HDR Bracketing: One or two stops?

[Timar]

It was in your line, "In a way I have done a lot of HDR work (working with the curves and sliders in Lightroom to map the dynamic range of the RAW file into the 8 bit output)". This made it appear to me that you were processing the image in LR as 8-bit files, not 16-bit, I believe that you would only change to 8-bit for web or printing at an outside lab.

Oh, I see. I actually meant the final JPEG output to which the merged DNG is reduced once you export it (but also to display it on the monitor). The DNG itself is 16 bit floating point.

As to the question of White Balance, you will tell me if I am incorrect, I believe that it is the only thing that is applied to a raw file as it gives a reference point to start with in all raw converters. In the development Mode in LR, it gives: as shot, auto, Daylight, etc., etc. it would seem that if you merged two raw files or more without any processing as it sound to me, then the result of the merged image would have the colour balance that the camera was set to. Is that not so?

Not quite correct. Raw files indeed have a color temperature assigned to them by the camera, but only in the meta data. The raw converter then applies the color temperature (as well as the black point and gamma curve) after demosaicing the sensor data and applying some basic corrections to them. This is why it is called "raw" file. It is really raw - the data straight out of the sensor, just accompanied by meta data that tells the converter what to do with it.

For more in-depth technical information see Guilliermos website or this blog article.

[DanK]

As to the question of White Balance, you will tell me if I am incorrect, I believe that it is the only thing that is applied to a raw file as it gives a reference point to start with in all raw converters. In the development Mode in LR, it gives: as shot, auto, Daylight, etc., etc. it would seem that if you merged two raw files or more without any processing as it sound to me, then the result of the merged image would have the colour balance that the camera was set to. Is that not so?

Your last sentence is correct. Timar is of course right that as long as the file remains in raw format, the white balance information is held in metadata. However, your statement is correct regardless. As far as I have been able to ascertain, in Lightroom, if you do nothing to the images you are going to merge, the software will retain the "as shot" WB it uses to render the initial images when it does the HDR merge. I rarely use the LR HDR feature and instead usually merge exposures using the Lightroom Enfuse plugin, but it behaves the same way.

[Manfred M]

Okay - I've decided to do my own experiment to see if the theory that two images should give me the same results as five (or more).

I shot some bursts a few weeks ago at sunset while I was in Ontario's Algonquin Park area at just before sunset. I shot +2, +1, 0, -1, -2 using the Nikon D800. As I cannot vary the automatic bracketing by more than 1 stop, I got the exposures 1 stop apart. The camera was set to Aperture Priority and the shots were taken at the CH (high speed burst) mode and I used a cable release to fire the camera; all shots were taken within around a second of each other. I shot on a very heavy duty and stable tripod (Feisol 3372 tripod) with a RRS BH-55 Ball Head. I can take multi-minute, rock steady shots with this gear.

All five shots were processed identically (16-bit NEF files output in the ProPhoto colour space). I assembled the +2 and -2 shots into an HDRI (again same bit depth and colour space) using the Adobe ACR CC 2015 HDRI engine (this is identical to the one in Lightroom).

This is one of the images. There is no PP work done to it (or the second image) other than converting from the ProPhoto colour space into sRGB 8-bit jpeg for posting on the internet.


1. This image uses 5 separate exposures that are 1 stop apart:




2. This image uses 2 exposures shot at +2 and -2 from metered "optimal"





I then took the two HDR images (16-bit, ProPhoto) into Photoshop and used the Difference blending mode. This does a pixel by pixel comparison of the absolute value of the difference of the two image. This is the resultant image. I did set the black point, white point and gamma adjustment as I normally would, so it is easier to see the the differences between the 2-shot and 5-shot HDRI processed image. The differences were visible in the original, but not as clearly as can be seen here.



Conclusion - the 2-shot and 5-shot HDRI images are NOT IDENTICAL.

Is the difference significant or meaningful? That I can't say. The difference could be due to movement between the two images (unlikely as there was little to no wind and the shooting platform was extremely stable). All I can say is that the algorithms used in the Adobe software do actually use the data in all the shots.

Thoughts or comments?

[DanK]

Manfred,

Very interesting. Thanks for doing this and posting it. At this resolution, the differences are subtle, but there are places where they are visible. Clearly, information is lost when you use 2 shots rather than 5.

I often shoot at 2-stop intervals. At some point, I will have to test that vs. 1-stop.

Dan

[Dave Humphries]

Conclusion - the 2-shot and 5-shot HDRI images are NOT IDENTICAL.

Is the difference significant or meaningful? That I can't say. The difference could be due to movement between the two images (unlikely as there was little to no wind and the shooting platform was extremely stable). All I can say is that the algorithms used in the Adobe software do actually use the data in all the shots.

Thoughts or comments?

Well, clearly noise is another area where they will differ, in fact I wonder if the averaging of that across five shots vs two is significant?

I am also sure that there will be some movement differences introduced between exposures, even without noticeable wind; e.g. falling leaves, water ripples from insects - of course these are all quite high resolution things, so I then wondered whether applying some kind of blur (Gaussian?) to both before comparing/differencing might be beneficial, but don't ask me what a sensible radius would be.

Switching between the two shots in LyteBox, I cannot claim to be able to see a difference between them, even though I know, thanks to the difference image, where to look. Dan must have good eyes (or mine are tired tonight)

If I were to try HDRI, I'd probably use three images two stops apart rather than 5 images, one stop apart, or two images five stops apart, although I appreciate you didn't have an option at shooting time.

[DanK]

Well, I might be seeing things , but after looking in the lytebox, I thought I saw some subtle differences, eg, a little more detail in the plants on the dark shore in the middle. However, it might be a placebo effect.

[Dave Humphries]

Well, I might be seeing things , but after looking in the lytebox, I thought I saw some subtle differences, eg, a little more detail in the plants on the dark shore in the middle. However, it might be a placebo effect.

OK, with serious pixel peeping at 100%, I began to see differences; the five shot has greater micro-contrast and finer detail.

Also, especially noticeable in the reflection of the distant, high altitude clouds roughly dead centre, was an 'area' of difference in blue sky peeking through a gap in the clouds - suggesting to me there were differences between first and last shots and these were more averaged across the five shot series than the two (first and last only).

[Timar]

I'm glad that you decided to test the theory for yourself, Manfred. It's a beautiful photo.

Conclusion - the 2-shot and 5-shot HDRI images are NOT IDENTICAL.

Of course they are not. How could they possibly be? First of all, they contain information captured in different moments in time, so they can never be identical, even if the scene seems compelely static. Assuming the vibrations in the tripod would be entirely neglegible, objects in a scene will undergo subtle movements (e.g. leaves in the trees, ripples on the water, clouds in the sky) which may be below the threshold to become apparent as ghosting artefacts, but they will decrease sharpness. Moreover, the more exposures you merge, the lower the noise floor for the overlapping (midtone) sections of the response curve will become if the merging algorithm employs some kind of averaging.

Yet your example images nicely illustrate what I wrote above - that there is no advantage whatsoever in using more than 2 exposures conveying the entire dynamic range of the scene. They look identical to a degree that you probably couldn't tell apart in a blinded* AB comparison that randomly switches between both images, and if you could, it would be most likely because of ghosting artifacts present in the 5 exposure rendering and you would therefore likely prefer the 2 exposure rendering.

This is exactly what I meant when I wrote that there is "nothing to be gained by using more than two different exposures, spread far enough apart to convey entire the dynamic range of the scene". If you think that the lower noise floor from avaraging would provide any visible benefit, remember what you are actually doing when tone-mapping the dynamic range of a HDR file: A lower noise floor in the midtones would produce a visible advantage only if you extended the dynamic range in that area - but what you do is actually the opposite - you compress the dynamic range into the 8-bit output. This is the very reason why Adobe wisely decided to merge HDR composites into a 16 bit floating point format.

*It is very important to compare such images blinded, i.e. without knowing which is "supposed to be superior". Otherwise you start to imagine perceptual differences where none exist. This is just the way the human mind works - there is no way to escape the placebo effect other than by being blinded.

[Manfred M]

Actually Timar, based on the result I see I would have to draw the opposite conclusion that you have. Like Dan I see a small but definite improvement in image quality when looking at the two images, so I am going to have to reassess my view on making HDRI in the future.

Double blind testing is a valid approach for some approaches, especially where efficacy of drugs is involved. In matters like these, I would prefer to have expert opinion from users who have good colour vision and a high level of expertise in colour accutance using calibrated and profiled wide gamut screens or prints.

[Manfred M]

Dave - Noise should not have been an issue as all the shots were taken at base ISO (100 across the range of all the images) and exposures varied from 1/15th to 1s. These are all in a range where noise is not an issue. Mirror slap is probably the one variable that might impact these images as all of the shutter speeds are in the range where I would often shoot with the mirror locked up.

I shot at f/13 so still in the range where diffraction is just getting to be an issue, i.e. 1/3-stop below f/11. Previous testing has shown to be the bottom end of the "sweet spot" on the f/2.8 24-70mm lens and is usually as small an aperture as I will shoot with to maximize DoF without getting too much softening from diffraction..

Focal length used in all the shots was 31mm, so fairly wide angle.

[dragon76]

Okay - I've decided to do my own experiment to see if the theory that two images should give me the same results as five (or more)

Conclusion - the 2-shot and 5-shot HDRI images are NOT IDENTICAL.

Is the difference significant or meaningful? That I can't say. The difference could be due to movement between the two images (unlikely as there was little to no wind and the shooting platform was extremely stable). All I can say is that the algorithms used in the Adobe software do actually use the data in all the shots.

Thoughts or comments?

Manfred, your post is one of my favourite posts in this forum as you actually go out, take the shots to prove your point. In regarding, 2 exposures HDR vs 5, 7, I haven't shot HDR as long as most people here in the forum but from what I have experienced, I can concur with you that 2 exposures HDR and 5 exposure HDR are not the same. I found that with 5 shots HDR, it almost always has more tonality than the 2 shots especially with low light scenes. But the tonal differences can be easily be compensated in post so I mostly do 3 shots to save more shooting time for another angle. Also, I learned something new today, I have used different methods & program to compare images before but never thought about using Diff blend to compare images.

Cheers

Dean

[Polar01]

Well I am going to add my 3-1/2 cents Canadian into the topic. As I do my HDR images differently than almost everyone else, but that is just me I will not talk on that. So as I see or assume that when Timar does his images he takes 3 shots however he only uses the +2 and -2 shots. He then takes those two images then using LR's HDR or maybe another HDR type of program creates an image, believing that this is the best method.
Now Manfred did a test and produced one image by Timar's method, then one using 5 shots at 1-stop spacing giving the end result of the same 4 stops range. Now looking at a jpeg that was posted to the web is not the best way to compare images.
Now for the main part of this post, when looking at both or either of Manfred's images every without this being about HDR the images look unnatural to me own personal liking. I will explain they are just to bright and have lost a solid black as a result of the programs defaulting to increase the shadow areas upwards. I have 3 different HDR programs, sometimes one of them will handle the same image better than the other two, it just depends on the image. I find from myself 3 images work well, however they are not +2,0,-2 but, 1 1 stop over exposure for shadow, neutral, and 1 2 stops under exposed for the highlights. These valves I feel give me a better control over the final look of my image.
I have 3 great examples 2 are from Manfred post #63 and Timar's post #53, all three images are all very well done are interesting. However to my eye, Manfred's images needs the highlights greatly reduced and the exposure also reduced 1/3rd of a stop overall, this is the result of the program use. Timar's image in post #53 screams HDR reasons why, if one looks at mainly the right side of the persons leg, back, and coat if appears like a reflector is being used to direct light on the backside of the person again the result of the program wanting to lift the shadow areas to much. If you were standing behind the person for real, you would not see those highlights there, the clothing would have almost no detail at all, the biggest tell to me is how light the shadow is from the boot and lower leg on the ground it should be a lot darker than it is. Almost finished, that is why I do not like the 2 up 2 down method and prefer my method as the shadows are not lifted as much and the overall image is not as bright, I use this method with two of my HDR programs, for Photshop's HDR I use a different method again.
One last thing Timar in your post #68 you state " but what you do is actually the opposite - you compress the dynamic range into the 8-bit output." this is probably right, however this in unimportant to me, I do not care and these are reasons why, I create a jpeg that is posted to the web or shown on a screen, to do that I take a 16 bit file in psd. format in ProPhoto RGB colour space, with the resolution set to 360 ppi/dpi (yes I know the difference it is how resolution is stated in "image size" in Photoshop). Now I Rip it to 8 bits, convert to sRGB and set resolution to 72 (that is all that most monitors will handle) these settings are for the web if the image was going to an outside Lab than the ppi/dpi would be likely set to 240. As I post to the web I do not care about the compressing off the dynamic range of a 8 bit file that is going to be viewed on a incorrectly colour balanced crappy screen in a colour space that is also crap, so to worry about that is not important. My work is printed from 16 bit files in psd. format in the ProPhoto Colour space at a 360 ppi/dpi resolution by myself.
I go long winded and sorry for jumping around in my thoughts I do that (a lot).

Cheers: Allan

[Dave Humphries]

A couple of follow up points from me.

First let me say I am looking at this in a theoretical sense, I don't shoot HDR myself, so bear that in mind when reading the following - IOW "Dave may not know what he is talking about"
That said, I obviously think my points are valid, or I wouldn't state them

Noise will be a pixel level difference between any two shots, because of its random nature, yes it may be insignificant to the overall image if shot at 100 ISO, but it will be an issue when you do things like the difference merging.

Surely, rather always use 'one up, one down', or 'two up, two down', the range ought to be determined by the dynamic range of the scene in front of the lens? I guess with experience, people will get comfortable with 'their' usual way of doing it and come to know when the scene will 'fit' in whatever their sensor's DR is, plus a number of stops necessary either side to be sure everything is captured with out blowing/crushing detail - and that's good enough.

[DanK]

One last post from me. I noticed two things. One is that Timar's argument is not the same as the original GUI argument. GUI wrote:

An additional middle shot would improve noise in the middle tones, but it is not essential.

Timar wrote:

This means that there is nothing(!) to be gained by using more than two different exposures, spread far enough apart to convey entire the dynamic range of the scene.

This got me thinking. There are actually two different issues bobbing around in this sea: what one needs to capture the entire DR, and what happens to all the stuff in between the extremes. Some of the discussion of the latter seems to assume tone mapping,e.g., this from Timar:

A lower noise floor in the midtones would produce a visible advantage only if you extended the dynamic range in that area - but what you do is actually the opposite - you compress the dynamic range into the 8-bit output.

However, you don't need tone mapping to create HDR images. You can do it (I almost always do) with exposure fusion. That doesn't compress anything, AFAIK. It simply selects pixels from each image based on its exposure.

So, the question for folks like me is what ranges the software selects from each image and what effects that has on the quality of the resulting composite. I'll have to do a test someday...

[Manfred M]

Dave - I'm going to go out and reshoot some images to check the noise issue (once I figure out a methodology that will let me isolate this variable). My suspicion going in is that noise at base ISO is going to be a non-issue, given the shutter speeds in use. The other issue is the movement that you pointed out and I would suggest that this will most certainly another possibility, but when I look at static areas in the image - tree trunks and shoreline, I would probably suggest this is a none issue as well.

The third possibility (and the one I consider most likely) is that the HDRI software will provide better results with more data. With two images all calculations will based on the two data points. Add more data and the software will have more data to work with. The software (other than perhaps at a global level) doesn't know how to weight the contribution of each image in the stack, but I suspect more is always going to be "better". Add the "magic" of the de-ghosting algorithms, this really means we have no idea as to how a specific value is assigned to each specific pixel.

All that being said, I rarely do HDR these days for very much the same reasons Allan gives; I don't think the stand-alone results work and a lot more PP work is required to bring out what I want the image to look like. Straight out HDRI looks flat (and of course this makes sense as we have taken a high tonal range image and brought everything down to a level that our display / print can handle. These days I generally work with three images; the mid-range image is the base I use and I then use the one that brings out the shadow detail and the one that brings out the highlight detail with layer masks to bring that detail back into the whole shot. I find it is much faster and gives me more realistic looking results.

[Dave Humphries]

When I said "Noise will be a pixel level difference between any two shots, because of its random nature, yes it may be insignificant to the overall image if shot at 100 ISO, but it will be an issue when you do things like the difference merging.", I could/should more accurately should have said "Noise will be a pixel granularity difference between any two shots ~"

You have only to look at your Difference blended image to see how significant the noise is in that special 'test case', but as I acknowledged, in the real world, it isn't significant. I don't think there's anything useful to prove here, at least not to me - as I said, I really don't know what I'm talking about.

Regards methodology, in post #65, I suggested a blurring to remove the noise artefacts when comparing to see differences that were not masked by noise, but later - in post #67, I realised this would remove the actual visual benefit of the extra images.

I think Allan suggested he couldn't see a benefit that can't be achieved in PP in some other way - and he might be right there; the tiny differences I see pixel peeping your two images might well be imitated with just a little additional Clarity in ACR/LR and/or very fine radius sharpening applied to the '2 pic' version.

I also agree with you that 3 is better than two (data points/images).

Yes, I am just 'splitting hairs'

[DanK]

All that being said, I rarely do HDR these days for very much the same reasons Allan gives; I don't think the stand-alone results work and a lot more PP work is required to bring out what I want the image to look like. Straight out HDRI looks flat (and of course this makes sense as we have taken a high tonal range image and brought everything down to a level that our display / print can handle. These days I generally work with three images; the mid-range image is the base I use and I then use the one that brings out the shadow detail and the one that brings out the highlight detail with layer masks to bring that detail back into the whole shot. I find it is much faster and gives me more realistic looking results.

Manfred,

Have you tried exposure fusion? It is somewhat similar to your method, although with less control. I wonder how the two compare in terms of the amount of work needed to get to the image you want.

Dan

[Manfred M]

Manfred,

Have you tried exposure fusion? It is somewhat similar to your method, although with less control. I wonder how the two compare in terms of the amount of work needed to get to the image you want.

Dan

I downloaded Enfuse some time ago and played around with it. If I were using Lightroom as my primary editing tool I would definitely be tempted to use it, but with Photoshop and Layers, I find my method is faster and gives me more control.

[Manfred M]

When I said "Noise will be a pixel level difference between any two shots, because of its random nature, yes it may be insignificant to the overall image if shot at 100 ISO, but it will be an issue when you do things like the difference merging.", I could/should more accurately should have said "Noise will be a pixel granularity difference between any two shots ~"

You have only to look at your Difference blended image to see how significant the noise is in that special 'test case', but as I acknowledged, in the real world, it isn't significant. I don't think there's anything useful to prove here, at least not to me - as I said, I really don't know what I'm talking about.

Regards methodology, in post #65, I suggested a blurring to remove the noise artefacts when comparing to see differences that were not masked by noise, but later - in post #67, I realised this would remove the actual visual benefit of the extra images.

I think Allan suggested he couldn't see a benefit that can't be achieved in PP in some other way - and he might be right there; the tiny differences I see pixel peeping your two images might well be imitated with just a little additional Clarity in ACR/LR and/or very fine radius sharpening applied to the '2 pic' version.

I also agree with you that 3 is better than two (data points/images).

Yes, I am just 'splitting hairs'

What I was planning to do was to shoot under 100% controlled conditions, i.e. set up a static subject and shoot it under studio light conditions. I was planning to fix ISO, aperture and shutter speed and vary the lighting in one stop increments. My lights have a 9-stop range (2.5 W-s to 640 W-s), By double banking (two studio lights) and by using ND filters and can work out a combination of well over 20 stops of exposure.

Unfortunately, I have a major project (painting the living room, dining room and front hall) on right now, so getting around to the experiment isn't going to happen over the next few weeks at least.