JUST OUT OF BETA TESTING: Topaz Gigapixel AI 6.1 (need to see!)

[fdisilvestro]

All one has to do is examine a DNG from Gigapixel in RawDigger. It's a Linear DNG (RGB not RGBG) data.
http://www.barrypearson.co.uk/articles/dng/linear.htm

Well, that does not seem to me as enough evidence. Linear DNGs out of DxO Pure Raw are RGB but not color space encoded and you can still apply a DCP Profile in LR / ACR. RawDigger still shows "Photometric Interpretation: Linear Raw" in the Gigapixel AI's DNG.

The first impression I had from the DNGs out of Gigapixel AI was that they were already color encoded, because you could not choose a DCP Profile in LR / ACR.

I tested an old image taken with a Nikon D300, processed from NEF into Gigapixel AI then analyzed the EXIF from the resulting DNG.

I think I have found a bug in Gigapixel AI. The field "Unique Camera Model" had "NIKON CORPORATION" (same as "Make") instead of "NIKON D300" (as in "Camera Model Name") as shown in the first image attached (I compared Linear DNGs out of DxO Pure raw to see what info they had in this field).

I used a HEX editor to change the information in the file, so that now the EXIF shows "Unique Camera Model" as "NIKON D300" as shown in the second image attached, and then when I opened the image in ACR, I was able to choose a DCP profile for the image. The third image shows a snip of ACR where I can select the DCP Profile.

So, in order to solve this bug, all Topaz Labs has to do is to change the way it handles this specific EXIF field.

[digitaldog]

Well, that does not seem to me as enough evidence.

Evdience of what? The data out of the product is not RGBG, it's processed.

The first impression I had from the DNGs out of Gigapixel AI was that they were already color encoded, because you could not choose a DCP Profile in LR / ACR.

This is simply a CR2 'opened' then processed in Gigapixel, it isn't the same raw data, it isn't saved as they tell us, as CR2. That part of the dialog used is a lie.

So, in order to solve this bug, all Topaz Labs has to do is to change the way it handles this specific EXIF field.

What bug? First, they tell us what they are doing which isn't what they are doing. They are not providing back the original data and when asked to provide a non DNG/"Raw" in the original color space which isn't possible, they give me ProPhoto RGB. What they are doing may not be a bug at all, simply a big fat lie in the save dialog.
The rendering from raw to DNG is awful; that may not be a bug, just really piss poor rendering.
Bottom line; they tell us they are going to provide something based on the settings they don't and can't. And what they provide in terms of color and tone processing of the raw, when viewed is butt ugly. Bug? Or just sloppyness.

[fdisilvestro]

I did some more tests and the results do not look god, so I guess you are right, it does not do a good job with raw files. I'm thinking that it is actually on purpose that they changed the EXIF value so you cannot choose a DCP profile.

It works ok for rendered files.

[Chris Kern]

. . . if they've got something proprietary going on that others cannot interface with, that might be part of the issue here.

As Andrew has already pointed out, everyone's demosaicing algorithms are proprietary.  Adobe's, Topaz's, DxO's, Iridient's—you name it.

I've done some more poking around today and, as far as I can tell (the usual disclaimers apply about my lack of demonstrable technical expertise in this area), when Topaz Gigapixel is used to enlarge the pixel dimensions of a raw camera capture into a DNG file, what you get is a fully-rendered image, comparable to a TIFF.  In other words, the DNG file format is just the container for image data that might otherwise be encapsulated into a TIFF or even a JPEG.

All perfectly legitimate, as far as I know, because you can stuff any image data into a DNG container.  But that means color information and, I believe, at least some tonal information have been "baked into" the output.

By contrast, Adobe's "enhance" (i.e., "super resolution" enlargement) emits a linear DNG with EXIF data that inform the post-processing program how subsequently to treat the file as though it was still a raw camera capture.

[Nikon NEF source]

$ exiftool DSC_5288-Enhanced.dng | egrep CFA
CFA Repeat Pattern Dim          : 2 2
CFA Pattern 2                   : 0 1 1 2
CFA Plane Color                 : Red,Green,Blue
CFA Layout                      : Rectangular
CFA Pattern                     : [Red,Green][Green,Blue]

[Fuji X-Trans source]

$ exiftool _XT43050-Enhanced.dng | egrep CFA
CFA Repeat Pattern Dim          : 6 6
CFA Pattern 2                   : 1 1 2 1 1 0 2 0 1 0 2 1 1 1 2 1 1 0 1 1 0 1 1 2 0 2 1 2 0 1 1 1 0 1 1 2
CFA Plane Color                 : Red,Green,Blue
CFA Layout                      : Rectangular
CFA Pattern                     : [Green,Green,Blue,Green,Green,Red][Blue,Red,Green,Red,Blue,Green][Green,Green,Blue,Green,Green,Red][Green,Green,Red,Green,Green,Blue][Red,Blue,Green,Blue,Red,Green][Green,Green,Red,Green,Green,Blue]

(The same metadata can be extracted with a commercial product, RawDigger.)

This difference between the semantics of the Topaz and Adobe output files doesn't have any inherent implications with respect to the quality of the neural network processing, which is presumably dependent on the particular algorithms used by the respective software manufacturers and—to an even greater extent, I suspect—the particulars of their machine-learning training sets.

But it does mean that after the demosaicing has been performed, you can treat the Adobe files as though they were still raw, which can be a significant benefit where color accuracy, or just getting the final image to look the way you want, is concerned.

[plugsnpixels]

Thank you once again Chris for your efforts here!

Between you and Andrew I am understanding that Gigapixel renders its results and wraps them in the DNG envelope (if one chooses that option) without the original RAW data that the user may want or need for further work in a RAW-friendly editor. Depending on the desired end use, I am going to assume this may or may not be an issue, and certainly isn't for those salvaging JPEG or TIFF sources, yes?

I'm also curious how other developers (ON1, Exposure, Skylum, Serif, etc.) are handling this workflow.

While we await the Gigapixel engineer's comments on the matter, I return to my original premise:

IMO and in my own testing, aside from the color profile issue, Gigapixel is giving superior upscaling results when compared to Adobe's Super Resolution and even ON1's new Resize feature.

I just did some tests using a CR2 image I shot YESTERDAY (usually I am scraping around in my almost 50 years' worth of rough image files for examples ;-) ). Here's what I got (images attached):

Adobe "Super Resolution". Very little control is offered and the results "don't seem" to be very good in either the preview or the exported DNG. Original details are more "natural" but still mushy. I don't understand how this different from their older methods.

ON1's new Resize at 6X (much better than Adobe in terms of clarity but some finer details are lost).

Topaz Gigapixel at 6X. Offers 5 different upscaling models (4 shown) along with full control of magnification factor, noise removal, etc.

Gigapixel result exported as DNG and reopened in Camera RAW: The colors don't seem garish (?). I left all values at default.

Here is my original CR2 file for those interested in testing my results.

So again, as I see it, we are dealing with two topics here: Color management and output quality. Adobe gets one right and not the other, Topaz the opposite.

[plugsnpixels]

For some reason Gigapixel's funky DNG output looks better in Photoshop than its JPEG output, regardless of JPEG color setting chosen.

[digitaldog]

Between you and Andrew I am understanding that Gigapixel renders its results and wraps them in the DNG envelope (if one chooses that option) without the original RAW data that the user may want or need for further work in a RAW-friendly editor. Depending on the desired end use, I am going to assume this may or may not be an issue, and certainly isn't for those salvaging JPEG or TIFF sources, yes?

My main beef, outlined in two replies to BP is that the save dialog doesn't do what it says it's going to do. Part of what it says it's going to do in a raw workflow is not possible.
Where is the written manual for this product? I couldn't find one but I can't say I looked exhaustively for it (nor should I have to)?

What is the recommended workflow for say a Lightroom Classic or Camera Raw user who's working with raw (DNG) data to and from this product? That might help explain what the product is doing and why but nothing it states would change a Save dialog that isn't at all accurate in what it states it will do.

Adobe "Super Resolution". Very little control is offered and the results "don't seem" to be very good in either the preview or the exported DNG. Original details are more "natural" but still mushy. I don't understand how this different from their older methods.

This is but ONE process in a long process of rendering raw data. A process that thus far, outside sizing up an image, Gigapixel gives absolutely NONE.

Gigapixel result exported as DNG and reopened in Camera RAW: The colors don't seem garish (?).

Is that a statement or a question?
In a rendered workflow, the raw to rendered processing is conducted, you can't put that toothpaste back in the tube. That is OK, you've rendered the raw as you desire. This is a process Gigapixel doesn't provide any control over for the user. On the other hand,  TIFF you send Gigapixel looks like the TIFF you get in terms of color and tone, color space, etc. This thus far absolutely isn't the case when I or BF (and presumably you) send a raw to the product and that's an issue.

IF the product either can't honor the color appearance from a raw converter (which is nearly an impossible task that even Adobe can't do from say a camera-generated JPEG which of course comes from raw), then not only shouldn't the product lie and confuse users as to what it's going to with raw data, it simply shouldn't accept raw data. Of course, it does. So the 'issue' is twofold:
1. A Save dialog that is a lie. I've outlined two lies earlier.
2. At least an initial workflow where it does a very poor job with no control over the rendering the raw it allows you to feed it. Above and beyond it's AI upsizing!

So what do the designers expect raw users to do before the product comes into use?

[Bart_van_der_Wolf]

[...]
So what do the designers expect raw users to do before the product comes into use?

Gigapixel is not a dedicated Raw converter, but it offers the possibility of opening Raw files (I think it uses the LibRaw engine/library). So for the best (and predictable) results, offer it a TIFF for input. The TIFF can come from one's preferred Raw converter (best is unsharpened, because sharpening artifacts will be upsized and become more visible).

Cheers,
Bart

[digitaldog]

Gigapixel is not a dedicated Raw converter, but it offers the possibility of opening Raw files (I think it uses the LibRaw engine/library).

That makes it a raw converter. Not a very good one IMHO; yours?
There is no reason it has to convert raw data. Or so poorly. Lots of products only process rendered images!
Nothing above also justify the lie told in the Save dialog I've outlined. Why Bart does it do this?

[Bart_van_der_Wolf]

[...]
Why Bart does it do this?

I don't know.

[digitaldog]

I don't know.

Worth repeating here:

"JUST OUT OF BETA TESTING: Topaz Gigapixel AI 6.1 (need to see!)" is more like JUST OUT OF BETA TESTING: Topaz Gigapixel AI 6.1 (needs a lot more beta testing!)

[Chris Kern]

There is no reason it has to convert raw data. Or so poorly. Lots of products only process rendered images!

And of course that is exactly what happens when you invoke Topaz Gigapixel from Lightroom's Edit In dialog.  Or, for that matter, when you invoke Photoshop or any other pixel editor or converter.  Lightroom transfers a rendered file for Gigapixel to process.  This allows you to edit the image the way you wish without modifying the original raw file ("parametric editing"), then use Gigapixel to enlarge it.  That workflow makes more sense to me than using the Topaz product as a raw converter.  Once the rendered image completes the round-trip, you can then make any further adjustments that may be appropriate since processing a photograph through a neural network may modify some of its attributes in ways that are impossible to anticipate in advance.

[digitaldog]

And of course that is exactly what happens when you invoke Topaz Gigapixel from Lightroom's Edit In dialog.  Or, for that matter, when you invoke Photoshop or any other pixel editor or converter.  Lightroom transfers a rendered file for Gigapixel to process.  This allows you to edit the image the way you wish without modifying the original raw file ("parametric editing"), then use Gigapixel to enlarge it.  That workflow makes more sense to me than using the Topaz product to as a raw converter. 

Yes it does. Yet the product allows us to open a raw and then it provides a Save dialog that's a lie, which is my beef and answer to BP:

There is a profile issue here I can't figure out yet. I work in Adobe RGB, which the Super Rez image from Lightroom is. Gigapixel seems to give me it's own profile.

This is a mess by Gigapixel's own (unnecessary) design.

[plugsnpixels]

Andrew, as for the quality of beta testers that you've mentioned twice, perhaps you can join in! It wouldn't take any extra time – you could say the same things there you are saying here (copy and paste even), the main difference being the devs who can do something about it will see your valuable comments directly.

I'm pretty clear on the RAW discussion as concerns Topaz's handling of the files and agree some different approach and perhaps clarification needs to be done. But what about everyone's opinion of the quality of the upsizing as compared to the other apps? That's a separate discussion we can have while we're here.

[Daverich]

It’s not clear to me why people are resizing the Raw files. I open my Raw files in ACR and make as many adjustments as I can there and open them as Smart Objects into Photoshop. I have no idea what resizing I’ll need until I’ve cropped the image and decided what output size i want. Gigapixel is the second to the last item in my workflow. Output sharpening is the last step and with the way I use it at least, Gigapixel works really well.

[plugsnpixels]

Thanks Daverich, that's a good point.

So I don't know whether people would feel slighted if they were not able to open RAW files in Topaz apps at all! Probably...

[deanwork]

Trickle down from Hollywood.

https://youtu.be/ixgFtjfO_7Q

I was using Topaz Gigapixel for quite a while before Adobe introduced its own machine-learning technique for increasing resolution.  The Topaz results never struck me as entirely satisfactory, but after trying several online alternatives it seemed like the best compromise.

Not so much anymore.

I have only a limited understanding of the design of neural networks, and I'm certainly no expert on color management, but my impression is that when fed a raw file, Gigapixel emits a completely rendered image (i.e., with respect to color, output-format-referred) even when you specify a DNG container as the destination file format, while Lightroom and Photoshop produce a DNG file that is still scene-referred so you can still make color adjustments without producing artifacts.

I still use both products, but as they continue to mature—I presume both software manufacturers are continually updating the "learning sets" they use to train their neural networks—I'm experiencing less need for Gigapixel except when I need to make massive increases in resolution.

The underlying machine-learning techniques are available to everyone, so the ability to manipulate raw RGB data and produce a manipulable output file is an important differentiator.  It seems reasonable that Adobe would have an edge here.  I suspect that how exactly Adobe manages to preserve the user's ability to manage the colors in the output file is a trade secret.  Somebody like Eric Chan could no doubt explain it, but I'm not holding my breath.

[Benny Profane]

It’s not clear to me why people are resizing the Raw files. I open my Raw files in ACR and make as many adjustments as I can there and open them as Smart Objects into Photoshop. I have no idea what resizing I’ll need until I’ve cropped the image and decided what output size i want. Gigapixel is the second to the last item in my workflow. Output sharpening is the last step and with the way I use it at least, Gigapixel works really well.

I didn't consider this, and it seems to work. My workflow out of Lightroom is the opposite: Super Rez first, all edits afterwards. Bypasses this stupid profile problem. Gawd, the processing is slow, though. Like, go make a little lunch slow.

[digitaldog]

Depending on if you know or not from the get go, you need to upsize the original raw or not, it makes a lot of sense to do this first and apply as many parametric edits on top than after. Just the new masking features alone will suffer or benefit from doing this step first.
If you don't know you'll resize up, then Daverich's approach makes sense since you're never getting a free lunch here in terms of quality data (SuperRez, AI, Gigapixel or otherwise). You may find you have to at a later date so yeah, you do it then and hope for the best.
If you know from the start, you've got to shoot an 8x10 transparency, you do that. If you can.

[plugsnpixels]

I didn't consider this, and it seems to work. My workflow out of Lightroom is the opposite: Super Rez first, all edits afterwards. Bypasses this stupid profile problem. Gawd, the processing is slow, though. Like, go make a little lunch slow.

Benny, what type of computer are you using? On the Mac side, moving to M1/2 makes a huge difference over Intel, whether you're using Handbrake to crunch videos, Gigapixel to enlarge images or anything else.

An even more serious computer than usual is getting to be a necessity these days... Certain art-type AI apps require high-end graphics cards, for example (I'm thinking of NVIDIA Canvas for one).

I did a Maya test render test awhile back, comparing a 2020 M1 MacBook Pro, 8 gigs RAM with a 2013 Mac Pro (trashcan) with 32 gigs RAM, 2 video cards, etc. The total render time was literally only seconds apart!