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

[Chris Kern]

Computational enlargement techniques like Adobe Enhance/Super Resolution and Topaz Gigapixel don't interpolate extra pixels the way traditional digital enlargement techniques do.  They create a new high-resolution image based on visual attributes the neural network identifies in the low-resolution image.  Depending on the sophistication of the neural network, and the extent and variety of the set of images that was used to "train" it, you may get a near-perfect enlargement for a particular source image or one that contains noticeable artifacts.  So whether you perform the enlargement before editing it or afterwards, a close (even 1:1) inspection of the resulting "rebuilt" picture is advisable, in my experience, because post-enlargement modifications sometimes significantly improve the final picture.

[Benny Profane]

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!

Well, that's nice, and, yeah, lots of great M1 reviews, but, I just dropped some major coin on a used trashcan recently, and I'm not moving up until I have to, i.e., I can't load the latest Mac OS and, therefore, keep up with Photoshop updates. Have you seen the state of the stock market lately? I'd rather save for that new 40 mp Fuji H2 rumored for, whenever. Then there's traveling to cool places to shoot...

[digitaldog]

Computational enlargement techniques like Adobe Enhance/Super Resolution and Topaz Gigapixel don't interpolate extra pixels the way traditional digital enlargement techniques do.  They create a new high-resolution image based on visual attributes the neural network identifies in the low-resolution image.  Depending on the sophistication of the neural network, and the extent and variety of the set of images that was used to "train" it, you may get a near-perfect enlargement for a particular source image or one that contains noticeable artifacts.  So whether you perform the enlargement before editing it or afterwards, a close (even 1:1) inspection of the resulting "rebuilt" picture is advisable, in my experience, because post-enlargement modifications sometimes significantly improve the final picture.

Be useful to see such near-perfect enlargement. Take say a 24MP/48MP capture, sample it down to 4MP or even 12MP, use such a product to get it back to the same size and show the results.

[Chris Kern]

I just dropped some major coin on a used trashcan recently. . . .  i.e., I can't load the latest Mack OS and, therefore, keep up with Photoshop updates.

Eh?  MacOS 12.X installs and runs fine on a 2013 Mac Pro.

[Benny Profane]

Eh?  MacOS 12.X installs and runs fine on a 2013 Mac Pro.

Not the one collecting dust in the corner of my room. Sob.

[Chris Kern]

Be useful to see such near-perfect enlargement. Take say a 24MP/48MP capture, sample it down to 4MP or even 12MP, use such a product to get it back to the same size and show the results.

Interesting experiment.

You definitely would not get a byte-by-byte correspondence between the original source and downsampled-upsampled destination files, so I presume the idea would be to make a 1:1 visual inspection of the two files and decide whether there were any objectionable visible differences.  Does that make sense?

[digitaldog]

Interesting experiment.

You definitely would not get a byte-by-byte correspondence between the original source and downsampled-upsampled destination files, so I presume the idea would be to make a 1:1 visual inspection of the two files and decide whether there were any objectionable visible differences.  Does that make sense?

Does not have to be byte for byte; just show no visible difference. Or very little compared to another upsizing method.

[plugsnpixels]

Well, that's nice, and, yeah, lots of great M1 reviews, but, I just dropped some major coin on a used trashcan recently, and I'm not moving up until I have to, i.e., I can't load the latest Mac OS and, therefore, keep up with Photoshop updates. Have you seen the state of the stock market lately? I'd rather save for that new 40 mp Fuji H2 rumored for, whenever. Then there's traveling to cool places to shoot...

Curious why you did that-? I support a bunch of those in Visual Arts (higher ed) and while they've held up 24/7 for 9 years (only a couple have failed with graphics glitches), even Apple has lost interest in them.

And yes, I do believe they can handle Monterey, but I'm about to post a separate thread about that...

[plugsnpixels]

because post-enlargement modifications sometimes significantly improve the final picture.

This may not be exactly what you are referring to but what I like most about doing "artsy" post treatments is: It doesn't matter what the quality of the original was, art effects cover all sins...

I originally took this scenic photo in 1985 in Switzerland in B&W on 35mm! Colorized it online and artsified it with Dynamic Auto Painter.

And relevant to our current discussion, I upscaled it with their companion app, AI Photo & Art Enhancer (I haven't written about this one yet, just got ahold of it). This app seems to be the next best thing to Gigapixel, sometimes better, sometimes not (my tests using the building photo were from a 1985 35mm slide scan).

[Chris Kern]

Be useful to see such near-perfect enlargement. Take say a 24MP/48MP capture, sample it down to 4MP or even 12MP, use such a product to get it back to the same size and show the results.

I thought this was an ingenious experiment, so I decided to accept the "Rodney Challenge."

I used a rather unsuccessful street photograph I made a few years ago in Kyoto as the test subject because it represents a situation where I would feel a need to increase the dimensions of an image file; it also has a variety of visual elements, including people and objects of various kinds, that I thought would be a good test of the ability of the computational enlargement methods to recreate different shapes.  I was on the other side of a busy downtown street with a lens that didn't have enough reach for my distant subject: a video photographer who had been shooting a group of young women wearing kimonos just outside the door of a busy storefront, presumably as an advertisement for the store.  I didn't have time to cross the street and move closer (as it was, I was a few seconds too late to get the picture I wanted), so I had to crop very aggressively in post.  I wouldn't have had enough pixels for anything larger than, say, an 8x10-inch (A4) print.

One of the attached files is the full-size image, exported directly from Lightroom.  I then reduced the linear dimensions of the image by half as a source for the other three attachments—in other words, the reduced file had one-fourth the pixels of the original.  One of the three remaining attachments was enlarged to the original image size through interpolation by Photoshop; I selected Photoshop's default Automatic enlargement method.  Yet another of the attached files was enlarged using Lightroom's computational Enhance method.  The final attachment was enlarged using Topaz Gigapixel's computational method.  I made no further adjustments to the any of the images following enlargement, so what you see here are JPEG exports of exactly what the respective software products emitted.

I can spot some differences among the various versions, but if I didn't know which was which I think I would find it difficult to identify the original, much less which of the enlargements was done through pixel interpolation and which through computational methods, whether I view the entire images or zoom in and examine them at 1:1.

I've given the files generic names in order to give those reading this thread an opportunity to take the Challenge.  I'll report their identities tomorrow.

[digitaldog]

I thought this was an ingenious experiment, so I decided to accept the "Rodney Challenge."

Can you gang them up, into one TIFF (san's JPEG) and maybe place on something like Dropbox so we can download and view them in one way inside say Photoshop? TIA.

[plugsnpixels]

Thanks Chris, great test image!

I downloaded each and used Preview (with all in one group) to scroll among them.

I am going to guess:

A/D = PS and LR or LR and PS
B = Original
C = Gigapixel

The main takeaway here is, the better your original, the less help it needs.

In my own case, I have a lot of old crap that is of no good use unless it gets a major overhaul. This example is an early-'90s 35mm negative film scan, treated in a slightly older version of GP.

[Chris Kern]

Can you gang them up, into one TIFF (san's JPEG) and maybe place on something like Dropbox so we can download and view them in one way inside say Photoshop?

Are you concerned about JPEG artifacts and want to see TIFFs instead?  If not, you can just click on the thumbnails to launch the website's image viewer and save the full-resolution files to whatever device you are using to view the forum thread.  I don't think there will be any perceptible visual difference between the full-resolution JPEGs and the corresponding TIFFs.

[digitaldog]

Are you concerned about JPEG artifacts and want to see TIFFs instead?

Mainly I'd prefer to see then all ganged up.
Secondly sure, I'd prefer to see a TIFF without unique JPEG compressions from each original.

[Chris Kern]

Mainly I'd prefer to see then all ganged up.
Secondly sure, I'd prefer to see a TIFF without unique JPEG compressions from each original.

You can download a composite containing all four images here.  From top to bottom, they are fileA, fileB, fileC, fileD.

[digitaldog]

You can download a composite containing all four images here.  From top to bottom, they are fileA, fileB, fileC, fileD.

Cool, makes viewing at 100% and syncing multiple views in Photoshop easier.
FileD looks the worst to me; needs some sharpening for sure compared to the others.
The others are more similar with differences seen in 'grain' (noise) with FileC looking like there is less of it than others, making me wonder if it isn't Gigapixel. It looks fine and in some respects 'better' with less grain/noise but somewhat processed appearing.
File A and B look the most 'natural' to me with B looking a tad sharper on this grain. My guess is it may be the original and A interpolated enough to soften it just a little. Now I wonder if a very little bit of capture sharpening after sizing (if so) might bring it back closer to B.
C and D look too 'smooth' without the lack of this grain, they appear processed at least to me. C being the worst offender. B is my preferred image.
None are 'bad' and I suspect they would all print just fine and each may need its own degree of post capture sharpening and output sharpening for an ideal output.

[Chris Kern]

One of the attached files is the full-size image, exported directly from Lightroom.  I then reduced the linear dimensions of the image by half as a source for the other three attachments—in other words, the reduced file had one-fourth the pixels of the original.  One of the three remaining attachments was enlarged to the original image size through interpolation by Photoshop; I selected Photoshop's default Automatic enlargement method.  Yet another of the attached files was enlarged using Lightroom's computational Enhance method.  The final attachment was enlarged using Topaz Gigapixel's computational method.

The file identifications are

FileA is the Lightroom Enhance enlargement;
FileB is the original (i.e., full-size) image;
FileC is the enlargement by Topaz Gigapixel;
FileD is a traditional Photoshop enlargement.

[Benny Profane]

You know, another little annoyance that loading this software into my computer has given me is that, anytime I directly double click on a raw file somewhere deep in the library, now, the default program is Gigapixel that opens the file, not camera raw, as before. Cmon, people, don't do that. Don't screw around with my little workflow. I'm busy. Now I have two choices, just to dump Gigapixel in the trash, or play computer detective and fix it so that Gigapixel doesn't intrude on my life like that. which is not making me happy. I hate being IT guy. The same thing happened with a third party software that I loaded a while ago, Boris FX Optics, which is a fun filter for effects, but, that's all it is, a side show, so, I had to tamp it down until it learned it's place I'm my world.
These companies think they are going to break the back of Adobe, and, ain't happening. Like it or not, Adobe rules. Monopoly power. Just stop with the games and learn your place in the world.

[plugsnpixels]

Benny, I can't answer whether Gigapixel caused this or not, but on your Mac, make a 2-second adjustment ("Get Info" on a Raw file and change the default app back to whatever you want, as shown here with a different example).

If after doing so it happens again, then we have something to report.

[Benny Profane]

That's kind of ridiculous. I'm not going to add a "two second" step to a very simple procedure just to accommodate this intrusion into my system. Bet Bye Gigapixel, because nowhere in your preferences do you allow me to disable this annoyance.