Histogram / RAW conversion comparison [Archive] - The-Digital-Picture.com Community

neuroanatomist

07-28-2010, 09:17 AM

Sure, I believe it. But it's not a valid comparison, I think. You're actually looking at different types of histograms, comparing apples to oranges so to speak.

The DPP histogram you're looking at (on the RAW tab) is pixel luminance represented on a base-2 logarithmic scale (meaning each 'unit' in that -10 to +4 range is a stop of light, i.e. 1 EV, as you can see when you use the brightness adjustment slider). The ACR histogram is an RGB histogram on a linear scale (0-256 for an 8-bit image, 0-4096 for 12-bit, or 0-16384 for 14-bit). Log vs. linear are very different ways to represent data.

If you want to compare the same type of histogram in the two programs, use the DPP histogram on the RGB tab instead of the RAW tab - you'll find that one looks quite different from the RAW tab histogram, but much more like the ACR histogram (without the secondary colors - cyan/yellow/magenta - superimposed on the display).

canoli

07-28-2010, 11:04 AM

Ah - thanks! and thank goodness there's a reasonable explanation, even if the answer only adds more confusion...

base-2 logarithmic scale (meaning each 'unit' in that -10 to +4 range is a stop of light, i.e. 1 EV, as you can see when you use the brightness adjustment slider)

I've been through the Help and I see nothing about the histo using a log scale. Am I just missing it? I'd like to make this knowledge useful but I don't understand how it helps me...(I do understand - at least basically - what a log scale is and why it's used)

Are the units you referred to the spaces between the vertical dotted lines?

Moving the Brightness slider to the right, from 0.0 to 1.0, moves the tones to the right but not an entire "unit" to the right.

Does the first 3 spaces on the left being so much skinnier than the rest of them have anything to do with anything? Why are they so much narrower? Why is one of the dotted lines in bold? And what's with the gray sections at either end? Are they aesthetic only? But then why are 3 of the units in the gray sections?

Better yet - why can't I find the answers in Help?? Arrrgh!

Thanks for your help!

neuroanatomist

07-28-2010, 12:53 PM

I see nothing about the histo using a log scale. Am I just missing it? I'd like to make this knowledge useful but I don't understand how it helps me...

I think the rationale for using a base-2 log scale here is simply to convert the luminance data to 'stops' - i.e., a change of one unit means a doubling of the intensity of that pixel.

Are the units you referred to the spaces between the vertical dotted lines? Moving the Brightness slider to the right, from 0.0 to 1.0, moves the tones to the right but not an entire "unit" to the right

The units are the numbers along the bottom; the dashed lines are every two units. So, changing the brightness sliderfrom 0.0 to 1.0 for your image will move the largest peak on the histogram from about -1.0 to about 0 (a change of one unit).

Does the first 3 spaces on the left being so much skinnier than the rest of them have anything to do with anything? Why are they so much narrower? Why is one of the dotted lines in bold? And what's with the gray sections at either end? Are they aesthetic only? But then why are 3 of the units in the gray sections?

The spaces on the left aren't narrower, the dotted lines are still every two units. The solid line at -9.0 isn't a division, that's the black point of the histogram; it's actually a moveable slider - you can drag it, and when you drag it into the portion of the histogram with pixels you'll see the contrast of the image change.

The dotted line at 0.0 is bold because that's the 'origin' - in this case, I believe that zero point is middle (18%) gray as the camera would meter it.

The gray zones at either end aren't just aesthetic - they represent areas of the histogram that fall outside the dynamic range of your camera's sensor. You shot this with an original 5D, correct? That has a 12-bit sensor, while newer cameras have 14-bit sensors which gives them a wider dynamic range. You can see that represented in a comparison of the DPP RAW histogram for a 40D (14-bit) vs. 5D (12-bit):

http://www.lsfoto.net/technical/14-vs-12-bit/images/histogram.jpg

Hope that provides some clarity - I have not idea why nothing along these lines are in the help documentation.

Johnny Rasmussen

07-28-2010, 02:20 PM

neuroanatomist,

Does your 40D really go all the way beyond -10EV ?

That´s strange. My 50D goes to -9EV and so does all the latest EOS models. I know this is a technical representation of the max dynamic range the sensor is capable of, defining "usable" DR below the gray point is another subject.

canoli

07-28-2010, 02:22 PM

Thanks! Very helpful indeed. DPP's Help really doesn't do that program justice.

Actually it's a 1DsIII capture (24-105 @ 28mm, 1/60 @ f/6.3 ISO 1600) inside B&H yesterday. Correct me if I'm wrong but I think that uses a Digic III (or maybe two of them?) at 14 bits/channel. In any case I understand now what the gray areas are for - almost.

The "almost" is this: your screen grab shows the wider DR of the 40D added (only) at the low end. Is that a "real" reflection of where the added DR is (from a 14-bit vs a 12-bit sensor)? So the 40D should be better than the 5D at capturing noise-free detail in the shadows but not any better at holding HLs?

Johnny Rasmussen

07-28-2010, 02:37 PM

canoli,

If you turn on HTP you will see a shift in the histogram, the highlight end will move up 1EV to around +4.6EV above middle gray, giving you about one stop better highlights, but at the same time the bar marking the shadow end will also move one stop up from -9EV to -8EV below middle gray.

neuroanatomist

07-28-2010, 02:42 PM

neuroanatomist,

Does your 40D really go all the way beyond -10EV ?

That´s strange. My 50D goes to -9EV and so does all the latest EOS models.

Sorry, but it's not my 40D. [:$] (I have a 7D, and had a T1i before that).

I pulled the image from this article comparing 12-bit and 14-bit images ("http://www.lsfoto.net/technical/14-vs-12-bit/).

But, in DPP with images from my 7D, there is no gray space on the left side, i.e. it's white all the way over, past -10.0. In fact, I may be wrong and this may not even be an approximation of sensor dynamic range, but rather just a function of the version of DPP - from some other articles online (especially one by Doug Kerr), it seems that an older version of DPP went from -9.0 to +3.5. I also ran across a post indicating that an even older version of DPP (v1.6) had actual luminance values on the scale as well as the EV numbers - those numbers ranged from 0-4095 (which makes sense for a 12-bit image - 2^12 = 4096).

neuroanatomist

07-28-2010, 02:52 PM

If you turn on HTP you will see a shift in the histogram, the highlight end will move up 1EV to around +4.6EV above middle gray, giving you about one stop better highlights, but at the same time the bar marking the shadow end will also move one stop up from -9EV to -8EV below middle gray.

Interesting. Just to add further confusion, in an image with blown highlights (lit up by the highlight alert), the histogram shows a 'cliff' at the right side, just a little to the right of +3.0, with a clear gap between the cliff and the white/gray border line at the right. But, when I increase the Brightness adjustment, it pushes that cliff even further to the right, so the histogram moves into the gap, eventually 'piling up' against the white/gray border line.

Hmmmm....

Johnny Rasmussen

07-28-2010, 02:56 PM

canoli,

I think this is an approximation of the sensors maximum DR. I have been looking at it lately and the numbers makes sense when you compare them with the numbers from the dxomark sensor measurement site. A total of around 11EV to 12EV for many of the latest EOS models.

I have downloaded raw files from Imaging-Resource and most models show only small variations in the highlight end. HTP is shifting the entire histogram to the right, about 1EV in both the shadow and the highlight region.

If all you see is white there may be something wrong with the installation? 7D files I have looked at are similar to other models.

Johnny Rasmussen

07-28-2010, 03:01 PM

Sorry, the comments above was to neuroanatomist :-)

neuroanatomist

07-28-2010, 03:31 PM

If all you see is white there may be something wrong with the installation? 7D files I have looked at are similar to other models.

What version of DPP are you using? I do see other screen captures that look like what I see - here's one posted on birdphotographers.net ("http://www.stanford.edu/~ahazeghi/Photos/birds/hist.jpg), for example.

I also ran across this in a tutorial video from B&H ("http://www.video.bhphotovideo.com/?fr_story=6705482e45f83bfb581744e6bbf1bee3836e657f &rf=sitemap) - starting at 52:28 he discusses the DPP RAW histogram (and why you should just leave it alone, in his opinion). He has gray space on the left of his histogram and is showing DPP 3.4. You'll also notice that there are only 4 sliders below the histogram. I'm using DPP 3.8.2 and there are 6 sliders on mine (as there are in the screenshot I linked above).

Johnny Rasmussen

07-28-2010, 03:56 PM

neuroanatomist,

I am using v3.8.1, the latest windows version. Earlier versions only had 4 sliders, the two new ones are the sliders for highlight and shadow correction. They are not very useful in my opinion, Lightroom or ACR are better for pulling back highlights or shadows, but I still use them occasionally.

neuroanatomist

07-28-2010, 04:16 PM

Odd! Although I'm using the Mac version, the screenshot I linked is Windows. Do you have a link to a 7D RAW file that shows up on your DPP with the white region of the histogram starting at -9.0?

Johnny Rasmussen

07-28-2010, 05:04 PM

/cfs-file.ashx/__key/CommunityServer.Components.UserFiles/00.00.00.22.85/7D-HTP-turned-off.JPG/cfs-file.ashx/__key/CommunityServer.Components.UserFiles/00.00.00.22.85/1Ds-Mark-III-HTP-OFF.JPG/cfs-file.ashx/__key/CommunityServer.Components.UserFiles/00.00.00.22.85/40D-HTP-OFF.JPG/cfs-file.ashx/__key/CommunityServer.Components.UserFiles/00.00.00.22.85/50D-HTP-OFF.JPG/cfs-file.ashx/__key/CommunityServer.Components.UserFiles/00.00.00.22.85/50D-_2800_same-scene_2900_-HTP-ON.JPG

From top to bottom, from left to right:

7D with HTP OFF, 1Ds Mark III with HTP OFF

40D with HTP OFF, 50D with HTP OFF

50D with HTP ON (same scene)

Johnny Rasmussen

07-28-2010, 05:42 PM

PS! The 7D file can be found here http://www.imaging-resource.com/PRODS/E7D/E7DFARI00100.CR2.HTM ("http://www.imaging-resource.com/PRODS/E7D/E7DFARI00100.CR2.HTM) other raw files can be found here http://raw.fotosite.pl/ ("http://raw.fotosite.pl/) but this site seems be having problems right now.

neuroanatomist

07-28-2010, 09:39 PM

The 7D file can be found here http://www.imaging-resource.com/PRODS/E7D/E7DFARI00100.CR2.HTM ("http://www.imaging-resource.com/PRODS/E7D/E7DFARI00100.CR2.HTM)

FWIW, for me, that file shows up in the DPP histogram like this, with no gray area at the left:

/cfs-file.ashx/__key/CommunityServer.Components.UserFiles/00.00.00.35.15/7D-RAW-histogram.jpg

Johnny Rasmussen

07-28-2010, 09:56 PM

Mac users....you guys just have to do things different...[:)]

Seriously, it is not important but still a little strange that Mac users can´t read the DR from raw files. Guess it is a design decision specially made for Apple.

neuroanatomist

07-28-2010, 10:11 PM

Mac users....you guys just have to do things different...

Seriously, it is not important but still a little strange that Mac users can´t read the DR from raw files. Guess it is a design decision specially made for Apple.

Well, yes, we 'think different' [:P] ...but that's not the reason for this. Here's that screenshot I linked earlier...now, I may be a Mac user by choice, but I use Windows when 'forced' to...and I know a Windows window when I see one.

http://www.stanford.edu/~ahazeghi/Photos/birds/hist.jpg

Got any other ideas?

Johnny Rasmussen

07-28-2010, 10:41 PM

neuroanatomist

07-29-2010, 08:51 AM

Hmm, that is odd. The only other thing I can think off is that there may be something wrong with the installation files. I know it sounds unlikely but I can not think off any other reason. Have never seen this before...I remember my wife had a problem with DPP a couple of years ago. The RAW tab or the RGB tab menus did not work properly. I had to do a clean install with "fresh" files from Canon to get it to work.

A clean install is a good idea, and might even work! So, if you decide to try it, let us know if that does the trick. [:P]

Actually, I'm serious (for once). Assuming we are correct in thinking the white area of the plot shows the theoretical dynamic range of the sensor, I suggest that you try it because I believe DPP on my computer is showing the correct data and DPP on your computer is not behaving correctly, for two reasons - one logical, and one empirical.

The logical reason goes like this: if we're correct that the 'white area' on the DPP histogram represents the theoretical maximum dynamic range of the camera, then an image from a camera with a 14-bit analog-to-digital converter (like the 40D)should have awider white area on the graph than an image from a camera with a 12-bit ADC (like the original 5D). (You can confirm the bit depths here ("http://photonotes.org/cgi-bin/camera-lookup.pl?camera1=eos5d&camera2=eos40d&camera3=*&c amera4=*&camera5=*).) So, the image below (not mine, as I said) would seem to be correct:

http://www.lsfoto.net/technical/14-vs-12-bit/images/histogram.jpg

The empirical reason goes like this: I downloaded a Rebel XS/1000D RAW file ("http://www.imaging-resource.com/PRODS/XS/XSFARI0100.CR2.HTM) from Imaging Resource, and the equivalent 7D RAW file ("http://www.imaging-resource.com/PRODS/E7D/E7DFARI00100.CR2.HTM) (thanks for the link above!!). Opening both in DPP 3.8.2 on my computer, I see this (XS on the left, 7D on the right):

/cfs-file.ashx/__key/CommunityServer.Components.UserFiles/00.00.00.35.15/1000D-RAW-histogram.jpg /cfs-file.ashx/__key/CommunityServer.Components.UserFiles/00.00.00.35.15/7D-RAW-histogram.jpg

It seems that my DPP RAW histogram shows a 2-unit wider dynamic range for the 7D (14-bit) than the Rebel XS (12-bit), starting at -11 vs. -9. The 2-unit difference on a base-2 log scale corresponds to the 2-bit difference between the ADCs of the two cameras [since log2(2x) = x].

However, the screenshots you posted from 14-bit cameras seem to be showing the same dynamic range as the 12-bit Rebel XS in DPP on my computer - that's what makes me think yours is not displaying correctly. If you're willing, try downloading the same two files linked above, and see what your computer shows (maybe there will still be a 2-unit difference, offset to -9 vs. -7 in your case, and we'll be back to square one!).

I hope the above makes sense. Thanks for taking the time to stick with this, Johnny!

--John

Johnny Rasmussen

07-29-2010, 05:11 PM

John,

You could be right of course. What I am seeing could be wrong for some reason.We know there is a huge technical difference between a 14-bit and a 12-bit file.Does it show ut in DPP...that is the question?

I downloaded both files from the link you gave mebut the DR is the same. I even checked my wifes computer and it is the same. She is using the previous DPP version, 3.8.0 for Windows.Then I downloaded the DPP manual for Mac, v3.8.2 and grabbed a screendump. As you can see itis different from what youare seeing, showing a DR from -9EV to around +3.6EV.

/cfs-file.ashx/__key/CommunityServer.Components.UserFiles/00.00.00.22.85/DPP_2D00_3.8.2_2D00_manual_2D00_for_2D00_Mac_2D00_ page_2D00_54.jpg

Johnny Rasmussen

07-29-2010, 05:24 PM

PS! Does the 7D file have the vertical bar to the left like the XS file did? If so can you drag it to the right and adjust DR?

neuroanatomist

07-29-2010, 05:56 PM

As you can see itis different from what youare seeing, showing a DR from -9EV to around +3.6EV.

True. I see that in the manual. But then again, I see that when I open up a 12-bit image in DPP v3.8.2. We don't know what camera the image used for demo purposes came from (in the section on viewing the shooting information, whoever wrote the manual specifically redacted the camera model/image size/etc. - it's a crop body, though, since it's using an EF-S lens, and it goes to ISO 3200 so it's not an XTi, but it could be a 20D, which is 12-bit).

We know there is a huge technical difference between a 14-bit and a 12-bit file.

Certainly. Canon even published an article on the benefits of 14-bit A/D conversion ("http://www.usa.canon.com/dlc/controller?act=GetArticleAct&articleID=2748).

Does it show up in DPP...that is the question?

It seems to...at least for me. RAW histograms from other 12-bit cameras (5D, XTi/400D) show up as -9 to +3.5 just like the Rebel XS/1000D, while RAW images from other 14-bit cameras (5DII, T1i/500D) show up with the wider dynamic range of -11 to +3.5 just like I see with the 7D.

I think if you don't see a difference between the two image types in the DPP RAW histogram, that's not correct.

PS! Does the 7D file have the vertical bar to the left like the XS file did? If so can you drag it to the right and adjust DR?

Yes, that moveable bar just initially overlaps with the left axis of the graph - hovering the mouse over that axis changes the cursor to a bi-directional arrow. The line can be dragged in from the left edge to adjust the DR.

Johnny Rasmussen

07-29-2010, 06:51 PM

neuroanatomist

07-29-2010, 08:14 PM

Interesting. I am beginning to think thatyou are seeing the true capability of the 14-bit sensor. Do you have some images from your own camera shot with HTP on? That should move both vertical bars 1EV to the right and give you -10EV (theoretical shadow DR) to +4.5EV.

I don't usually shoot with HTP, since the highlight preservation comes at the expense of increased shadow noise and the 7D doesn't really need any 'help' in that department. But, I just grabbed a couple of test shots -/+ HTP - the desk lamp against the closed curtain was intended to result in a histogram with peaks at both ends of the DR. The one on the left is without HTP, the one on the right is with HTP.

/cfs-file.ashx/__key/CommunityServer.Components.UserFiles/00.00.00.35.15/7D-RAW-_2D00_HTP-histogram.jpg /cfs-file.ashx/__key/CommunityServer.Components.UserFiles/00.00.00.35.15/7D-RAW-_2B00_HTP-histogram.jpg

As you can see, the DR 'window' does shift 1 EV to the right with HTP, compressing the shadow end and giving some additional headroom on the highlight end. It's worth noting that in the non-HTP image of this contrived scene, there are pixels in the RAW data that span the full dynamic range represented in the DPP histogram.

As an interesting side note, parts of this simple image can be easily resolved on the histograms - the peak at the far right is the bulb, and the next one just to the left is the lamp shade. The big mushy peak on the left is the curtains. It's interesting that there is a series of peaks in between (-3 to -1 EV), seen only on the +HTP histogram. They might represent the forward folds of the curtains that are brighter than the backward folds, and get progressively less intense further from the light - but it's certainly not an obvious visual difference between the images, so it's interesting that HTP resolves them. I wonder what that says about the algorithm. But then again, those peaks may be an artifact of the processing, which is probably intended for 'real world' scenes and is here being applied to something pretty artificial. Some noise reduction algorithms can actually create patterns in a blank region of an image. One of those peaks is at -1 EV, and there's nothing in the image that stands out to me as a patch of something one stop darker than middle gray - that sort of supports the artifact idea, but really, I have no idea what's going on there.

Regardless, in any case, HTP does appear to offset the DR to the right for me.

I still don´t understand the differences between the rendering of the histograms. Think I have to ask Chuck Westfall about that.

Great idea! If you ask him and he responds, please let me know what he has to say.

Thanks again!

--John

Johnny Rasmussen

08-03-2010, 10:17 PM

John,

I received an answer. He said that Canon USA are aware that the shadow end appears to be longer on the Mac version of DPP but they have not received a technical explanation from Canon HQ. They say that the actual DR captured by a given sensor is the same even if the representation is different for Mac users. He also said that the white portion of the scale represents the approximate DR of the camera in use and that they do not have detailed information about the scale it´s using, it is referred to as highlights and shadows only. He further said "when attempting to adjust the left end of the histogram on the Mac version, the line for the deepest shadow seems to jump to the same starting position on the scale as the starting position for the Windows version of DPP."

What do you think? Any other ideas? Seems like it will not be easy to get to the bottom of this. I guess Canon HQ has it´s reasons not to publish all the details but I have no clue why.

Do you think this has to do different "country versions" ? When you install a new version you are asked to choose your location, (Europe, Asia, America ). Running out of ideas.

Since it seems like you have the camera and computer with the best dynamic range why don´t you ship it to me? [G] I volunteer to test it. I have a few portraits to make and better DR is always welcome..[:D]

neuroanatomist

08-04-2010, 07:36 AM

Hi Johnny,

Thanks for checking that out with Chuck!

He said that Canon USA are aware that the shadow end appears to be longer on the Mac version of DPP but they have not received a technical explanation from Canon HQ.

I've got one problem with that explanation - it fails to explain that the same histogram as I see is also seen in Windows (at least, I think it's Windows...but it's certainly not the Mac version of DPP!):

http://www.stanford.edu/~ahazeghi/Photos/birds/hist.jpg

He further said "when attempting to adjust the left end of the histogram on the Mac version, the line for the deepest shadow seems to jump to the same starting position on the scale as the starting position for the Windows version of DPP."

I do notice the 'jump' but I still wonder what's going on, since there are clearly pixels at both ends of my wider histogram. Since I see a difference in the width of the active part of the histogram between images from 12- and 14-bit cameras, and you don't, I'm still thinking that what I'm seeing is more accurate.

Seems like it will not be easy to get to the bottom of this.

Agreed.

neuroanatomist

08-04-2010, 08:36 AM

I'm still thinking that what I'm seeing is more accurate.

I was thinking more on this as I drove into work this morning. I do think the representation I'm seeing (where the DR of a 14-bit image is displayed as being wider than the DR of a 12-bit image) is more accurate than what you're seeing, Johnny. But, I should point out that neither is truly accurate, and so the differences are probably irrelevant in a practical sense. As I stated above,log2(2x) = x, so if the scale being used is log2of pixel luminance, a 14-bit image should have a DR of 14 units. The range of -11 to +3.8 (or so) is ~14.8, nearly a full unit wider than it should be for a 14-bit image; likewise, I see a 12-bit image displayed with a range of 12.8 units.

Regardless of the actual units, DPP is not 'throwing away' any of the dynamic range of the RAW image (until it's converted to an 8-bit format like JPG, that is). So, practically it doesn't really matter if the scale is-11 to +3.8,-9 to +3.8, 0-16,384, -20 to +20, or even if there are no units at all. So long as the pixels are binned across a range (and those bins are sufficiently narrow as to allow meaningful adjustments), the histogram serves its purpose.

Johnny Rasmussen

08-05-2010, 09:21 AM

I came across a discussion in the dpreview forum. It seems like we are not the only ones that have been wondering about this. One guy said that the differences seen in the histograms is "canons way of showing us the wider DR of the 14-bit raw-files". He argued that a 14-bit file should be able to show a 2 unit wider DR. I went to Canon Japan, looked at DPP tutorials and they had the same wide DR as your version.

Anyway, you are right, even if the representation of DR differs it does not really matters. I can live with it [:)]

I was thinking about something else, John. You mentioned you did not shoot with HTP on because of increased shadow noise. Is it really that visible? I have not shot with the 7D but it is hard to see any difference in files from the 50D. If I enlarge the image 200% I can see a small difference. Bob Atkins tested the HTP function http://www.bobatkins.com/photography/digital/canon_eos_40D_review_6.html#highlight ("http://www.bobatkins.com/photography/digital/canon_eos_40D_review_6.html#highlight) and particularly looked at the shadows. The link is for the 40D but he has also tested the 50D and the 7D. He said that the differences between these models are insignificant in the shadow region. I know that sites like dpreview criticized the 50D compared to the 40D but I wonder if this is because of the methodology they are using.

neuroanatomist

08-05-2010, 10:18 AM

You mentioned you did not shoot with HTP on because of increased shadow noise. Is it really that visible? I have not shot with the 7D but it is hard to see any difference in files from the 50D.

It's rather situationally dependent. I have seen it in some shots, not in others. I suspect it's a combination of factors. Even Bob Atkins stated (from the article you linked), "Looking closely at images shot with Highlight Tone Priority turned on and off, there is an increase in the shadow noise if you look closely enough." As I stated, I don't usually use it. But, I do use it occasionally (in fact, it's one of the settings I put in My Menu) in situations where looking at the scene I expect issues from blown highlights at the exposure I plan to use.

Daniel summarized it nicely in a previous discussion ("/forums/p/1886/14429.aspx#14429) on 'tweener' ISO settings:

<div>

It's a pretty minor effect by itself. Just like HTP is a pretty small effect, and ALO is a pretty small effect. But when you start combining all these small factors, it results in a big factor. For example, ISO 250 by itself may not increase shadow noise enough to notice. And ISO 200+HTP by itself may not either. But combine 250+HTP and the result will be much more noticeable: the shadow noise is as bad as ISO 640. Add ALO into the mix and you can get people wondering why their ISO 250 shot looks like ISO 1600.
</div>

Daniel Browning

08-05-2010, 01:17 PM

You mentioned you did not shoot with HTP on because of increased shadow noise. Is it really that visible?

It depends on the ISO. HTP works in the context of a constant display brightness and a constant exposure. In that context, low ISO has more noise than high ISO. In other words, when you keep your exposure the same, and the raw conversion is made to come out with the same brightness, higher ISO has less noise (sometimes a lot less) than low ISO. This is the opposite of how we normally think about ISO because we don't normally keep exposure fixed. In essence, it separates ISO-as-a-camera-setting from ISO-as-an-exposure-index.

So ISO 3200+HTP (which is actually just ISO 1600) has the exact same noise as ISO 3200 without HTP, but ISO 200 has less noise than ISO 200+HTP (which is actually just ISO 100).

The way HTP works is pretty simple. Say you shot a scene at ISO 200 f/4 and developed it with default raw converter settings, and it came out with too much blown highlights. Then you shot it at ISO 100 f/4. Now the highlights are better, but the overall image (midtones, shadows, etc.) are too dark. So you push the midtones and shadows without blowing the highlights (e.g. by using a curve). This makes the increased noise of a lower ISO setting visible (you wouldn't see the difference if you left the overal image darker). That is what HTP does.