For generating future NWP outputs, the inputs need to include the NWP data, but possibly also satellite imagery
If the inputs to the diffusion models are images, then the inputs can be a multi-channel image of NWP data at 0.25 degree (or whatever resolution) inputs, so it would be easy to layer on top of that downsampled satellite imagery, although it would need to then take in all the main geostationary satellites, which have different numbers of channels, and update times.
For generating future NWP outputs, the inputs need to include the NWP data, but possibly also satellite imagery
If the inputs to the diffusion models are images, then the inputs can be a multi-channel image of NWP data at 0.25 degree (or whatever resolution) inputs, so it would be easy to layer on top of that downsampled satellite imagery, although it would need to then take in all the main geostationary satellites, which have different numbers of channels, and update times.