laplace-torch v0.3 aims to bring extensive support to diffusion models. These models are challenging due to their (often images) output dimensionality. Considering that the linearized Laplace requires the (n_outputs, n_params)-shaped Jacobians, this can be very challenging (e.g. images have n_outputs = width * height).

All features we aim for are:

• Efficient, universal, ultimate Jacobian backend
• Functional Laplace with subset-of-data support
• Heteroskedastic Laplace
• Fast last-layer predictive