huguesva/semantic-transport-generation

Generate in Reconstruction Space, Match in Semantic Space: Transport Geometry for One-Step Generation

This repository contains the model presented in Generate in Reconstruction Space, Match in Semantic Space: Transport Geometry for One-Step Generation.

The method introduces one-step image generation where a generator maps noise to an image in a single forward pass inside a pretrained autoencoder's latent space. A frozen self-supervised featurizer provides semantic features, and the distributional matching loss used to train the generator is defined in this semantic feature space. At inference, only the generator and decoder are needed: no featurizer and no iterative sampling.

Code is available at: https://github.com/Genentech/semantic-transport-generation.