Paint holes into a field. The backbone encodes only what is left, an irregular cloud of points, as one embedding of the whole field, and decodes every pixel from it, the holes included. The field is 256 × 256 pixels and always one partition, so the presets are fields one embedding represents accurately.
Checking the server…A preset hides large holes over whole crests and troughs; with Paint your own, everything you paint on the input is hidden from the model.
Only the observed pixels are passed to the model: an irregular point cloud, normalized and encoded as a single embedding ζm of the whole field.
The decoder answers any coordinate, so ζm is decoded at every pixel, observed or not: the holes are filled from the same embedding as the rest.
The baseline is the piecewise-linear interpolation of the same observed pixels. The peak signal-to-noise ratio (PSNR) is measured on the hidden pixels only.
The backbone was never trained to inpaint. One embedding captures only the lowest frequencies of a field, so the presets are smooth fields it represents accurately.