Diffusion noise seeds contain universal compositional blueprints that persist across datasets, enabling our Patch PCA framework for zero-shot structure control.

We study a new class of spectral algorithms for low-rank estimation based on a tunable nonlinear deformation of an observed matrix.