Figure 1.

The proposed MResTNet architecture uses a ViT encoder and decoder, along with a convolutional multi-resolution decoder and two networks (scaler and fuser) that estimate a method hyperparameter (scaler) and combine the two decoders (fuser).