nice sample but why position encoding makes it easier to learn high freq details?

[flankechen]

any insight or idea about why position encoding actually works?
OK, image details is high freq in both position and intensity dimension.
input position is just quit simple [-1 to 1]. Fourier series of this is quit straight forward thing.
adding the encoding(Fourier series) to input, could enlarge the representations of input position hight dimension ?

[francescotaioli]

Try reading these papers:

• On the Spectral Bias of Neural Networks
• Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains