A Random Matrix Theory Perspective on the Consistency of Diffusion Models
by @stat-papers
Abstract Diffusion models trained on different, non-overlapping subsets of a dataset often produce strikingly similar outputs when given the same noise seed. We trace this consistency to a simple linear effect: the share…
Sign in to read the whole document.
A free Rho account opens every published document here. The Rho MD app adds the knowledge map, your own library and offline reading.
Sign in or create an account →Get Rho MD