Energy-Tweedie: Score meets Score, Energy meets Energy
The paper derives a denoising-to-score identity for a broad class of energy-based noise distributions, extending Tweedie’s Gaussian result.
The Energy-Tweedie identity connects the full denoising posterior to the noisy distribution’s Stein score through a kernel scoring rule defined by the noise potential. It provides a route to estimating scores from posterior samples and a criterion for estimating unknown noise parameters. It also enables diffusion-style sampling along user-chosen paths through noise-parameter space.
HF Daily Papers' note
The Energy-Tweedie identity connects the full denoising posterior to the noisy distribution’s Stein score through a kernel scoring rule defined by the noise potential. It provides a route to estimating scores from posterior samples and a criterion for estimating unknown noise parameters. It also enables diffusion-style sampling along user-chosen paths through noise-parameter space.
HF Daily Papers' note
score 4