I've run neural-style on my GPU at home and each frame at high res took ~ 15 minutes to converge. So 5 * 60 * 24 * 15 = 1800 Hours GPU Time * 65 cents an hour on EC2 = $1170. So costs aren't that low.
If it looks like what the client is looking for (a very big if), those are very cheap frames. If you hire someone to manually mangle 5 min of video it will usually end up costing way more than that.
At this price point, it's time to start looking at alternative approaches. The original iterative optimization approach is not the only, nor is it remotely the fastest. (There's an upscaling CNN version which runs at like 20FPS but I can't seem to refind the citation at the moment.)
Yeah, that worried me. The pessimist in me suspects that maybe they don't want to make a lot of money selling it, maybe they just want this hypothetical price point to exist when they go out suing people doing similar stuff (the deepart.io people have patented some of the early research in style transfer).
I wonder if there are ways to reduce the GPU RAM requirements on this. Even the commercial offering doesn't go beyond 720p - I guess that fills up 12GB on a Titan X?
On stills style transfer you can process the image in multiple tiles, but I don't expect that would work with video with the temporal consistency
I'm pretty certain I saw this exact set of videos posted on another article yesterday, however.
Edit: My bad, it was via a youtube link I saw elsewhere, thats also been posted in this thread.
19 comments
[ 4.0 ms ] story [ 45.9 ms ] thread- video: https://www.youtube.com/watch?v=Khuj4ASldmU
- paper: Manuel Ruder, Alexey Dosovitskiy, Thomas Brox, "Artistic style transfer for videos" https://arxiv.org/abs/1604.08610
- code: https://github.com/manuelruder/artistic-videos
https://arxiv.org/abs/1603.08155
On stills style transfer you can process the image in multiple tiles, but I don't expect that would work with video with the temporal consistency