I recently purchased and started using this Topaz app. I love what I can do in SA with supersizer, but when it comes to edges, it can create a "scribbly" loose effect. That does not work for me with geometric shapes and typography in where accuracy is critical. The Topaz software is great. It allows enlargement with color accuracy up to 600%. It has different categories for photos, art and architectural with even facial recognition. There is a default setting or one can adjust sharpness and noise independently. That is good but it also can sanitize the image and make it too perfect. It can appear too cold or slick. I find my way through this maize by using Supersizer at first to a smaller than needed end point and finish it with Gigapixel. I then have the best of both worlds. With large files, however, Gigapixel is very slow and requires patience when saving. I definitely favor the warmth and humanity of my imagery in SA. And the Topax product is excellent and helps a lot. They have a free download for testing. Last time I looked it was good for a month.
We're pretty actively looking into putting neural net interpolation directly into Studio Artist in a future release. Hopefully with our own unique twists.
But if you have access to something like Topaz now, i would really recommend considering it in your workflow. I think there are a lot of advantages to working with a smaller working canvas size to build up your image. Then use neural interpolation to get it up to your final size. As opposed to starting with a working canvas that is the final size you want (unless that is your working canvas size).
Here's one specific example below. But keep in mind that everything gets slower, and in many ways more clunky to work with as you endlessly keep increasing the canvas size. There is a better optimal working canvas size for getting work done quickly. That's going to be a function of your computer's speed, the size of your display, personal preferences, etc.
We've noticed this as an issue sometimes for people working with vectorizer effects. Just endlessly making the canvas size bigger isn't necessarily the right approach. Because the source image is interpolated to that canvas size as well for use as input to the Vectorizer algorithms. So you can end up with a different looking image because of this, and usually the smaller typical working canvas size is the better looking one.
And of course the Vectorizer is a memory hog, so it's going to run slower as you move up to really large canvas sizes.
Of course with Vectorizer output you could output it to svg, and then import that svg file back into Studio Artist and have it import it into a much larger canvas size than the one you worked with. So there's the solution to this issue today for anything you can output as vector svg files from Studio Artist.
You get what the canvas looked like at the working canvas size beautifully scaled up to your final desired finished canvas size.
What's nice about neural interpolation is that it works with raster images. And you can think of it as a much smarter more modern algorithm for generating interpolated images. Part of how this works is that the neural net has been trained on a large database of images. Trained to interpolate in a natural looking way. Natural means 'look like the images it was trained with'.
The way the training works is that you start with high quality natural images at the resolution you are shooting to interpolate to. You then generate a copy of this database of images that you resize to a smaller resolution. You can also add noise and or further augment that set of original smaller sized training images. You then train a convolutional neural net to learn how to interpolate the image. And it does.
In a real abstract way you can think of convolutional neural networks as super-charged versions of the non-linear filtering operations inside of Studio Artist. But these non-linear algorithms are able to learn off of training data. They can also in theory 'replicate any transformation'. The non-linear model is able to learn any arbitrary mapping. That's pretty amazing.
In the case of neural interpolation, the nonlinear model is learning the correct transformation to enlarge images that look like the images it was trained on.
Now if your particular image doesn't necessarily look like what the neural net was trained with, then it's going to convert it to look more like what it was trained on. And that would introduce artificial looking visual artifacts.
Another interesting observation is that the neural net is actually synthesizing a new image from scratch. As opposed to more traditional interpolation algorithms that are very closely tied to how their input looks, and then essentially smooth it out to make it larger. Neural nets are generating a whole new image from scratch. And there are some interesting implications associated with that.
Because the 'image synthesizer' could also be trained to do other different things at the same time it was trained to do good interpolation that looks natural.
And with that thought to ponder, i'll conclude for now.
I have been using the Topaz Gigapixel and have further thoughts and feelings about it. To be blunt, I don't like what it does with my images. It actually changes the structure of the pixels and does not represent my original art. It sanitizes the linear quality to a perfection that changes the nature and overall appearance (and feeling) of the art. I have tried adjusting it manually and that helps some but not enough. One size does not fit all. I suppose the algorithm(s) they are using will please those who like a slick perfect appearance. But one of the reasons I love Studio Artist is that I can create art that has a natural feeling that does not have the appearance of something computer generated. The woodcuts are an example of that.
I would be wary of using this neural approach in a way that changes the nature of the original art and alters the artist's intention.
So..what to do?... Fortunately there is another alternative that I already had ( I actually "forgot") and am very happy with. On1 Photo Raw 2020 has a resizer that is truly excellent and appears lossless. (These are the people that started with Genuine Fractals many years ago.) The results are more truthful and there are many more options that also allow me to add grain, etc. It is also much quicker than Gigapixel, and does not drain the energy from my computer. The learning curve is steeper but what you get is much greater and there are so many more options.
Topaz has developed a business plan that was very smart. Rather than offer a "Photoshop like" alternative, they market individual modules that end up being expensive. Price wise On1 software is reasonable and so much more more bang for the buck.
I suppose it would be interesting to build a neural interpolation algorithm that generates fractal information as it expands. So i will think about that as well for the future.
I was going to do a tutorial on how to take a raster woodcut preset and turn it into vector, so that you could output svg vector file and then re-render it to the final raster output size you want.
I would love to have that tutorial. Thanks, John.
Will love too