DYffusion is a reimagining of the diffusion processes to more naturally model spatiotemporal sequences by replacing the forward process $D$ and reverse process $R{\theta}$ of standard diffusion models with a stochastic interpolator network $I{\phi}$ and a deterministic forecaster network $F_{\theta}$
Context
Advantages:
Efficient (Memory footprint): Given a time horizon $h$, DYffusion requires only $xt$ and $x{t+h}$ to train, resulting in a constant memory footprint as a function of $h$
Relations to ODE: The neural networks $I{\phi}$ and $F{\theta}$ are trained to construct a differential equation that implicitly models the solution to $x(t)$. DYffusion models the dynamics as
where $s$ is the time variable.
Disadvantages:
The model doesn't support the form where output space is different from the input space
Arxiv/Blog/Paper Link
Arxiv Blog Code
Detailed Description
DYffusion is a reimagining of the diffusion processes to more naturally model spatiotemporal sequences by replacing the forward process $D$ and reverse process $R{\theta}$ of standard diffusion models with a stochastic interpolator network $I{\phi}$ and a deterministic forecaster network $F_{\theta}$
Context
Advantages:
where $s$ is the time variable.
Disadvantages: