Generating synthesized digital images utilizing a multi-resolution generator neural network

A system comprising:

• one or more memory devices comprising a multi-resolution generator neural network comprising:
• a base encoder that generates a feature set from a semantic label map;
• a bottom-up encoder that extracts multi-resolution features and generates a latent code from the feature set;
• a top-down encoder that determines a two-dimensional spatial feature tensor for the multi-resolution features by aggregating the multi-resolution features with lateral connections;
• a decoder that generates a digital image from the two-dimensional spatial feature tensor and the latent code;
• and
• one or more processors configured to cause the system to generate, utilizing the multi-resolution generator neural network, a synthesized digital image with a semantic layout of the semantic label map.

A computer-implemented method comprising:

• extracting, by at least one processor, a plurality of feature sets at different resolutions from a semantic label map comprising a plurality of labeled objects via a plurality of sequential downsampling neural network layers;
• determining, by the at least one processor, a two-dimensional spatial feature tensor for the plurality of feature sets by aggregating the plurality of feature sets via a plurality of sequential upsampling neural network layers and feature combination layers;
• determining, by the at least one processor, a latent code from a reduced resolution feature set of the plurality of feature sets via a plurality of fully connected neural network layers;
• transforming, by the at least one processor, the latent code corresponding to the reduced resolution feature set based on a feature distribution associated with learned parameters of a generator neural network;
• and
• generating, by the at least one processor utilizing the generator neural network, a synthesized digital image based on the two-dimensional spatial feature tensor and the transformed latent code.