<ul><li>Diffusion models are widely used for generative sampling but can be prone to memorization, leading to overfitting of the learned score on finite datasets.</li><li>This study introduces a kernel-smoothed empirical score to reduce noise variance and analyze its bias-variance trade-off in denoising diffusions.</li><li>Regularization on the score helps prevent memorization by increasing the effective size of the training dataset, leading to better variance control.</li><li>Experiments on synthetic and MNIST datasets demonstrate the effectiveness of the proposed approach in enhancing generalization and mitigating memorization.</li></ul>

Kernel-Smoothed Scores for Denoising Diffusion: A Bias-Variance Study

Discover more