sagiebenaim/DistanceGAN

Combines DiscoGAN and CycleGAN architectures to enable one-sided unpaired image translation using a distance-based loss function, eliminating the need for cycle consistency constraints. Supports flexible training modes including unidirectional mapping, optional reconstruction loss, and self-distance variants across diverse datasets (edges, objects, faces, animals). Integrates with Visdom for real-time loss visualization and includes both discriminator-based and CycleGAN-based implementations for various domain mapping tasks.

195 stars. No commits in the last 6 months.