Car-Worning Flow Reparameterization aligns with the goal of better flow simulation
Active conditional models aim to learn the conditional distribution of data from samples containing pairs of data. In this case, perturbation methods and flow-based methods have achieved powerful results. These methods use a learned (Flow) model to transport normal Gaussian noise that is agnostic to the conditional data distribution. The model is therefore required to study both the Transport injection and the conditional coupling. To ease the demand on the model, we suggest that Reparameteriction to know the matching variable (car flow) – A simple change, learned that the useful conditions are the source, the target, or both of the distribution. By moving this distribution, the autopilot reduces the probability that the model has to learn, resulting in faster training in practice. In the low synthetic manual, we visualize the effects of the flow of the car. With high quality natural image data (imagenet-256), sit sit-XL / 2 with car-flow reduces fid from 2.07 to 1.68, while introducing 0.6 additional parameters.
- ** Work Done while at Apple
