Well, not so much to say at this point. At the lab, we have started working on the problem of style transfer (without the use of CNN or course :p). And we have asked David Revoy, a real artist, to help defining what is really a successful style transfert, and more generally what is the “style” of an image, etc… so we can try to model it mathematically, but from an artistic point of view.
That’s really the start, and I hope we’ll get something to put into G’MIC some day
There is a new paper (22.03.2017) with a not painting style approach. Seems it is very good on lightning afterwards, like in night for day shots, or the other way around.
There are some functions about the method in it, but its all jibber jabber for me.
But maybe David Tschumperlé knows more about it and how to use it in GMIC.
It would be really great addition to Natron, and open source in general.
Adobe has got his fingers in it, so its likely to see in Photoshop, Premiere Pro and/or After Effects…
One of my hobbies is reading about image processing. It is a love-hate relationship because I don’t have a background in mathematics or programming, and as a non-scientist, I have limited access to papers, code and resources. In terms of papers, I really dislike the research (patent) paywall (red tape).
Anyway, here are some thoughts that come to mind after skimming through this thread:
Machine learning (ML) is certainly a hot topic nowadays. What I am wondering is: Does G’MIC have the capability to address machine learning? If not, is it feasible to extend it to the point where it can? Or is there another open source project that could better address the problem?
@iarga brings up a good point that shouldn’t be missed.
The genius of ML is that it can be adapted for (m)any type of problem(s). It doesn’t have to necessarily be about the transferring of artistic style. So, I think we can split the discussion into two problems: one about ML and the other about transferring artistic style. As @David_Tschumperle put it:
As for yet-another-paper, a quick search yielded this one from IPOL: http://www.ipol.im/pub/art/2016/150/. I know it isn’t exactly about artistic transfer and I don’t have time to read it and play with the demo to see if it is on-topic, but we can always learn from how others approach a similar problem.
When it comes to ML, we don’t necessarily have to start from the beginning. There are many pre-trained databases out there that can be used with some caveats after some tweaking. It will be a compromise but a small community with limited resources probably couldn’t match their scale and quality.