Diffusion models can pull noisy samples toward the dataset’s center, a user notes
When a sample is mostly noise, the model’s best guess is close to the mean, the post says—even when no data exists at that center.
TLDR
A user describes an early detour in diffusion sampling: when a sample is mostly noise, the model’s best guess is close to the mean. According to the post, that can pull samples toward the dataset’s center—even if no data exists there—before they rebound toward the data manifold, where the data lies.
Combined views
28.6K
1 Source, first seen 23d ago