Score Matching and Score-Based Models

Denoising Paradox

Phenomenon where training a score model on noisy data (denoising) yields better results for estimating the score of clean data than direct training.

← Indietro