[Code Idea]: Diffusion Predictor

[galegozi]

Summary

The idea is to replace the predictor portion of the model with a diffusion model. This takes inspiration from popular image generator models, like stable diffusion.

Motivation

This is designed to help with repeated timestep prediction, where the outputs from the model are fed back into itself to predict future timesteps.

Affected Area

This affects the predictor section of the model, while still maintaining mechanics from climaX/Aurora papers to ensure the diffuser receives a constant size input.

Proposed Approach

Look at how diffusion models are built in other documentation/papers.

Alternatives Considered

This is intended to be a separate model and people using Yoke can select which model to use.