ~9.34. This seems about right since I don't think it uses any kind of highlight recovery.

When I balance ISOs 100-800 to the same exposure, they're the same clipping point because 800 hits the max highlight value form that point on. The noise floor actually looks the same to me as well.

In this case, ISO 1600 and up are not adding anything in the highlights.

ISO 3200 clips earlier than the rest, so ~9.9 vs 16.1, but the noise floor is about the same as ISO 800. If you think about it, there's a net gain here because the highlights are holding about 1/3 stop less than 6400 but the noise floor is -1 stop less. So it's the same as 800, which is really just ISO 100+3 stops). 

So to me, the way to take advantage of this is to underexpose 3200 by at least 1 stop so you're rating for at least 6400. The effect is that the noise floor is the same as 6400, but the highlights now clip a bit higher.

This ISO difference is the same behaviour with the Ninja V at 3200. 

I'm uploading two waveforms:

1. ISO 3200+1 stop vs ISO 6400.  = +1/4 stops more highlight detail. Max value 19.86 vs 16.48 in the red channel, noise floor the same.

2. ISO 3200+2 stops vs ISO 12800 = +1 1/4 stops more highlight detail. Max value 39.7 vs 16.48 in the red channel, noise floor the same. 

For reference, Alexa's sensor clips at 55 in linearized logC which is only half a stop more, so it seems that the shadow detail would hold up okay if the ISO rating was pushed further.

I was going to upload a split screen of the shadow noise, but it honestly looks the same in each image when inspecting each channel individually so take my word for it.

It would definitely be interesting in more test images showing the same ISO range at the same exposure level (adjusted with either lens or ND), and also protecting the highlights. It does seem like there's plenty of room to push shadow detail way down and still recover a decent image.

EDIT: I just want to clarify that in the +2 stops waveform image (LHS), the highlights are not getting clipped but they're in the upper code values of the log range and can be pulled back.

 

[Llaasseerr]

15 minutes ago, Llaasseerr said:

 

I'm just noticing the shadow levels are getting a little lost at the bottom of the waveform for the pushed 3200 image, so despite the fact it looked visually the same to me when inspecting noise per channel, it evidently is throwing a bit of detail out in service of the highlights. It will be interesting to see how the deepest shadows hold up if the image is shot -2 stops or more underexposed.

Also, sorry for the numerous typos in the above post.

[OleB]

For what it is worth. FCPX indeed is interpreting the fp RAW files as V-LOG, but is not applying the RAW to LOG conversion and not LOG to V-LOG LUT.

But if you select both options to Panasonic V-LOG the picture will look identical to both options off. That is great, because you can proceed to edit with ARRI Log C as LUT if you like. Gives the colors a little more pop.

[Llaasseerr]

5 hours ago, OleB said:

For what it is worth. FCPX indeed is interpreting the fp RAW files as V-LOG, but is not applying the RAW to LOG conversion and not LOG to V-LOG LUT.

But if you select both options to Panasonic V-LOG the picture will look identical to both options off. That is great, because you can proceed to edit with ARRI Log C as LUT if you like. Gives the colors a little more pop.

Yes that makes sense, it would just be looking at the metadata tags in the ProRes RAW file that say it's V-log and applying the linear to log transform for display. But as you say since it's not baked in, you can easily change the transform to another log curve/gamut available from the drop-down list. I guess Atomos just decided that for generic cameras they will set it to V-log.

It's not meant to make any difference what log curve and gamut you use as intermediate encoding if you have a colour managed pipeline. For example in ACES if you use the V-log IDT for V-log or the Alexa IDT for LogC. But it's worth noting that V-log clips at 46 in linear space vs 55 in LogC. Probably not a problem for the Sigma fp though!

[Ryan Earl]

15 hours ago, Llaasseerr said:

It would definitely be interesting in more test images showing the same ISO range at the same exposure level (adjusted with either lens or ND), and also protecting the highlights. It does seem like there's plenty of room to push shadow detail way down and still recover a decent image.

As I did the ISO ramp to 12,800 I went backwards to 100, but only 100, 800, 3200, 12,800

I added 5-8 second clips so you can see the noise, though I didn't match exposure perfectly, I was moving the shutter and aperture for approximately middle gray on the card in the right of the frame. 

https://we.tl/t-squvulFUOY

Thanks for your other feedback, I've been testing 3200 ISO in real world daylight scenes with 4 - 5 stops of ND and will try a few 1 stop underexposed along with your findings. 

[Ryan Earl]

17 minutes ago, Ryan Earl said:

I added 5-8 second clips so you can see the noise, though I didn't match exposure perfectly, I was moving the shutter and aperture for approximately middle gray on the card in the right of the frame. 

https://we.tl/t-squvulFUOY

Above link doesn't work, this is the correct one:

https://we.tl/t-fLzZl7fRNy

[Llaasseerr]

4 minutes ago, Ryan Earl said:

Above link doesn't work, this is the correct one:

https://we.tl/t-fLzZl7fRNy

For sure, shooting 3200 with that much ND seems like a really interesting possibility if the black levels hold up!

Thanks for re-uploading.

[Llaasseerr]

I know this is getting off topic of the thread which is about the external recorders, but as far as using an SSD for internal DNG recording, if you use an NVMe stick and can get 1GB/s across USB-C then couldn't Sigma do a firmware upgrade where the camera could potentially output 6k RAW or HFR 4k 12-bit?

It could potentially do 72fps.

[TomTheDP]

27 minutes ago, Llaasseerr said:

I know this is getting off topic of the thread which is about the external recorders, but as far as using an SSD for internal DNG recording, if you use an NVMe stick and can get 1GB/s across USB-C then couldn't Sigma do a firmware upgrade where the camera could potentially output 6k RAW or HFR 4k 12-bit?

It could potentially do 72fps.

4k at 60p would really be something.

[Llaasseerr]

9 hours ago, Ryan Earl said:

As I did the ISO ramp to 12,800 I went backwards to 100, but only 100, 800, 3200, 12,800

I added 5-8 second clips so you can see the noise, though I didn't match exposure perfectly, I was moving the shutter and aperture for approximately middle gray on the card in the right of the frame. 

https://we.tl/t-squvulFUOY

Thanks for your other feedback, I've been testing 3200 ISO in real world daylight scenes with 4 - 5 stops of ND and will try a few 1 stop underexposed along with your findings. 

My quick finding from these clips is that at the same exposure, ISO 3200 is only marginally noisier than 800 so my prior thesis still kind of holds. I'm trying to clarify what the advantage of 3200 is, rather than the stream of conscious observation in my prior long post, So I think the simplest way to describe it with the limited dataset is:

ISO 3200 = about the noise level between 800 and 1600, probably due to the change to the higher base ISO. But also, lower highlights capture.

ISO 3200 +1 stop = same noise level as ISO 6400, but 1/4 stop more highlight capture.

I'm just speculating for now, but if you were going to play the underexpose game then underexposing ISO 3200 by one stop more than underexposing 800 will give the best DR. I think. So maybe pushing it to the limit would be comparing underexposing 3200 by 5 stops vs underexposing 800 by 4 stops. Then back off from there one stop at a time if the black details are getting destroyed. I'm not including 1600 or 6400 because they don't seem relevant.

 

 

[Llaasseerr]

29 minutes ago, Llaasseerr said:

I'm just speculating for now, but if you were going to play the underexpose game then underexposing ISO 3200 by one stop more than underexposing 800 will give the best DR. I think. So maybe pushing it to the limit would be comparing underexposing 3200 by 5 stops vs underexposing 800 by 4 stops. Then back off from there one stop at a time if the black details are getting destroyed. I'm not including 1600 or 6400 because they don't seem relevant.

Just to put things in context, I measured ISO 3200 +1 stop with still a clean noise floor and clipping at a bit less than 20 in RGB, and Arri RAW clips at about 60. So pushing +3 stops with ISO 3200 could put it about 1/3 stop extra highlight detail over an Alexa, with the huge caveat of we would need to see how the shadow detail is holding in the raised noise floor.

And obviously you could play this whole game of underexposing with an Alexa to a degree, but I'm just looking at a baseline.

[Llaasseerr]

9 hours ago, TomTheDP said:

4k at 60p would really be something.

I'm quite into 72fps because it's a clean multiple of 24 and there are advantages to that, but 60 is obviously handy too.

[TomTheDP]

9 hours ago, Llaasseerr said:

I'm quite into 72fps because it's a clean multiple of 24 and there are advantages to that, but 60 is obviously handy too.

48 would be good enough for me.

Also on another note I just bought this. Hoping it will completely eliminate breakage on the micro hdmi.
 

[Llaasseerr]

36 minutes ago, TomTheDP said:

48 would be good enough for me.

Also on another note I just bought this. Hoping it will completely eliminate breakage on the micro hdmi.
 

Does that turn it into a full size HDMI port at the top?

Yes, 48 would def be nice too.

[TomTheDP]

20 hours ago, Llaasseerr said:

Does that turn it into a full size HDMI port at the top?

Yes, 48 would def be nice too.

Yes it does.

[OleB]

Took the public holiday today to investigate further, results before were not good enough.

You can disregard my tables. Dead end street...

Llaasseerr you were right, if I underexpose middle grey the post production is a becoming a big headache.

Now what I did today is set up lighting (camera ISO 800) and my color checker passport. Turned on false colors and tuned the lighting so that the patch which is representing middle grey was green (44 - 47 IRE) in the Atomos. (the one marked in the picture)

False color screen showed the following values for those patches, starting with white...

93 IRE, 84 IRE, 75 IRE, 59 IRE, 46 IRE, 26 IRE

Loaded the file into FCPX, but instead of pulling the highlights down as usual I did something different this time. On import the file looks totally over...

The new idea was to create a curve to bring down the patches to their appropriate brightness.

The curve, as smooth as I could get it, looks like this for all ISO values expect for ISO 3200.

ISO 800:

ISO 3200:

Last but not least I now had to search for the correct light meter values to get the exposure matching with my middle grey patch of the color checker. 

Camera - light meter

ISO 800 - 250

ISO 1600 - 640

ISO 3200 - 2000

ISO 6400 - 2500

ISO 12800 - 5000

ISO 25600 - 10000

That's it 🙂 I am happy now. Can use the false colors now correctly and even take a meter reading which is matching the Ninja. Since I have saved my curves as a preset it is now a one double clic action to get the file in FCPX exactly matching what I saw when recording.

Basically it seems I have created my own brightness calibration file. Might take a further investigation in regards to colors, but for now the result seems pleasing. Obviously I have tested it with various of my previous recordings and it is just so much better.

Maybe this could be of help 🙂