NEW Topaz upsampling product: A.I. Gigapixel

[plugsnpixels]

Release date: August 16.

A.I. Gigapixel is a new image resizing app contender, for both JPEG and RAW images. See my blog post with examples. An unadvertised discount is available.

[Manfred M]

It looks like the rather ancient OnOne Perfect Resize has a competitor with this plug in.

How does this technology compare to the Photoshop's up-sampling algorithm? It's not clear that you have tested this aspect of it in your article. To me that would be the question to ask.

My other comment is that this product is likely only of interest to people who make large prints, as this level of resizing should never be an issue for people who post to the internet, given the relatively low resolution (when compared to print) of even 4k screens.

[plugsnpixels]

Thanks Manfred, I had begun doing the PS comparisons and will continue later today.

True about prints being the main output method of such technology, though in extreme cases a really teeny original might be needed for bigger web use.

[plugsnpixels]

I updated my blog post with some A.I. Gigapixel vs. PS results.

[DanK]

Very interesting. This and Richard's thread leave me with two questions:

1. What is the software actually doing? Phrases like "deep neural networks" don't tell us what edits the software is making, and it seems clear from the examples I have seen so far that it isn't simply uprezzing. It seems from the examples that this includes at the least some noise reduction and sharpening.

2. Mike, you offer a discount. What's the connection? Do they sponsor you in some way?

[plugsnpixels]

Hi Dan, this page gives some background info of how A.I. Gigapixel works.

This page explains how my website works. Topaz first reached out to me over a decade ago and I have been using their plug-ins/apps ever since.

[DanK]

Mike,

Thanks. That link has some information, but not explicitly what I was looking for:

A neural network is exposed to a large number of high-resolution and low-resolution image pairs. This network gradually learns to synthesize plausible detail in the enlarged image based what it has seen.

The key thing, it seems to me, is what criteria it uses to determine "plausibility" and what edits that selection process implicitly entails. There is a complaint in the comments about unnaturally smooth skin textures, the response to which was to retry without the "enhance image" option checked. That gives one part of the answer: at least with that option checked, smoothing is part of the process. That makes sense to me; one of the most obvious differences between the AI and PS images you posted is smoothness.

Along the same lines, one of the developers' responses to a comment was this:

Turn off the LR noise reduction since A.I. gigapixel usually do much better jobs.

The difference in noise was the first think I noticed in the images you posted. That suggests that a more informative comparison would be Gigapixel to the combination of uprezzing and noise reduction in PS.

The link includes technical information that potential users would find useful.

Re your final link: at the end of that page, you note that some developers, including Topaz, give you a referral fee. It's good that you include that information in your blog, but it would be best, IMHO, to note it here when you recommend a product from one of those developers. That's not to imply that anything you write would be different if there were no referral fee, but it seems best to me to be transparent.

Dan

[plugsnpixels]

Thanks Dan, once my workweek calms down I'll do some site revisions.

In the meantime, this just in:

An update is available:

Update V1.0.2 has the following changes:

New Features:
- none

Bug Fixes:
- Custom ratios did not work correctly
- Start button was not enabled when size was measured in inches
- Installer would stall after user accepted terms and conditions
- Saved as 72ppi instead of original resolution
- Underexposure of processed images has been improved


GUI Modifications:
- none

[Manfred M]

The issue I have when it comes to products that use Machine Learning / AI / Neural Networks is that the developers really do not have a particularly complete idea of what is happening "under the hood" with the software. In old fashioned programming the code was an explicitly written set of instructions that would effect a change. When I look at what OnOne did with Perfect Resize, they used fractals to define the elements of the resize.

With the new technology, the system behaves far differently. The input is something referred to as a "training set" where the computer software is taught by showing the input and desired output. Do that a significant number of times (likely tens of thousands if not millions of examples) and the program "learns" how to create output given a new input. If the training set has issues (built in biases, for example) the output is not going to be what is required.

Based on the comments read here, the software does not appear to be ready for "prime time" just yet because the output is not "good enough" under certain input conditions. The question that I would have is whether the issue is with the actual software or with the data it has been trained with. One issue is easier to fix than the other. If the software is not giving the desired output, the learning algorithms need to be changed but if the issue is with the data, then the learning sets have to be updated or tweaked or expanded to provide the desired outcome.

I think I will wait on the sidelines for now. The technique itself shows promise, but it certainly sounds like the software may have been released before all the kinks have been ironed out.

[DanK]

This is stretching what I understand about machine learning to the limit, but if I understand it, the training set has in it an implicit set of rules that may be enormously complex but that the machine can figure out with enough time and comparisons. One issue is the one Manfred pointed out: whether the training set is good for uncovering the implicit rules without bias. Another is that what is whether a falls within the range of what the training set is designed to represent. For example, I would want different amounts of smoothing for faces than for macro shots of a bee's eye.

This is off topic, but a lot of human learning works like this. Years ago, I watched a lecture by someone who taught new teachers how to teach reading--a difficult skill to learn in English because we have a "deep orthography", that is, no simple relationship between letters and sounds. He asked everyone in the room to give him a rule for pronouncing the letter "y." He would then write on the board an example that fit that rule and one that didn't fit. Someone would then offer a more complex rule that included the second example, and he would repeat the process. Eventually, the group had generated a rule that, although probably incomplete, was far more elaborate than anyone in the room could have produced alone. He concluded by saying that everyone "knew" the rule in that they could read English with very few errors, but none could articulate it fully. Rather, we had learned to apply it from exposure to many examples over time.

Just last night I heard another example that folks who speak German better than I do can either verify or contest. It was a comment by a German computer scientist, a native speaker of German, who said that he usually can guess the gender of an unfamiliar word even though German has relatively few obvious rules (particularly for distinguishing neuter from male). He said that there must be patterns that explain this, but he has no idea what they are.

In any case, back to the topic at hand, I prefer to know what the editing software is doing and to have the option to control specific edits directly.

[Manfred M]

Just last night I heard another example that folks who speak German better than I do can either verify or contest. It was a comment by a German computer scientist, a native speaker of German, who said that he usually can guess the gender of an unfamiliar word even though German has relatively few obvious rules (particularly for distinguishing neuter from male). He said that there must be patterns that explain this, but he has no idea what they are.

I'd agree with that.

German is my mother tongue and I did not start learning English until I started school and spoke German at home with my parents throughout my childhood and well into my late teens. I also took German in a Heritage Language program on Saturdays for a number of years.

I remember taking some German courses in high school and university and would run into problems when the teacher or professor would ask me what rule I used to use the word / gender (German has three genders, masculine, feminine and neuter) / sentence structure. I did not know the rules, but could intuitively figure this out correctly through years of practice.

The French are not nearly as lucky with their two genders (masculine and feminine), so people have to memorize the word with its gender. They also get to carry around the grammatical rule book; Bescherelle due to the complexities of the language.

[plugsnpixels]

I did some more tests today, this time with DSLR-sourced images and including ON1 Photo RAW in the mix.

[JohnRostron]

The results look good, but are they worth $85 or £66 (with the discount)? I would not want it for print upsizing, but for enlarging small images (or extracts).
John

[plugsnpixels]

John, each person would have to answer that depending on their needs. My examples show exactly what you are looking for (enlarging image extracts); on another forum a person with high-end equipment was very pleased with the results he got from enlarged and printed images.

[plugsnpixels]

I received an email today asking why I hadn't also tested BenVista PhotoZoom Pro 7 against Topaz Gigapixel. Well, here you go ;-)

[JohnRostron]

I do have Ben Vista which I was quite impressed with. Many years ago, I scanned a number of slides, but i scaled them down for use. I would now like to scale them up again, hence my interest.

John

[plugsnpixels]

Some more fun on a quiet Sunday, changing pixel mush into something usable with Topaz Gigapixel!

600% upsampling of the lower left of this storefront image (shown here in reduced size for reference) as viewed in Photoshop at 100% (native pixels/Gigapixel).

[JohnRostron]

I tried out AI Gigapixel and discovered that it will not work because my graphics card is not up to it! The website does list incompatible cards and mine is one of them. Is is going to be worth upgrading my old card just to use this? I would guess that most modern cards will work, but I am looking at around £100+. Would there be any other advantages in a new card? I am not a gamer.
John

[plugsnpixels]

John, how old is your overall system?

At any rate, hold off on the new card until at least next week...

[JohnRostron]

John, how old is your overall system?

At any rate, hold off on the new card until at least next week...

I would guess that my system is around ten years old, although I have added more memory. The specs are:
CPU: AMD A6-3670 Quad core
RAM: 16.0GB DDR3 @ 666MHz (9-9-9-24)
Motherboard: ASUSTeK COMPUTER INC. F1A55-M LE (FM1 )
Graphics: 2047MB NVIDIA GeForce GT 630 (Undefined)

I am in no rush to upgrade my Graphics card, although I was planning upgrading to a SSD later this month.

John