In my previous post, I mentioned how drawing instructions could be executed as soon as GPT produces a token. Yet, this approach is also not ideal because geometric shapes would appear on the screen immediately, and it doesn't look natural.
I used a technique that transforms a signal of x and y coordinates into a Fourier series, which produces epicycles that are more visually interesting.
As an example, given this test image:
The final result looks more advanced. Instead of individual geometric shapes, it produces a list of a single-line drawings.
I believe that NNs can identify correlations between clusters of epicycles, provided there is a dataset large enough. I may start with the QuickDraw dataset to test things out, it includes a list of 345 categories, but this is likely not enough to generate more advanced and interesting drawings.