in praise of RawTherapee, "Amaze" demosaicing algorithm

Started Jun 10, 2011 | Discussions
OP RussellInCincinnati Veteran Member • Posts: 3,201
Raw Therapee great. Amaze no better at "normal" print sizes

GaryW and Ck3 have raised the question of the significance of Amaze demosaicing at normal print sizes (rather than huge, 100% magnification pixel-peeping of 3 foot high images on our monitors). Thus have prepared below (what's amazing is how long these comparisons take to prepare) 5 images, taken on a tripod, with remote release, from a pretty perfect multicoated enlarging lens, being used at its designed-for optimum distance, at F9.5:

  1. Typical AHD demosaicing in RT Raw Therapee, no unsharp mask sharpening as is quite typical, at huge magnification. Nex does nicely.

  2. Amaze demosaicing in RT, RL Deconvolution mild sharpening but also no unsharp mask, at huge 1 meter high print size. Noticeably crisper than AHD demosaicing, and at the same time a touch less noise. Seems like a win for Amaze+RLD, and the Nex looks to be larger sensor DSLRs a run for their money.

  3. Picture number 1, but the AHD-demosaiced image has now been put through typical radius 1 unsharp mask sharpening. Now the AHD image looks as sharp as the Amaze image number 2. But, the noise is of course still tiny a bit worse than in the Amaze image. Not to say the noise is real bad in any of the images.

  4. Typical AHD demosaicing plus unsharp mask, at a reasonable print size. Raw conversion took just a few seconds.

  5. Amaze demosaicing plus RL Deconvolution plus unsharp mask. Raw converstion took at least 1 minute. Appears identical in every way to AHD demosaicing, at this normal print size. Darn, didn't think of taking a JPEG at the same time, but this was a real assignment not a test session.

My original post praised Raw Therapee plus the Amaze demosaicing algorithm. My opinion after this test is that Raw Therapee, with all its defaults, does an almost perfect job of "developing" raw files. For example the above AHD images are all straight out of Raw Therapee with everything defaulted, and the color and brightness are almost perfectly natural (well the darkest tones need a bit of pulling down). Of course someone might want to add more color saturation or contrast to their images, but it's nice to start with natural accuracy, and it's much easier to contrast and saturation than it is to take them away when you've overshot.

But can now see that Amaze demosaicing and RL Deconvolution are only advantageous at huge magnifications. At normal print sizes from not-very-much-cropped Nex raw files, typical AHD demosaicing results are indistinguishable from the output of the new Amaze demosaicer.

Unless there is a moire false color problem, which (shown in this thread) Amaze does a bit better with at any print size. But then most images do not have moire problems.

ck3 Senior Member • Posts: 2,925
Thanks for the test (nt)

no text

OP RussellInCincinnati Veteran Member • Posts: 3,201

GaryW:Why bother with deconvolution or any sharpening then, if you're not going to zoom in on that level of detail?

Good question, leading to some useful experiments, and agreed. For example have just posted that there is no advantage to the Amaze demosaicing plus RL deconvolution at normal print (i.e. display) sizes, unless there's a moire problem (which is not frequently the case).

As for the huge print sizes, my original post was showing that there is more sharpness that can be extracted by Amaze demosaicing plus RL Deconvolution from Nex raw files, than you can get with AHD demosaicing. And my unsharp mask samples showed that you can't make up for that sharpness in an AHD image simply by applying an unsharp mask.

You keep hitting on the point that you don't like my unsharp mask application to the magnified sample images, because it "oversharpened" them. I must not be making it clear that the unsharp masked examples were only shown on this forum as part of a demonstration that the lost sharpness of the AHD demosaicing could not be simply recovered. I was not meaning to recommend that everyone should apply the same unsharp mask to real images that I applied to my samples. Thus my comment to you about the "oversharpened" sample images is not really to argue with you, it's just to point out that it's not real important that the sample fragments are oversharpened.

GaryW: I'm just saying, if you're going for the best pixel-level resolution, don't overcook the sharpening and mess it up.

OK, agreed, when am not posting exaggerated-fine-detail examples on the Sony Nex forum, it would be better for me not to oversharpen my client's images.

Russell: Anyway, the overall point of interest in this thread is not to show a workflow that is "good" in any absolute sense.

GaryW: That's fine, but I take opportunities to try to get people to think about sharpness, as most people overdo it, IMHO.
OK, I see.

GaryW: But...[in arguing for not showing us oversharpened test images] I like to make comparisons of real-world, this-is-how-I-would-do-it over perfectly-and-arbitrarily-"fair".

Agreed that it would have been better for me to show exaggerated-for-fairness-and-didactic-purposes examples, and also should have shown real world workflow examples. Have indeed done a bit of that in my post just a couple of minutes ago, showing my real workflow on a real job resulting in normal sized images. And interestingly in that realistic, normal-print-size workflow there is no advantage (or even perceptible difference) to Amaze demosaicing and RL Deconvolution.

GaryW: But as long as you are clear that you're just forcing an equality of procedure to prove a point, that's OK
Well I must not have made that perfectly clear.

GaryW, but since deconvolution is just another kind of sharpening, I'm not sure you really have to run it through the separate sharpening technique. As I said, if you really think it needs touching up, I'd try a small radius. Often I find deconvolution to be sufficient.

My results with Amaze+RL deconvolution suggest that while it does induce some jaggies (as CK3 I think noted), it doesn't produce as much haloing around edges of things as does unsharp mask. RL Deconvolution is a much different sharpening algorithm than unsharp mask, quite of the digital era and related to the technology of the new "lightfield" camera. Unsharp mask is an imitation of procedures well known in the film era, and while I love it dearly, it does mangle the images and increase noise a bit more for a given level of crispness than RLD. Once again now am clear that AHD+unsharp mask disadvantages only noticeable at huge print sizes.

Russell: Adopting RT Raw Therapee...solved maybe half of Nex moire problems. Without blurring anything.

GaryW: I agree -- using RAW and taking that extra time is well worth it, at least for shots that matter. I also like DxO for its noise handling. I find it superior to the free RAW converters. It has its own deconvolution, but only for certain lens/camera combinations, unfortunately.

However that DxO deconvolution should be better than Raw Therapee's, in the case where you're using a lens and sensor combo that DxO knows about, and there's no motion blur or other complexifying factors that make the DxO super precise deconvolution less noticeable.

Russell: RL Deconvolution sharpening ... does not sharpen so much as to screw up later resizing of an image

GaryW: It can extract a bit of detail where you thought there wouldn't be any more to get. You can definitely overcook deconvolution, though. It's not a cure-all.

Fortunately have found Raw Therapee's deconvolution with its built in defaults to be not overcooking anything. In fact have no desire to ever touch those defaults, since the default sharpening is so great and the slight jaggies produced from the default RT workflow seem inconsequential.

GaryW: On the other hand, normal sharpening is easily overcooked and messes up your low-level detail.

Agreed, and agree that several of my posted samples in this thread show overcooking with the unsharp mask.

IdiotBird Junior Member • Posts: 25
Re: RawTherapee looks promising, but unfortunately no good results with ARW on Linux

Clicked wrote:

I've tried to get RawTherapee working with the Nex .ARW files, but unfortunately there is no way to get a normal look. The difference with JPEGs is huge, colours are completely off, and RawTherapee's forum doesn't offer any solution.

Good to see that on Windows things are fine, but seeing as it's free software, I thought it might be good to mention that there are still major Linux issues, in any case on Ubuntu-like distro's.

I use Ubuntu, and I too felt that RawTherapee produces very poor results. I then found out that we have very old versions of the software in the repository (3.0 alpha1). Most Windows users would be downloading RawTherapee now, and therefore, using the latest software. It's sad that the person responsible for RawTherapee on Launchpad has not made the effort to remove his builds.

Anyway, the latest build of RawTherapee seems to need a library that's not in 11.04's repositories. I had to execute the following commands to get the library:
sudo apt-add-repository ppa:dasprid/rawtherapee
sudo apt-get update
sudo apt-get install liblcms2-dev

(This information can be found via their homepage.)

I did all this, and on starting the program via rtstart, it tells me that it is "Loading thumbnails..."

I'll wait for a while and see how things go. Maybe you can do this too in the mean time, if you're curious.

 IdiotBird's gear list:IdiotBird's gear list
Sony Alpha NEX-3 Sony E 18-55mm F3.5-5.6 OSS Lenovo Moto G Plus
IdiotBird Junior Member • Posts: 25
Re: RawTherapee looks promising, but unfortunately no good results with ARW on Linux

I did all this, and on starting the program via rtstart, it tells me that it is "Loading thumbnails..."

Apparently, the "Parsed Extensions" was blank by default in this version. Add a few extensions manually. Just add the letters of the extensions, and not as "*.jpg" or ".jpg".

 IdiotBird's gear list:IdiotBird's gear list
Sony Alpha NEX-3 Sony E 18-55mm F3.5-5.6 OSS Lenovo Moto G Plus
Keyboard shortcuts:
FForum MMy threads