How many partitions P does the backbone need to reconstruct a field within 1% error?
Checking the server…Smooth functions of many families, from waves and bumps to vortices and saddles, and sums of radial sines of several bands, on 256 × 256 pixels.
Every field was reconstructed at every power of two P from 1 to 1024 with the backbone’s inference partitioner; the answer is the fewest partitions with a relative L2 error below 1%.
Every partition is normalized and encoded on its own, so more partitions leave less variation in each, until each looks like the smooth functions the backbone was trained on.