Heads-up for the LAB proponents

On another forum, I was pointed to this very interesting article.

It discusses the severe quantization that occurs when converting from 8-bit sRGB to 8-bit LAB. In short, you lose about 87% of your levels of colour.

Adobe RGB isn't discussed, it would be better, but I imagine it's still pretty nasty.

The upshot is: Don't ever convert in and out of LAB in 8-bit.

The reason I bring this up is I often see people recommending a conversion to LAB for various editing steps, particularly sharpening on the L* channel. I've always thought this was dangerous advice, and now I'm sure of it. Obviously, it's ok to do it on 16-bit files, but not in 8-bit.

Personally, I've never found any reason to convert to LAB mode, because the "Luminosity" and "Color" layer blend modes simulate it adequately for my needs.

So, to all those people who are habitual LAB converters, I say this: I don't care what you do in the privacy of your own home. You can quantize the crap out of your images for all I care. BUT ... if you choose to advise another POTN member to convert to LAB as part of their editing process, please always add the caveat that it must be done in 16-bit.

Damo77 wrote in post #5981284
On another forum, I was pointed to this very interesting article.

It discusses the severe quantization that occurs when converting from 8-bit sRGB to 8-bit LAB. In short, you lose about 87% of your levels of colour.

Unfortunately, that is an extremely flawed conclusion. The only logical conclusion one can reasonably draw from that rather brief article is that for that specific image - one that bears very little resemblance to an actual photograph - severe quantization occurs when converting from sRGB to LAB mode. However, it does not necessarily follow that the amount of quantization in that one, unique image also occurs in other photographs.

"Disclaimer* I am not one of the people who you derisively describe as "habitual LAB converters." I have also never recommended that anyone convert to LAB at any stage of post processing, the reason being that I don't care one way or the other.

However, it would be exceedingly close-minded to say never convert to LAB in 8-bit mode. And, in fact, that advice contradicts that of acknowledged "gurus" like Dan Margulis - a prepress expert, and Deke McClelland, a graphics artist and best-selling Photoshop author...both of whom are Photoshop Hall of Famers.

Its quite possible that Margulis and McClelland are wrong but the article you link certainly does not make that case. Just this week McClelland released a video titled "Stealth 'Shop, The Virgin Histogram" where he advises converting images to LAB to avoid the honeycombing effect that occurs when certain luminance adjustments are made in Photoshop. Its impossible not to notice that this video takes exactly the opposite stance about LAB mode. Also, he also recently released a video on Lynda.com titled "Photoshop CS3 Mastering Lab Color" ... perhaps he's just misinformed and really knows nothing about working in LAB mode.

Margulis, or course, it the author of several well-received books like Photoshop LAB Color: The Canyon Conundrum and Other Adventures in the Most Powerful Colospace. Its rather foolhardy to dismiss his advice on the basis of one rather sketchy article.

Its rather ironic to me that you claim not to care what others do yet you seem quite passionate about this issue to the point of suggesting that disclaimers ought to be added to posts that suggest LAB mode conversion in 8-bit.

There is no way you lose 87% of colours moving between RGB and LAB in 8 bit. While that test may tell you something about what happens to an image that contains every possible colour in the 8 bit RGB colour mode whilst in the sRGB colourspace, it says nothing of what happens to images in real world usage. As an audio engineer, I'm forever drawing comparisons between digital imaging and digital audio, and I would liken this to someone making the claim that they can take analogue white noise captured at 16bit/48khz, run it through spectral analysis, convert it from wav to aif, re-analyse and then show how the waveform has undergone significant change. My response to that claim would be twofold - firstly "well, I can't hear any difference" and secondly "who listens to white noise anyway?". My responses in this instance would be largely similar. Who sits around looking at images that contain 1 pixel each of 16 million different colours? Unless I see the difference with my own eyes, in a real world example, I couldn't care less if the results came back telling me that 99.9% of colours were lost due to quantisation.

I believe it was Dan Margulis who wanted to check out what kind of loss is inccurred upon swapping modes, so he converted back and forth between the two multiple times. If memory serves, he ran this process over 100 times before noticing even the smallest amount of loss.

I honestly think you've gone somewhat overboard in the scare-mongering and alarmist tone you adopted in your OP Damo, especially as a self-professed non LAB user, and therefore someone who has little to no experience working in the colour mode. As Fedka has already pointed out, there are authoritative tomes in print addressing this colour mode, and they cannot simply be dismissed by one short article describing 1 isolated experiment.

Damo77 wrote in post #5981284
Personally, I've never found any reason to convert to LAB mode, because the "Luminosity" and "Color" layer blend modes simulate it adequately for my needs.

you've seriously never tried converting to LAB to run a saturation boost using curves or levels? It's SO much superior in every way to increasing saturation in RGB that I wouldn't personally know where to start (other than strongly recommending you go out and read Margulis' book). Note - there ARE issues when desaturating in LAB, due to the nature of the mode and how the A and B channels work together. A classic example of this is when blue turns to purple on desaturation. For that kind of transformation, you're better off doing it in RGB. It's just a case of knowing what mode is best for what adjustment. But try this - create 2 versions of the same document and fill them with RGB 0 0 255 for a fully saturated blue. How can you saturate blue beyond 255 in RGB? Obviously you can't. Now take a copy of the image and do the same thing in LAB.

I for one will not be adding this disclaimer until I see compelling evidence that this article has any weight of authority to it. I don't consider the recommendation to work in LAB mode to be in any way irresponsible. While I'd agree that working in 16 bit is always preferable regardless of colour mode, the idea that LAB mode and 8 bit don't play well together just doesn't sail for me right now. As always, I'm open to changing my opinions, but this article has failed to convince me that my current standpoint is in need of revision.

OK, I thought I'd go ahead and put this to the test, because, frankly, I've never tested it before and I was curious as to whether or not there was any validity to the claims made in the OP.

I wanted to see exactly how much of an effect on an image converting from RGB to LAB back to RGB has. To this end I took the following steps:

1) Opened an 8 bit RGB image in the AdobeRGB colourspace into Photoshop
2) Took a crop from the image to make the results postable in here
3) Placed 3 colour samplers in the image over different regions of colour and luminance
4) Created a snapshot of this state and named it "Original - no conversion"
5) Converted the colour mode from RGB to LAB, performed no modifications, then converted back to RGB again. I counted this as one conversion. I repeated the process for a total of 2 conversions and created another snapshot
6) I went back to the "original" snapshot and recorded an action whereby I converted from RGB to LAB and back 10 times. I used this as the basis for creating a further series of snapshots
7) I then proceeded to create a series of snapshots that stored copies of the images that had been through the following number of conversions - 0,2,10,20,30,40,50,100​,200
8) For each snapshot, I made a note of the RGB colour values for each of the 3 colour samplers in the image, and also made notes on the subjective appearance of the image after x number of conversions
9) I created a new image from each snapshot, converted the profile to sRGB, and exported it as a jpeg with a quality setting of 79, the abridged results of this you will find posted below. As an interesting point of note, the filesize gradually increased from the original to the 200 conversion image to the point that the original was approximately half the size of the last image in the series, the reasons for which ought to become fairly apparent when viewing the images.

I drew the following conclusions from this test: the conversion of an image from the RGB to LAB to RGB colour models is a "lossy" process as was already established by a number of leading experts in the field. However, under normal, or anything approaching reasonable circumstances, this lossiness is so minute as to be invisible. Over progressive conversions of the original image, the results of this lossiness begin to manifest themselves. Subjectively this loss takes the form of a slight hue shift, and the progressive introduction of a film-grain like noise - at extreme numbers of conversions this grain develops the characteristics of fine chroma noise, and inspection of a 100% crop will reveal this quite readily. Viewing at less than 100% masks the issues significantly. However, and this is also subjective, the quality of the degradation is such that it may be likened to a degree of mild analogue distortion in recorded music - that is, the effect itself, while noticeable, does not necessarily have a displeasing appearance, especially when compared to other problems that plague images such as jpeg artefacts. Furthermore, in my own personal workflow, I do not recall a single time that I have ever converted back and forth between colour models more than 4 times at maximum. The average per image I would say is closer to 1 or 2. To achieve the results you see at the extreme settings would require either going through a deliberate process such as I have done, or being the proud owner of the most haphazzard, unproductive post-production workflow on the planet.

Anyway, the best thing that you can do is to draw you own conclusions. At the bottom of this post you will find the corresponding values for each of the 3 RGB colour samplers that were placed in the image. I chose to place them together rather than under each image, as I feel it makes for better comparison between the data.

So, without further ado, on to the pictures (title above each image).....

Original Image - zero conversions

HTTP response: 404 | MIME changed to 'image/gif' | Byte size: ZERO | PHOTOBUCKET ERROR IMAGE

HTTP response: 404 | MIME changed to 'image/gif' | Byte size: ZERO | PHOTOBUCKET ERROR IMAGE

HTTP response: 404 | MIME changed to 'image/gif' | Byte size: ZERO | PHOTOBUCKET ERROR IMAGE

HTTP response: 404 | MIME changed to 'image/gif' | Byte size: ZERO | PHOTOBUCKET ERROR IMAGE

HTTP response: 404 | MIME changed to 'image/gif' | Byte size: ZERO | PHOTOBUCKET ERROR IMAGE

HTTP response: 404 | MIME changed to 'image/gif' | Byte size: ZERO | PHOTOBUCKET ERROR IMAGE

HTTP response: 404 | MIME changed to 'image/gif' | Byte size: ZERO | PHOTOBUCKET ERROR IMAGE

Screen capture. Just posted this up for you all to show that this is indeed an 8 bit per channel image in the Adobe RGB colour space, and so that you can see where I placed the 3 colour samplers and what their starting values are. For reference, the colour samplers were set to sample at a 3x3 average.

HTTP response: 404 | MIME changed to 'image/gif' | Byte size: ZERO | PHOTOBUCKET ERROR IMAGE

Finally, just to illustrate the difference between sRGB and Adobe RGB, here are 2 copies of the image after 100 conversions. The first is the Adobe RGB image from my previous post. In the second image, I converted the profile to sRGB at the outset of the process, so that all the colour mode conversions were done with sRGB as the working colourspace. Judge the differences for yourself.


Adobe RGB - 100 conversions

HTTP response: 404 | MIME changed to 'image/gif' | Byte size: ZERO | PHOTOBUCKET ERROR IMAGE

HTTP response: 404 | MIME changed to 'image/gif' | Byte size: ZERO | PHOTOBUCKET ERROR IMAGE

Wow, you really took the bull by the horns on this one. As someone who has used lab sharpening for quite a while now I found this thread very interesting. Thanks for all the work.

Interesting read...
One thing I'm curious about: Did Bruce Lindbloom use Dither when converting?
Did you, Irreverent?

If so, how would this change things? I could imagine dithering loosing you colors, specificly when you use an image that has one pixel of every color.

Ok, looks like I went off half-cocked on this one.

But yes, I agree that Dither would make a difference. It makes sense that a colour scientist like Bruce would leave Dither off for his experiment, in order to achieve "pure" results.

And certainly the eventual damage in the 100+ conversion examples that Irreverent posted look much more like noise than banding.

Apologies to all.

I still won't be incorporating LAB conversions into my workflow, though, I'm afraid.

René Damkot wrote in post #5987408
Interesting read...
One thing I'm curious about: Did Bruce Lindbloom use Dither when converting?
Did you, Irreverent?

As fas as I was aware, dithering was a process used when moving between bit depths and to indexed colour mode. As there is no dithering option going between Lab and RGB I would assume the process isn't necessary, as we are translating color values from one color model to another as opposed to downsampling them. Whether or not dither is applied by default within Photoshop I do not know (but doubt it). To be honest, I don't know a great deal about dither in the digital imaging world. The only steps I took were to go to Image>Mode>RGB, and Image>Mode>Lab. No other options or anything else was carried out for the duration of the test.

There is ditering (or at least the option) when you go through convert to profile.

Ahhh, I wasn't converting to profile René, I was changing mode. I don't think anyone who swithces to LAB in their workflow does it by converting to profile. I can't think of a particularly good reason to convert the profile to LAB for standard PP work.

I can maybe set up an action that runs the profile conversion a few times later on, and post the concise results up here for you to see for yourself if you like.

Would be interesting