Thanks for the news tip, @shodan5000,
Topaz Labs might be interested in this. Though I think they follow what NVIDIA is doing because it’s what business demands.
As for me, I’ve already read it, but I don’t use it because I designed my own PyTorch-based system, which is more advanced than the solution described here. I prefer to develop my own work rather than copy someone else’s.
Pytorch is more convenient when designing your own neural networks, TensorFlow is more complex. I personally dropped out of TF a long time ago and am now using P. Opportunities keep growing.
GANs are still impressive, but this is a very old idea.
You can do great things with GAN, but we developers, inventors, and graphic artists need something a bit more modern. This idea is slowly coming to an end. There are an endless number of programs that use GAN, but they’re still just toys. Few hobbyists like me use it professionally.
There are two reasons:
- you must have a very large interdisciplinary knowledge and artistic skills to be fluent in it and do whatever you want. I mean, to be able to really use and see the possibilities that are hidden there (you need to learn: programming several languages, graphic environment, hardware, electronics, optics, physics, geometry, neurology, photography, even holography, because there are some things useful in other fields, on: the mathematics of complex numbers, useful in the field of FFT and the whole great field of Digital Signal Processing and 100 other things. Having to study all of this puts off many artists. It’s a serious barrier.
Meanwhile, when you assimilate it all, you not only have a job in almost every corporation (although I have never worked this corporation way, because I bet on my own business), but above all you have proficiency, pleasure and incredible ease in creating new things.
I think I wrote about it before when testing something, but now I can link it to your topic, which makes more sense.
It’s not a problem to copy and run someone else’s script if you only have a little knowledge. But that’s not what I meant. Anyone can copy, but it’s not fun and good. The most interesting thing is creating your own stuff, isn’t it? So I encourage you to learn and develop your own interdisciplinary knowledge.
- more complex architectures like SPADE or MUNIT require fantastically powerful computing machines. Your home graphics cards are too weak for this. On ordinary computers you will wait 2 months for the models learned, and then create 256 x 256 px images.
You can, of course, buy cloud computing, where the configuration is strong, but this in turn requires paying for access. If you want to make a book like me, with 300 dpi quality and dozens of illustrations of 4000 x 3000 px - instead of a micro image for fun - then you have to invest thousands of dollars, because in the cloud, calculations are counted by time. You will pay the most not for your finished design and image, but for a huge amount of trials and tests to get the right model and architecture of the network. Because if you’re an inventor like me and you’re designing new types of AI, you have to rehearse like Edison. This is the essence of invention, not copying someone else’s scripts from github. It is these tests and research that cost so much.
- There is also a third reason, an additional one: all this knowledge about AI is still fairly superficial (despite many scientific studies). Even though the theme is trendy, there are very few commercial programs such as PC or MAC that use AI to create art. Because? because of  (I am linking to the same paragraph here recursively). And because of .
It’s a bit like gold. This is a rare metal, so it costs a lot. That’s why there are no good commercial desktop software. Instead, “businessmen” and combiners try to make money on the minimum knowledge they already have and the so-called “kitsch”, because it always sells well
The knowledge is small, but it is enough to make another hopeless “AI app” to turn a friend’s young face into an old man, or to draw tattoos or ears on his face in real time during a chat. So there are online applications or hundreds and thousands of rubbish and stupid toys for mobile phones like Prisma, which sends your images to undefined Russian servers, plus “face fixers” (to look better on Tinder or Facebook), automatic wrinkle removal, etc.
It all creates a pile of rubbish and worthless adult play gadgets instead of real tools.
Likewise, spam, kitsch and violence dominate television rather than valuable programs. People just like it, so the producers deliver it. This is the essence of how capitalism works. Deliver what sells.
No sane and normal person likes violence or war. Despite this, people like to watch crime fiction and movies about the war, for example with the family over dinner. From the point of view of science, it seems we humans just have a series of bugs in the mind’s software.
But I will not expand this thread because I did not write this text to save the world
So the architecture of these toys is still primitive because  and it will be a long time before we have a real AI for artists.
Against this background, Topaz Labs products seem like a light in the tunnel and a certain consolation. You don’t have to learn anything or bother with complicated installation. They just wrap all this complicated technology in a nice little box and you get what’s easy: a few icons and push buttons.
In addition to praising Topaz, I also have a proposal that this company should make more innovative products for artists, and not only improve a few of the same applications (sharpening, denoise) - because everyone does it.
It’s like a supermarket: why do I need 20 kinds of jam, ham or cream? They all taste the same. Only the logo changes. Similarly, there are more and more applications for sharpening, magnifying and de-noise on the market.
Do something really original and you’ll be ahead of the competition (only for a while, of course). But you can always be a few steps ahead of the others, right?
For example, there is still no commercial AI application for making high-quality pencil sketches from photos. What is on the market is so primitive that it is not suitable for commercial use.
Or another model: turning photos into line art (used in coloring books). AI has great possibilities here, because traditional algorithms based on edge detection or even FFT filtering are not suitable for this, because creating an artistic drawing does not consist in simplifying the image and simple filtering, but in decomposing the image into elements and synthesizing - that is, applying process of abstraction. For this you need to use several coupled cascade models, a supervisor and an image library (for example).
Another application: creating photos, also high-resolution drawings from ready-made elements and a handwritten sketch of the user or from a small image (or even from a fragment of another photo). This is not about superresolution like in Gigapixel, but about redrawing large pictures from small sketches and photos.
Or the user provides not one but several small images. AI looks at it all, recognizes objects, then selects its own objects from the library, and from low-quality sample images produces one large “redrawn” or “re-photographed” or “synthesized new image”.