Differentiable weighted reward. Returns a callable `fn(images, prompts) -> tensor[B]` that computes a weighted sum of the configured differen

Predict the sample from the previous timestep by reversing the SDE. This function propagates the flow process from the learned model outputs

EMA update from detached tensors (e.g. another wrapper's ``ema_parameters``).

Submits images to DeQA and computes a reward.