Variational inference often relies on tractable families, but semi-implicit variational inference (SIVI) lifts that restriction by defining the variational distribution as a mixture or expectation over a conditional distribution. The added expressiveness comes at a price: the marginal score of the semi-implicit distribution is generally intractable, so optimizing the evidence lower bound (ELBO) is hard. Existing unbiased methods, such as Unbiased Implicit Variational Inference (UIVI), estimate the score through reverse-conditional sampling or approximations, which can be computationally costly and unstable.

Denoising Implicit Variational Inference (DIVI), from Wang, Wu, Yu, and Zhang, takes a different route. Instead of estimating the score indirectly, DIVI learns the marginal score directly via denoising score matching. The learned score network is then used as a plug-in estimator in the pathwise ELBO gradient, replacing inner reverse-conditional sampling with a single network evaluation. The authors analyze DIVI as inexact stochastic gradient ascent and show, under stated assumptions, that the averaged stationarity measure is controlled by the usual optimization term plus the average score-matching error.

The paper reports experiments on synthetic distributions and several real-data Bayesian inference tasks. DIVI improves over current UIVI baselines while reducing the cost of ELBO-gradient estimation. The findings come from a single arXiv preprint accepted as a Neurips 2026 camera-ready paper; there are no conflicting sources to compare.