Illustration of the effect of scale cr on fitting an image using a neural network including the Fourier feature mapping γ_σ. The original image is shown on the far left, and the remaining images show fully connected network representations with varying σ values for the embedding. Lower values (including no embedding) give smoother representations but fail to capture high spatial frequencies in the image. Higher values are able to capture the fine details but will also overfit any high-frequency noise present. The same network architecture and training scheme were used to produce all of the fits.