2024-04-13

Last modified: 2024-04-13 19:33:56

Neural net images

Another idea I had was to do an initial training where it has just one picture, the same picture at every z coordinate, and then after that has reached an acceptable quality level, then limit the first picture to z=0 and add the second picture at z=1.

But first I'm going to get a collection of images that are sort of "similar" to each other and make it learn them all at different z coordinates. Another way of thinking of z is that it's t, i.e. you're giving it keyframes and it's learning an animation.

Here's some F1 driver headshots, let's try these:

I wonder if there's a better sort order? For example have the same uniforms together so that it has half a chance of identifying "like" parts of the uniform and moving them over, instead of fading one out and fading the next in.

Looks interesting so far, needs a lot more training though: https://www.youtube.com/watch?v=ajeWBiwLKkw