Image reconstruction for deep learning is typically performed by training a neural network to map from noisy to clean images by minimizing a loss or a sum of losses. For a sum of losses, a hyperparameter needs to be used to weight the contribution of the losses. At deployment, these models will only produce a single clean image based on the loss it was trained on - this is suboptimal because different losses produce different visual reconstructions. How can we eliminate this dependence on the loss we select at training time (i.e. be agnostic to the loss function) and also get multiple reconstructions for a single noisy input at test-time?

HyperRecon addresses this by using a hypernetwork to produce multiple reconstructions corresponding to different loss functions. This means that at test-time, a user can sweep over a continuous range of hyperparameter values and get a dense set of recontructions at test time!