Here at Smells Like Revenge Productions We’ve always tried to combine Math, Science, History and Art in, hopefully, interesting ways. Now we’re taking it a step further. Blake Ellender has now finished coding an Ai to speed up rendering animation frames, or in layman’s terms, how fast we can make videos. This Ai takes frames from the animation software we use and ups the resolution. Essentially the Ai takes a pixelated image and then draws in the corrections based on images it’s looked at before.
As of now this Ai is doubling the resolution of each frame (quadrupling
the number of pixels). This is equivalent to taking a 540-pixel by 960-pixel
video to a 1080-pixel by 1920-pixel one.
Here is an example of a successful upscaling done by the network:
From the upcoming music video for Where the Mountains Meet the Sky by To Say the Least. Animation by Blake T. Ellender |
For the super nerds:
This Ai is based on various VDSR CNN architectures. It is
much shallower than most I found mentioned online, only a dozen or so layers
deep, but works quite well for these purposes. This is helped by the fact this
application is looking for what looks best NOT WHAT IS 100% accurate. This
creates a purposeful “bias” that comes out in the lighting and field of depth
within the scene. This is achieved by have a larger allowance of loss in the
Convolutional Neural Network than may be desired in non-artistic applications.