Highly recommend you watch everything on his channel. He’s super smart and very entertaining
A nice use of SDFs. I wonder about some sort of regularization or constraint on the outputs from training, because SDFs should be continuous, which induces a triangle inequality on every pixel relative to the implicit edge of the font. Edit: After spending a few minutes at the whiteboard, I think that the inequality is on every pair of pixels, and might only be useful for pixels whose values in the field have different signs.
Valid SDFs have a spatial derivative with an amplitude of 1: The square of the x and y gradients should add up to 1. SDFs are a special case of eikonal equations.