HowTo? Best practice for using Dual-ISO?

[l_d_allan]

I'm unclear on how to use Dual-ISO to best advantage. I think I may be close to a good workflow, but I wanted to check for sure, and also learn about "the error of my ways" from others on this forum with more experience and expertise.

• Appropriate for scene with high contrast, where HDR would help.
• Even with RAW histogram, there is "valuable content" on the left and/or right side.
• Otherwise, "why bother?" if the levels fit within the histogram ok (low to moderate contrast)
• For an example using ISO 100/1600:
• Dual-ISO starts at OFF
• (but does Auto-ETTR take Dual-ISO into account, so this isn't necessary or appropriate?)
• Or does the camera's normal metering end up being appropriate?
• Determine best exposure for bright areas at ISO 100, perhaps with ETTR, trial-n-error, etc.
• ...
• The RAW histogram would tend to look like it was ETTL (exposed to the LEFT)
• Exposure will tend to be underexposed, with few if any blown highlights, but will probably appear to have "blocked shadows" on the LCD.
• Enable Dual-ISO (this example uses 100/1600)
• Take picture(s)
• Process with cr2hdr.exe
• Open with ACR, LR, or other
• Expect to see an image that looks rather underexposed
• Image should have quite a bit more usable dynamic range, so the exposure/shadows/blacks can be increased significantly without as much noise happening

Or not?

Am I "unclear on the concept"? Do I have a flawed understanding of how Dual-ISO is supposed to work?

Is my workflow more or less ok, but sub-optimal? Would a different technique work better? Am I leaving something out? I haven't really been following the progress on ETTR and Auto-ETTR, but my guess is that those capabilities are now somewhat integrated.

My speculation is that once you "get the hang of it", you don't have to disable Dual-ISO, but can more go by the metering and/or Auto-ETTR.

[a1ex]

ETTR with default settings should be a good starting point. If it enables dual ISO, it's likely to get noticeable improvements. Indeed, your best bet is an image that looks very dark if you expose it to the right (because in this case you will probably want to raise the shadows).

It also depends heavily on how are you going to color grade your pictures, and how good is your denoising software. If you raise the shadows a lot, it will help. If not, the improvement may not be noticeable.

If you use DxO PRIME for denoising, you probably won't see any improvement except for really extreme cases. For me, 45 minutes for denoising a single image is a bit too much (but the results are fantastic IMO). With a GPU, I think it only takes a few minutes from what I've read.

[Audionut]

Appropriate for scene with high contrast, where HDR would help.

Correct.

Even with RAW histogram, there is "valuable content" on the left and/or right side.

The RAW histogram is just giving you a correct representation of the captured detail.  The content would only be "valuable" if it falls outside the bounds of what the camera can capture in a single exposure, and you decide you don't want it blown to white (overexposed), or to far in the shadows (underexposed).

Otherwise, "why bother?" if the levels fit within the histogram ok (low to moderate contrast)

Exactly.  Unfortunately, dual-ISO doesn't come for free.  It has limitations that must be considered.

For an example using ISO 100/1600:
Dual-ISO starts at OFF
(but does Auto-ETTR take Dual-ISO into account, so this isn't necessary or appropriate?)
Or does the camera's normal metering end up being appropriate?
Determine best exposure for bright areas at ISO 100, perhaps with ETTR, trial-n-error, etc.

There are 2 exposures with dual-ISO.  In simplest terms, the lower ISO captures the highlight detail and the higher ISO captures the shadow detail.  Where both ISOs capture detail, full resolution is retained.  Where only the lower ISO captures highlight detail (higher ISO becomes overexposed), then you suffer reduced resolution (aliasing).

The RAW histogram would tend to look like it was ETTL (exposed to the LEFT)

Generally yes.

Exposure will tend to be underexposed, with few if any blown highlights, but will probably appear to have "blocked shadows" on the LCD.

The only blown highlights will be those that you choose to blow.  You retain control over exposure with dual-ISO.
It will tend to look very dark though due to the way cr2hdr (correctly) merges both exposures.

Take picture(s)
Process with cr2hdr.exe
Open with ACR, LR, or other
Expect to see an image that looks rather underexposed
Image should have quite a bit more usable dynamic range, so the exposure/shadows/blacks can be increased significantly without as much noise happening

Correct.

There are some tutorials here and here which may be useful.

[l_d_allan]

Thanks for the input. Whew. It seems like I am not too far off.

It also depends heavily on how are you going to color grade your pictures

I've seen "color grade" mentioned, but I am ignorant about what that means. I did find the post on "Auto Color Grading", but it was mostly over my head. Also, it seemed dependent on quite a few tools that I'm unfamiliar with.

Is this a valid summary: gets all pictures "in the ballpark" as a starting point.

[Audionut]

I've seen "color grade" mentioned, but I am ignorant about what that means.

In a nut shell, adjusting the image in post.  Exposure, contrast, highlights, shadows, saturation, etc, etc.

When you take a dual-ISO image and raise the shadows, that's grading.

[l_d_allan]

Exactly.  Unfortunately, dual-ISO doesn't come for free.  It has limitations that must be considered.

I wonder if my list of limitations is correct: complicates and slows down workflow, some loss of resolution .... others?

There are 2 exposures with dual-ISO.

Now that's different from my newbie understanding. I had thought that during a single exposure, different scan lines of sensor pixels had different tone curve characteristics (i.e., ISO).

Are there two actual exposures? Is the sensor turned on and back off twice? I suppose that happens quite quickly, but with a moving target, could you get some ghosting? For a 5d2 with Digic-4, any idea of the time delay between the exposures?

Where both ISOs capture detail, full resolution is retained.  Where only the lower ISO captures highlight detail (higher ISO becomes overexposed), then you suffer reduced resolution (aliasing).

I recall the original post describing Dual-ISO mentioned resolution loss, but I thought it was more straightforward than what you describe.  Your explanation fills in some gaps.

So .... if you had Dual-ISO enabled with a relatively low contrast scene with no blown highlights and no blocked shadows, then the "ON" and "OFF" files would be very, very similar? Little or no loss of resolution?

[l_d_allan]

In a nut shell, adjusting the image in post.  Exposure, contrast, highlights, shadows, saturation, etc, etc.

So, the "Auto Grading Script" seems like it might be a smarter "Auto button" than ACR and LR have? Or more flexible with more user control?

[Audionut]

I wonder if my list of limitations is correct: complicates and slows down workflow, some loss of resolution .... others?

I only consider the resolution problems as a limitation.  Increased workflow is just that, it doesn't limit the rendering capability of the scene.

Now that's different from my newbie understanding. I had thought that during a single exposure, different scan lines of sensor pixels had different tone curve characteristics (i.e., ISO).

Are there two actual exposures? Is the sensor turned on and back off twice? I suppose that happens quite quickly, but with a moving target, could you get some ghosting? For a 5d2 with Digic-4, any idea of the time delay between the exposures?

My apologies.  There aren't 2 exposures as in the shutter actuates twice.  There are 2 exposures as in you should consider both ISO's when determining correct exposure.

For instance you might use ISO 100 for the highlights, but you have to consider what your other ISO is doing to wanted detail.  The higher ISO controls shadow detail, but it also determines where aliasing becomes a problem in the highlights.

I recall the original post describing Dual-ISO mentioned resolution loss, but I thought it was more straightforward than what you describe.  Your explanation fills in some gaps.

Resolution loss will occur where either ISO is overexposed, since there is no longer full detail from both exposures.

So .... if you had Dual-ISO enabled with a relatively low contrast scene with no blown highlights and no blocked shadows, then the "ON" and "OFF" files would be very, very similar? Little or no loss of resolution?

I would expect the dual-ISO shot to be a touch better.  Because you clearly haven't ETTR the lower ISO and the higher ISO will now be ETTR.  Of course, you could have just ETTR the low ISO and called it a day

[Audionut]

Resolution loss will occur where either ISO is overexposed, since there is no longer full detail from both exposures.

The resolution loss comes from the line skipping.  Where the higher ISO is overexposed, you have half the lines with detail (the lower ISO) and half the lines with no detail (overexposed, the higher ISO).
Just like video, the actual amount of aliasing is determined by the detail of the scene.  Flat textures are easy to fill in, sharp detail is harder.

[a1ex]

So, the "Auto Grading Script" seems like it might be a smarter "Auto button" than ACR and LR have? Or more flexible with more user control?

Yes. I don't use ACR or LR though, so I have no idea how smart they are (though I could try a demo). I've compared my script against rawtherapee's auto: https://www.dropbox.com/sh/013ewzjdjtj7p55/rBM16yO5Rh

[l_d_allan]

I would expect the dual-ISO shot to be a touch better.  Because you clearly haven't ETTR the lower ISO and the higher ISO will now be ETTR.  Of course, you could have just ETTR the low ISO and called it a day

Thanks for your patience on providing a number of clarifications.

Based on what you've written, there doesn't seem to be much, if any, downside to having Dual-ISO enabled all the time. With static images, you'd take the time to get ETTR at a good setting.

If I understand what you've written, with a low to moderate contrast scene (as evidenced by fitting within the histogram), most or all of the information would come from the low ISO part of the exposure. The high ISO info wouldn't apply.

But when you did need the expanded DR of Dual-ISO, it would be there.

Somehow, I think that conclusion reflects a simplistic understanding on my part.

And to me, a more complicated and slower workflow is a liability (not a limitation, as you point out).

Some geek speculation:
The impression I have is that during the actual exposure, a ML routine is invoked that examines a pair (or quadruple?) of scan lines. In real time during the exposure, it is deciding whether to used just the low ISO pixels, or to merge the high and low ISO info.  It almost seems like that would be on a pixel by pixel basis.

But I have my doubts that is really going on.

Also, I'm unclear how "aliasing" comes into it? Do you have a situation were two (or four?) pixels are averaged and/or interpolated? I observe that the thumbnails have a huge amount of moiré, but that's gone after cr2hdr.exe works its magic.

But I guess I don't have to understand how it works to be able to use. Whew.

[Audionut]

If I understand what you've written, with a low to moderate contrast scene (as evidenced by fitting within the histogram), most or all of the information would come from the low ISO part of the exposure. The high ISO info wouldn't apply.

It would where the higher ISO becomes overexposed.  And this is determined by the low ISO exposure and what higher ISO setting you have.
A low contrast scene that has been ETTR with the lower ISO, is going to overexpose the higher ISO (lose detail).

Also, I'm unclear how "aliasing" comes into it? Do you have a situation were two (or four?) pixels are averaged and/or interpolated? I observe that the thumbnails have a huge amount of moiré, but that's gone after cr2hdr.exe works its magic.

The lines you seen in a thumbnail are different to actual aliasing that may or may not be present.

Consider a very early example here.

Those lines all through the photo are aliasing.  Dual-ISO scans 2 lines at 1 ISO and the next 2 lines at the other ISO.  In the example linked, every other 2 lines of detail was overexposed.  Because those every other 2 lines where overexposed, there was no detail to fill in the gaps.  a1ex has made leaps and bounds in the processing since that early example, but the issue will still be there in limited situations.

But I guess I don't have to understand how it works to be able to use. Whew.

A better understanding of how it works, leads to a better ability to use it!  It has limitations, and these need to be considered.

Follow the tutorial here and the links there.  With Auto ETTR + dual ISO + SNR limits set, this becomes a set and forget affair for probably 99% of all situations.

[Audionut]

If you use DxO PRIME for denoising, you probably won't see any improvement except for really extreme cases. For me, 45 minutes for denoising a single image is a bit too much (but the results are fantastic IMO). With a GPU, I think it only takes a few minutes from what I've read.

Unfortunately, denoising cannot capture more detail, and is bound to destroy fine details  Dual-ISO can capture extra detail without the noise. 

[a1ex]

The detail is already there, just that noise is stronger. A good denoising algorithm will try hard to keep the detail; the non-local algorithms won't averae the neighbours of each pixel, like traditional algorithms do; they will find patches that look similar and will average those.

Here's an animation: http://www.cs.tut.fi/~foi/3D-DFT/

A similar algorithm can be used for super-resolution (that is, do a really good guess of missing pixels).

A comparison would be nice. My subjective impression is that DxO reveals roughly 1 stop of shadow detail, and noise is reduced by roughly 3-4 stops.

[Audionut]

I regard detail as the highest priority, fine detail especially.  I very rarely apply any luma denoising on my images.

There are denoising algorithms that work well, but they come at the expense of to much fine detail loss for myself.  I also prefer to use the subjective increase in detail that noise produces where possible.  I also very rarely view my content at a 1:1 ratio.  It's either resized for web (averaging), or printed as such that the noise is not visible when viewed at reasonable viewing distances.

edit:  At the noise levels present in the first 2 images in the PDF you linked, denoising would be preferred.

[a1ex]

Got a little test for DxO PRIME:

original raw from Greg
dxo output (dng)
a raw with shutter speed slower 2 stops

and jpeg files developed with ufraw from the above files (click for full-res images):


Forgot a small detail: here's how fast is DxO without a GPU:


It would be great if you or anyone else can do a similar test with dual ISO vs DxO.

[Audionut]

As expected, DxO has destroyed fine (and medium IMO) detail.

Above that AC control knob but below the radio there is a band of textured fabric/plastic.
In the 2 stop slower shutter shot, detail is present.  In the DxO sample, it's been destroyed.  Not to mention all of the artifacts in the DxO sample.   

I'll run some tests myself, but don't expect them until Sat/Sun GMT+10.

[a1ex]

Yeah, this is why I said it only brings back roughly 1 stop of detail. Still, I find it much cleaner than any other denoiser I've seen.

Dual ISO also removes some fine shadow detail in deep shadows (but it still does well on test charts, where noise gets hidden by strong detail). Not sure how it compares; the theory says 3 stops, but I didn't actually compare it with a 2x3EV bracketed shot.

Regarding combined DxO and dual ISO, I've noticed we have almost reached the limits of the 16-bit DNG format. That is, I've experimented with some highlight recovery algorithms, and if I reserve two bits for recovered highlights, the deep shadows (where the noise is strong enough to benefit from DxO) will get destroyed.

So, if we ever manage to implement dual shutter or a denoiser comparable to DxO, we should look at some file format that handles HDR better. I've looked at OpenEXR and floating point DNG, but both of them would require a major change in my workflow.

[Marsu42]

I've looked at OpenEXR and floating point DNG, but both of them would require a major change in my workflow.

... but no doubt this is the way to go, I'm processing all my HDR as EXR recently in apps and then convert to floating point DNG in Lightroom. This is a major quality boost when doing further postprocessing over 16 bit, and the EXR-based floating point hasn't got an as hilarious filesize as floating point TIFF:

[a1ex]

I can create floating point DNGs, but none of the open source programs I've tried could open them. Of course, I could convert them back to plain DNG with Adobe converter. Also, dcraw & friends do their internal processing as 16-bit, so even if I patch them to read floating DNGs, it wouldn't solve anything.

RawTherapee has a floating point pipeline, and I think darktable too.

[Marsu42]

Also, dcraw & friends do their internal processing as 16-bit, so even if I patch them to read floating DNGs, it wouldn't solve anything.

Indeed, and since traditional raw has only 14bit there's probably little reason for them for change this anytime soon. But maybe it's feasible to update the dcraw code to floating point yourself as you already know about the floating point dng format?

Yes. I don't use ACR or LR though, so I have no idea how smart they are (though I could try a demo). I've compared my script against rawtherapee's auto: https://www.dropbox.com/sh/013ewzjdjtj7p55/rBM16yO5Rh

Btw the ACR/LR "auto tone" is a complete catastrophe since they switched to pv2012, they silently keep tuning it with every update, but it still looks awful with every shot I've done. The only usable "auto" of ACR is auto exposure (double click the ev slider), if you really want working "auto" use DxO as they've got many good and actually working presets.

[a1ex]

There is already a dcraw-float version, will look into that. If I can mod it to use Kelvin WB and the soft film curve from ufraw, that's pretty much all I need for my scripted workflow.

[Audionut]

Btw the ACR/LR "auto tone" is a complete catastrophe since they switched to pv2012,

Agreed, I pushed the auto button once and couldn't believe what had happened    Luckily I prefer to shoot manual and can mostly batch process a bunch of shots with the same settings.  Also, the highlight recovery of pv2012 is excellent.

[a1ex]

Also, the highlight recovery of pv2012 is excellent.

Let's see:

Original raw (blue channel is not clipped at all, red clipped a little, green clipped a lot)

ACR 8.3, -3EV, developed by Greg:


LR 5.3 demo, -3EV, developed by me (side question: why it's everything green at 5000K and zero tint?!)


A dumb algorithm I'm experimenting with (jpeg from ufraw):


DNG created by my algorithm (yep, same as the one from ufraw bug)

[Audionut]

That's a good test case.  Normally ACR/LR will recover detail if only one channel is blown.  In Adobe applications you should use the highlight slider first to recover highlight detail, although in this case it makes no difference.

Have you made further changes to ufraw?  How does it fair now on this?

edit:  I should have probably read more of that thread