Multi-Diffusion

Note

The multi-diffusion (patch-based diffusion) utilities are under active development and will be released in an upcoming version of PhysicsNeMo.

Multi-diffusion is a technique for scaling diffusion models to large spatial domains by splitting the full domain into overlapping patches, running the diffusion model on each patch independently (potentially in parallel), and fusing the generated patches back into a coherent full-domain output. This is particularly useful for problems where the full domain does not fit into GPU memory.

For a preview of the approach, see the positional embedding architectures SongUNetPosEmbd and SongUNetPosLtEmbd in the Diffusion UNets documentation, which provide the backbone support for patch-based methods. The CorrDiff example demonstrates a full application of patch-based diffusion for weather downscaling.