2024-04-08

Last modified: 2024-04-08 22:20:06

Neural net images

Following on from https://incoherency.co.uk/blog/stories/lenna-neural-net.html

What if instead of (x,y) => (r,g,b) we mapped (x,y,z) => (r,g,b), where z=0 for some base image, and z=1 for another, and then we can animate varying the value of z to get a cool transition between the first base image and the second? We won't train on values of z other than 0 and 1, but we can still sample it as a continuous function. What would the intermediate positions look like?

And we can also generate images/animations where the value of z is not constant across the entire frame. For example you could start with every pixel having z=0, but sweep a wave across from left to right where it changes to z=1 gradually. What would that look like?

This easily generalises to more than 2 images. For example, train on a third image at z=0.5.

I've got it running with a picture of me at z=0 and Richard Feynman at z=1, saving out frames for z=0,0.25,0.5,0.75,1 after each round of training. I'll let this run for a while and if it looks interesting I'll try to make a higher-resolution video of transitioning between 2 other pictures.

It's not obvious to me whether the results would be most interesting when the 2 endpoints are very similar, or when they're very dissimilar.

Also, what happens if you extend beyond the bounds of [0,1]? Not only in z, but also x and y for that matter!