Previously my understanding about one-to-many RNNs was the input at a single timestep is "cloned" multiple times to make multiple timesteps, and feed in
