Image Details
Caption: Figure 2.
Graphical representation of conditional denoising score matching. A neural network gθ is trained to denoise noisy model parameters xt generated by passing prior samples x0 through the forward SDE Equation (4). The neural network does so given t and observations y generated from x0 by the forward model. It outputs an estimate ﹩{\hat{{\boldsymbol{x}}}}_{0}﹩ of x0, and the dashed line represents the loss function Equation (10).
© 2026. The Author(s). Published by the American Astronomical Society.