High Frequency Separations - Nice try but No Cigar

[cathodeRay]

Following from my recent burning and dodging post, and in particular whether PL can do something similar to the SW that Cannot Be Named method for generating the HF layer (Image > Apply Image LF layer to original: 8 bit: subtract, scale 2, offset 128; 16 bit: invert, add, scale 2, offset 0) which appeals to me at least on a theoretical level as a logical subtract (adding an inversion being in effect a subtraction) then add back for the final image sequence, I have established that once again PL can do anything the SWCBN can do - but something isn't right as the results are not entirely as expected. The images below show the 8 bit and 16 bit methods and the whole thing is based on a Draw Isolated group, much of which I hope is evident from the images. For the 16 bit version, I did the inversion explicitly, and used Linear Dodge (Add) as the blend mode ('Linea...'), having confirmed in other experiments that Linear Dodge (add) does really appear to be a simple (a+b) blend. The 'green funnel' layer is a custom filter which does the scale/offset bit - in both cases a 3x3 matrix of ones, div=2 and for the bias 128 of 8 bit and 0 for 16 bit. This should replicate the Image > Apply Image scale and offset settings.

As can be seen, both methods produce a HF layer, but neither is right. In each case, adding them back produces a vile image, one suggesting the girl needs to seek medical attention, the other that she needs to seek urgent medical attention so I am not in a hurry to post them! The original image has been used in a number of online FS toots (so there are benchmark separations available online) and is available here for anyone who wants to have a go.

I have as it happens come across a suggestion that given the same r (size/radius) value, mathematically High Pass filter = Original - Gaussian Blur filter, so HP + GB = Original so maybe my original appeal to 'no data loss' isn't so important, but then there may be times when r((HP) doesn't equal r(GB) for some reason or other and in any event I have now got the bit between my teeth on this, not to mention it is always fun finding new ways and means in PL. I've even come across a SWCBN Channel Mixer method for doing the separations that says it provides a dynamic HF layer (tweak the LF layer, HF layer adjusts in response, if I understood it right) but cracking that one will have to wait for another day. In the meantime, any thoughts why these HF separationss (8 bit first, 16 bit below) aren't quite there yet?

8 bit HF Sep.jpg

16 bit HF Sep.jpg

Thanks in advance for any thoughts

cathodeRay

[bkh]

Following from my recent burning and dodging post, and in particular whether PL can do something similar to the SW that Cannot Be Named method for generating the HF layer (Image > Apply Image LF layer to original: 8 bit: subtract, scale 2, offset 128; 16 bit: invert, add, scale 2, offset 0) which appeals to me at least on a theoretical level as a logical subtract (adding an inversion being in effect a subtraction) then add back for the final image sequence, I have established that once again PL can do anything the SWCBN can do - but something isn't right as the results are not entirely as expected. The images below show the 8 bit and 16 bit methods and the whole thing is based on a Draw Isolated group, much of which I hope is evident from the images. For the 16 bit version, I did the inversion explicitly, and used Linear Dodge (Add) as the blend mode ('Linea...'), having confirmed in other experiments that Linear Dodge (add) does really appear to be a simple (a+b) blend. The 'green funnel' layer is a custom filter which does the scale/offset bit - in both cases a 3x3 matrix of ones, div=2 and for the bias 128 of 8 bit and 0 for 16 bit. This should replicate the Image > Apply Image scale and offset settings.

The main problem here is the allowed range for the colour values (0..255 in 8 bit). If you work with 32 bits of colour depth, everything will be very straightforward (LF layer is by Gaussian blur, HF layer is original - blurred, adding it back using LinearDodge(Add) will give you the original, note that the "divide by 2" layer is superfluous). Well, this isn't perfect, because you can't see the negative values in the HF layer.

The way to work around this is to offset the negative vaules (by 128, as in your recipe). To compute the HF layer from the LF layer, you'll have to make a copy of the original and divide the values of the original by 2, then add 128. Then divide the values of the LF layer by two, and subtract the results from the modified original layer. The result is your HF layer, change that to Linear Light to add LF + HF to get back the original.

Dividing by 2 and adding 128 is most easily done by a histogram correction where you set the output range to 128..255, dividing by 2 is the same as setting the output to 0..127.

I've included a sample file where I have created the histogram adjustments as adjustment layers so it's easier to see how it's done. Of course, you'll have to merge the HF layer before you can make adjustments.

Alternatively, if you want to use a high pass, you'll have to divide the layer by 2 before applying the high pass, and add the high pass layer using "Linear Light".

If I remember correctly, there are also actions somewhere in the forum which create the layers for you.

Cheers

Burkhard.

[cathodeRay]

Burkhard - very interesting and enlightening (sic). I normally work at 16 bits, and only included 8 bit partly in case anyone had to work at 8 bits and also to try and make sense of the layer maths (subtract = 'add an inversion' as in a - b = a + (-b) etc). Apparently, if I've even vaguely got my head round it, this jiggery pokery comes about because 8 bit = unsigned 0-255 (with no whole number mid-point, and you can't have half a pixel), 16 bit = signed (ie can be either plus or minus as actual value) with a range from approx -32k to +32K, with a real meaningful mid-point Being mid/neutral grey of 0 (zero).

I've just run your histogram approach (adjustment layer, set output to 127-255) at 32 bits (thanks for answering another question I had/test to run: are virtual duplicates OK - they are) and got a very clean and perfectly working subtraction based layer stack. Now, one of the clever things here is because the blurred LF copy is a virtual copy, it mirrors changes in its 'parent', so if we now make the GB a child adjustment layer to the first (full) copy of the original (the virtual blur copy), we now have a fully dynamic non-destructive fully editable at any time perfectly working LF/HF separation pair. It is almost magical watching the HF layer adjust as you tweak the blur on the LF layer! Image attached below showing the full stack. The layer blend modes are all Normal except for the LF Virtual Copy which is Subtract, and the top ('name' - I should have renamed it 'HF') Group layer (which is the HF layer deliberately left visible in the screengrab below) which is set to Linear Light. If the LF (Image) layer visibility is turned back on, we get back to the original. As before, the top Group layer has been set to 'Draw Isolated'.

At a stroke PL has overtaken the PS voodoo complexity of using the blue channel via the channel mixer to hold HF data. Instead, we have a simple visible and comprehensible but at the same time dynamic stack. That is not to decry the hard and clever work done by those still stuck to PS, more to underline the flexibility of PL which makes it so adaptable. Brilliant! Now, where's that big cigar?

32 bit HF Sep.jpg

cathodeRay

[bkh]

16 bit = signed (ie can be either plus or minus as actual value) with a range from approx -32k to +32K, with a real meaningful mid-point Being mid/neutral grey of 0 (zero).

Just a remark: 16 bit is still unsigned in PL, so subtraction still causes the same problems as in 8 bit (and you can use the same adjustments that work in 8 bit, PL translates them correctly for 16 bit, multiplying, as usual, by 257). Afaik, PS uses signed ints, but I don't know if negative values after a subtraction are possible. But even if that works, there's still the problem that you can't see those values below 0.

Btw., if you are up to experiments, try "Adaptive Soften" instead of the Gaussian blur – in my experience, that's much better for separating small detail from coarser structures (which is what one really wants).

Cheers

Burkhard.

[cathodeRay]

Burkhard - thanks for the further pointers about the underlying maths. Others (Hoogo?/Ken?) mentioned 'Adaptive Soften' and since the above stack is so flexible it took but a moment to modify the blur layer adjustment layer from GB to AS but it seems to me it might retain too much edge detail (the right eye eyelashes are a particular case in point, even with the AS threshold set very high - high because low values are meant to retain more detail). But no worries - we can experiment to our heart's content because all this is happening non-destructively. I should perhaps mention, for anyone who cares to follow this method, that any actual retouching should also be done on non-destructive (extra, not shown in the above stack) adjustment layers.

cathodeRay

[photoken]

Excellent work, guys! I especially like the idea of using Virtual Copies.

Any chance of getting an action to create the layers?

(I had used Adaptive Soften for an example image because, in that case, it did a better job of retaining the shape of the highlights in the eyes -- the reflection of a window.)

[cathodeRay]

Ken - thanks for the compliments. Whilst I've done a few actions of my own for my own use, I'd be rather worried at this stage of my PL experience of putting one in the public domain lest it do something horrible to someone's precious images. Maybe Burkhard will do one, but if he doesn't, it really is very simple setting up the layer stack - about a minute and a bit if you are not racing at it, less if you are. I think all the key steps are in the posts above but if not what I can do is put together a verbal (written) recipe.

The beauty of this stack is it is so flexible. I was able to swap the GB adjustment for an AS adjustment layer in seconds and the whole thing updates in real time and so you can decide there and then whether it is an improvement (this time I didn't think it was, and so swapped back to GB, again in a few secs), but I am sure there will be times when AS is better than GB.

cathodeRay

[Hoogo]

Excellent work, guys! I especially like the idea of using Virtual Copies.

Any chance of getting an action to create the layers?

(I had used Adaptive Soften for an example image because, in that case, it did a better job of retaining the shape of the highlights in the eyes -- the reflection of a window.)

I posted an action lately in this thread with the old picture, that seems to fit your needs. I remember that it was somehow not very practical and I wanted to add something, but meanwhile I forgot what I was doing there...

[photoken]

... what I can do is put together a verbal (written) recipe.

Yes, that would be great if you would.

Since this layer stack is new to me, it would be a matter of trial-and-error to ensure I got things right. That's not a good way to record an action....

[cathodeRay]

First off, I should say that while I opened this thread, with a question, it was Burkhard who provided the answer here, and of course there is the earlier work by Hoogo and Ken and others, so the usual only see further by standing on the shoulders of giants applies to what follows.

OK, here goes. This is a step by step live record of what I did to create an identical dynamic stack to the one above (plus a visual check layer added at the end). Obviously it took a bit longer than a minute to do as I was making notes, but since the steps are logical enough and not too complex they can almost be memorised and done very quickly, or indeed an action recorded for future use (BTB my slight concern about actions is their slightly lively behaviour and lack of transparency, as in there isn't a script you can open and view/edit in a text editor. Actions are in effect macros and I am more used to something like VBA where you can jump in and directly edit the code and/or write the code from scratch). Abbreviations should be obvious eg FS = frequency separation, HF/LF = high/low frequency, GB = Gaussian blur etc. Most of the steps are easily done via buttons at the bottom of the layers panel and/or context menus. Lots and lots of words to try and get it all down in the finest detail, but actually really quite simple:

1. Open the image you wish to add the stack to. This is your background layer and will not be edited. Name it 'Original'.

2. Convert to (RGB) 32 bit via layer panel. This step is probably crucial, because it ensures the 'right sort of maths' happen during the blends. Going to 32 bits will substantially increase file size, but against that we are going to use virtual copies in a moment, and that will help decrease file size.

3.Duplicate the layer: rename it from Original (1) to LF (or whatever you want to call the LF layer). This will be your LF layer, but don't so anything to it yet (you will add a blurring adjustment layer very shortly).

4. Make Virtual copies of both the Original layer and the LF layer. These will both helpfully be marked Virtual Copy. Change the name from Original (1) and LF (1) by stripping off the (1) (you know they are copies because under the name it says so).

5. Drag to Original Virtual Copy until it is directly under the LF Virtual Copy.

6. Check the tick box on the top two Virtual Copies layers and Group them. An new layer will appear, which is the group (helpfully labelled as such). Rename the Group layer HF (I didn't do this in the version in the screengrab above).

7. Using the layer panel, turn on (check) the Draw Isolated box. This means the group members will 'do their own thing', and send the result to the group layer.

8. Using the arrowhead, expand the group.

9. Add a new Histogram adjustment layer. This will appear at the top of the stack. Its associate dialog/settings box will also open. As it is open it is convenient to do one of the crucial steps at this time: set the Output levels (the ones we don't often touch) to from 127 to 255. Don't change anything else. This is a short cut to doing some of the maths, so you don't have to do the calculations explicitly. It remaps 0-255 to 127-255, ie black-white to mid grey-white. Don't worry about how the stack thumbnails look at this stage.

10. Drag the HG adjustment layer to be directly over the group layer (its outline will match that of the group layer, and drop (release the mouse button). It will become part of the group, the top layer of the group.

11. Add another adjustment layer, this time using your chosen blur method (I used GB). It will appear somewhere in the stack, quite possibly next to the HG adjustment layer. As the GB dialog/settings box will open (and we are lazy or rather prefer to be efficient), use the opportunity to set an initial blur ((size/radius for GB) value, say 5 (it's not critical we can change it later). Drag it to be over the LF Image (not Virtual Copy) layer. It will become a child (I think that is the right term) adjustment layer to the LF Image layer. Continue not to be distracted by how the stack thumbnails and main image look.

12. Now its tome to work the magic. We are going to subtract the LF Virtual Copy (which is now mirroring the LF Image layer) from the Original Virtual Copy, to get the HF layer delivered, via the HG adjustment layer, to the group layer. All we need to do to achieve this is select the LF Virtual Copy layer, and set its blend mode to Subtract. If all is well, an obvious HF layer will appear at the top of the stack, obvious in the thumbnail (if your thumbnails are large enough, and in the main image window if you have left all layers' visibilities on throughout.

13. Take a moment to assess the HF layer. If you want to adjust it, double click the LF Image layer adjustment (in this case GB) layer icon to open the layer dialog/settings and tweak the blur. The HF will magically adjust in real time (if it doesn't, make sure the Preview in Document button (bottom LH) is down/on). Close the dialog when you are happy enough with the HF layer.

14. Now its time to put it all back together, and (optionally, but sensibly) perform a visual check that all is happening as it should. Start by selecting the bottom BG layer (the known unchanged original) by checking the LH tick box for it (and only it - uncheck any others as need be.

15. Again do a Virtual Copy of the BG layer. Move it to the top of the layer stack, rename it Visual Check and turn off its visibility (for now). You will now again be looking at the HF layer.

16. All we need to do now to get back to the original image is blend the HF group layer with the LF image layer (the one near the bottom, with the blur adjustment layer) using Linear Light as the Blend mode. Make sure the HF layer (and only it) is selected (one of the reasons for things going awry is that it is not too difficult to accidentally set the blend mode for the wrong layer) and set the blend mode to Linear Light.

17. Bingo! Your original image (or rather the blend copy formed from the LF/HF layers appears! To check it really is the same as the original image, toggle on & off the visibility of the Visual Check top layer of the stack (we put it there to make this easy). If all is well, it and the underlying layer will be visually identical. You will not be able to tell one from the other. If you are an*lly inclined, you could even use a quick Difference blend mode to really really confirm there really is no difference. Just make sure you set the Difference blend mode on the right layer.

18. Proceed to do any FS LF/HF layer editing. While you could do this directly to the LF/HF layers, it makes sense to add adjustment layers as required. Lets say you want to do some cloning (copy brush)/healing on the LF layer. Add a new blank transparent layer and place it directly above the LF Image layer. Make sure that new layer is the active layer (check box ticked). Depending on which brush you are using yon can either just jump right in (the healing brush) or first set the source to be 'until active layer' for the copy brush. For effects/brushes that can't work (jump across) layers, you may do better to duplicate the LF Image layer (rename the duplicate LF Working Layer) and work on that (so leaving the 'original' LF intact in case you want to go back to it. The same applies to working on the HF layer: some tools will allow working on an added blank transparent layer; others need a working copy. For the HF layer, to get an working image copy of the HF layer alone, you need to (yes it's a bit fiddly): (a) turn off the visibility of the lower layers (but leave the blur adjustment (not image) layer visible (otherwise you turn the blur off), (b) select the HF layer and do a merged copy (so it picks up the group effects but becomes one image/layer; if you just copy the group you will, err, get a group) and then paste as layer (and rename HF Working Layer and use it in the final Linear Light blend after doing any editing.

I have to confess I did think I read somewhere that you could use things like the brighter/darker brush non-destructively on a blank transparent layer but today that eludes me, hence the use of copy layers detailed above. But it's still an 100% non-destructive workflow, as in you can go back quickly and easily to where you were, before you so disastrously over-egged the pudding.

Hope that helps!

cathodeRay

[photoken]

cathodeRay,
Thank you very much for the description of the steps. I'll work with this later, when I have some time.

This technique raises some interesting questions that go a little further afield than this discussion, so I'll start a new discussion about them.

[bkh]

I've modified Hoogo's action (thanks!) slightly, to use even more virtual copies and added some transparent layers for retouch of low and high frequencies (the HF layer is in the top group). As usual, remember to use "Merge" when loading the action into PL (or just drag and drop it).

If you like, you can replace the Gaussian Blur layer (in the LF group) by an Adaptive Soften layer or whatever.

Cheers

Burkhard.

[photoken]

Burkhard,
Thanks for the action, but the "Retouch Low Frequencies" layer should be part of the "Low Frequencies" group to make the layer stack consistent....

Also, cathodeRay mentions that the first thing to do is convert the original image to 32-bit, but your action does not do that. Is converting to 32-bit necessary?

[bkh]

Burkhard,
Thanks for the action, but the "Retouch Low Frequencies" layer should be part of the "Low Frequencies" group to make the layer stack consistent....

That would be nice from a user perspective, but it doesn't work (because then the content of the retouch layer would be incorporated in the computation of the HF layer). Otoh, the other retouch layer needs to be in the group if you want to use the copy brush, remove object brush or similar. EDIT: I thought I had tested this, but for some reason, the HF retouch layer doesn't work as expected (see also http://www.pl32.com/forum3/viewtopic.php?f=3&t=5126).

Of course, it would be possible to group the LF layer and the LF retouch layer, but that would be another level of grouping.

Cheers

Burkhard.

[cathodeRay]

Also, cathodeRay mentions that the first thing to do is convert the original image to 32-bit, but your action does not do that. Is converting to 32-bit necessary?

I based this statement on Burkhard's earlier remarks above (first reply to my original post) including "If you work with 32 bits of colour depth, everything will be very straightforward" and I'm all for "very straightforward". Elsewhere (they got too complicated to reproduce in full here) I made some notes about the ways the various bit depths are handled mathematically eg signed/not signed (has implications for addition subtraction eg if it is signed then a + (-b) is actually a sort of subtraction (this is how I think the PS uses an addition (of an inversion) to do a subtraction and yes I get very confused too); where the actual mid point is (you can't have half a pixel eg at 8 bit mid grey should be 127.5 but instead has to be 127 or 128) and partly this is why the PS (HP sucks) methods get so obtuse (scales here and offsets there). 32 bits 'just works'. Of course the simple thing to find out whether other bit depths work is to do the procedure at those bit depths (I've yet to do that...).

The other useful thing I have picked up from Burkhard is the idea of using Histograms (and I'm sure Curves would do the same thing, and more) to do calculations: instead of a calculation/custom filter (and in PS the Image > Apply Image settings) with cryptic entries, you have a visual/graphical mapping of values. To my non-mathematical mind this is much easier to understand, and so see what is going on.