When training flow matching, it seems like the model predicts the noise all at once without using time steps, but during inference, it uses 10 time steps. Is this approach used because training with a harder problem and performing inference with an easier noise prediction leads to better performance?
When training flow matching, it seems like the model predicts the noise all at once without using time steps, but during inference, it uses 10 time steps. Is this approach used because training with a harder problem and performing inference with an easier noise prediction leads to better performance?