Neural Deep Learning

[Sam]

Please see this :
https://macdaddy.io/style/
they are trying to bring deep learning style paintings to desktop (like pikazo for iOs),

Adobe is trying too:
https://research.adobe.com/leveraging-deep-learning-to-fix-images/

Is it the time for SA that looks these and does some integrating too,
Or you think Deep Learning Is not good enough?

thank you

[Farshore]

Topaz Labs has a similar product: https://topazlabs.com/ai-remix/

[Paul Perlow]

I have looked at these software alternatives and am grateful for the information. They raise significant questions. Is one satisfied to travel the path of a copyist or venture into the realm of true creativity?

I am constantly surprised and happy to uncover/discover the new, the realm of creativity as I experiment with the myriad of options and possibilities offered in Studio Artist. Why convert or use Studio Artist as a copy machine? I am constantly amazed by the results that come from experimentation in the digital playground offered in Studio Artist. Leave the domain of copy to these other softwares.

Paul

[Yuya Takeda]

Ok personal opinion time: Style and AI-REMIX programs use what's known as the fast neural style (feedforward network). The tendency for this algorithm is that it makes the style transfer very fast on large images, but it also makes quite a mushy, kinda hideous, but accurate-to-style look. The generated results often introduce a bunch of chatter and unnecessary noise which results in an overall dimension-less, gargled image.

The code is pretty much free for you to run it yourself, there are already a gajillion programs that do the same.
https://github.com/jcjohnson/fast-neural-style

To be fair, the pace at which this sort of NN (neural network) is developing is incredible, we see improvements in just a few short years (even months at times). So it'd be probably unfair to say the same in a couple years time. Algorithms, methods, and ideas will definitely improve the overall image quality so I hope to see that.

For higher quality results currently we have to resort back to the original context of this research. Originally this took a crap load of time to render and output, especially for larger images. But personally, this is far better than the current generation of feedforward network results.

There are a couple websites that offer high resolution output, one in which is:
https://art.app/

The resulting output from art.app is not only impressive, it's far better than anything that the fast-neural style can produce as of today. The problem with the image generation is that it takes a huge amount of processing power to accomplish it. It takes some tremendous amount of GPU power to just get these fantastic results. I personally gravitate towards this however, artistically, quality-wise incredible. Below are some extreme test examples.

By the way, these are the same source image, just with different styles. One was processed far extreme than the other.

This isn't to say that those fast-transfers are bad or anything. I'm just saying that the aesthetic output kinda feels at a loss somewhere. However, it will be fantastic to see something of this nature in StudioArtist!

[Bernard Bunner]

I have tested many neural networks software, Pikazo, Art.app, Artisto, DeepArt Effects, Style, etc. The most interesting, with the best result is Pikazo. I played with it 2 years, and posted results, here and here (Pikazo, SA, Filter forge pictures).

Yes, as Yuya wrote, (The generated results often introduce a bunch of chatter and unnecessary noise which results in an overall dimension-less, gargled image) it's a problem. But it you modify your source and style pictures, the results are better. Sometimes, I deleted black or whites regions, etc...Or just turn 90° my style..Some interesting results also when you revisit 2 artists, like Picasso with Kandinsky : 

or a lion head, "mixed" with a metal structure :

or something like this, with :

Art.app : the deep effect is more "regular"

I have my own methods to make large pictures.

But there are no settings where you can modify some parameters.

Yes, a "deep effect" tool in V6 would be great....

[Sam]

I think as a tool just like other tools of SA, Studio Artist is better think of Deep Learning as soon as possible, It must ride it and not stay under the wave. 

Dear Bernard,
you can see my works here:
https://deepdreamgenerator.com/u/smmsamm

but I like your pikazo result and tested Pikazo too, never could get Large and high quality from it,
Please tell us how to enlarge with this quality?

thank you

[Sam]

thanks, I found two website that can enhance and increase the resolution by Deep Learning algorithm:

https://deep-image.teonite.com/
But It shows me error "fetching filters failed" maybe in my country it is banned.

and

https://letsenhance.io/boost
that I checked and liked it.

I hope if you know better ways tell us here.