@MichaelHale These days, one might always consider trying obj["Methods"] or obj["Properties"] to see if WRI has hidden anything into any new data type they introduce.
Does anyone have a link to a semi-technical description of how this new Prisma app works? I know it uses neural networks, but I'm curious about the details. I know very little about neural networks.
@MichaelHale Mentioned here on chat. Also, I think I have ran into it earlier on somehow...
@Szabolcs I doubt there's description of Prisma specifically anywhere. I guess they have somehow written specialized neural networks for specific filters (it's definitely dramatically more efficient than neural-style).
The fact that they use large image classification network to separate content and style from each other is sort of mysterious (==I'm not informed enough) to me.
To me this stuff is sort of "OK, this explanation makes me feel I understand it, but no, I couldn't really invent constructs that accomplish similar things, only reuse them."
I fed character-by-character recurrent neural network couple megabytes of my writings earlier this year. I think I understood reasonably how training in that case worked. The network babbled in reasonably coherent words, somewhat coherent syntax, but only marginally understandable content. :I
Interesting nonetheless.
As far as "painting generators" of neural network style go, they are based on the fact general object recognition neural networks encode knowledge of what world and objects in it look like in them, and this general information is used for those artistic interpretations... but that's awfully non-technical interpretation.
@Szabolcs, it doesn't seem like Moiseenkov has released any technical details, and I'm pessimistic that he will, considering how successful his app is.