>>4872551

There's not much I can say, it's not my program and I only know enough to run it. I only guessed that the temporal consistency feature can be generalized because the developers made a faster version using a different technique, though it's almost useless atm because the developers never released the tools needed to use your own style images. https://github.com/manuelruder/fast-artistic-videos

There was another promising technique for temporally consistent video used last year, but it was never released publically.. https://www.youtube.com/watch?v=BcflKNzO31A

There are generally 3 types of style transfer programs, the ones that sacrifice speed (>>4827535), quality (>>4836008), or the ability to use arbitrary style images. For the latter, you often have to spend hours training the software on whatever image you want to use for stylizing. Everyone's focused on improving speed over quality, so the highest quality program is the oldest, by user jcjohnson. It's the program I used for >>4827535
https://www.reddit.com/r/MachineLearning/comments/8o846n/d_what_is_the_best_implementation_of_neural_style/ (sorry for plebbit)

This is a good starting place for stylizing still images, packaging two techniques for style transfer plus the "multiresolution" technique for jccohnson's program to improve quality further, though ti does require CUDA, Linux and Docker. It's gonna be amazing when this stuff reaches the masses. https://github.com/albarji/neural-style-docker