Let Assumptions1 and 2 hold.

In this way, the model can mine distributional knowledge about the target scenario and quickly adapt to it with enhanced generalization performance.

Recent works [76, 33, 70, 72, 17, 7, 66, 20] started to explore sparse-view novel view synthesis, specifically focusing on generating novel views from alimited number of input images (typically 2-3) with known camera poses. Some of them [33,70,72,17,7] introduce additional priors into NeRF, e.g.

Thefirsttwocolumns showresults for two different step-sizes, and the third one using the best step-size chosen retrospectively. The plots show the final ELBO achieved after trainingfor40000stepsvs. stepsizeused. Figure 11: VI using a diagonal Gaussian, with the best step-size chosen retrospectively. Bayesian logistic regression: We use a subset of700 rows of thea1a dataset. In this case the posterior p(z|x) has dimensionality d = 120.

Flexible variational distributions improve variational inference but are harder to optimize.