Observe a field through a window. The backbone encodes the window alone, as one partition and so one embedding, and decodes the whole field from it, beyond what it saw.
Checking the server…Only the pixels inside the window are given to the model; everything around it is hidden.
The window's pixels, an unstructured cloud of points, form one partition: they are normalized into its frame and encoded into a single embedding on the sphere.
The decoder is a function of the coordinate, so it can be asked about any point: every pixel of the image, inside the window and outside it, is decoded from the one embedding, which continues the field past the window's edge.
An interpolation cannot reach beyond the observed points, so the baseline repeats the nearest observed pixel. The peak signal-to-noise ratio (PSNR) is measured outside the window only.
The backbone was never trained to outpaint. A smooth field continues well from one embedding, since it describes a smooth function that does not stop at the edge of what was seen; a detailed field or a photograph does not, as nothing outside the window says what is there. The presets are smooth fields that one partition represents reliably.