We propose score-based diffusion framework with explicit G-modeling, a nonparametric empirical Bayes method that achieves near-parametric score estimation guarantees and state-of-the-art denoising for multivariate heteroscedastic Gaussian mixtures.