Builds on top of Flow Matching.
The network still predicts velocity but it takes in one additional input or the distance to jump towards the input data from the current .
The improvement here is that since the model is now parameterized on as well, we donβt need to take fixed jumps anymore.
How is it trained?
We have:
, which is the same step size as normal flow-matching. For in , we have a bootstrap loss:
where,
where,
The intuition here is, learn how to jump by learning how to jump twice. You can do this recursively until you hit the base case , which is just regular Flow Matching
To choose the distances that we sample from, we take a which is a power of two and sample like so:
This means the smallest time step we can take is .
The shortcut model is more expensive to train than flow matching, but far cheaper for inference.