junyanz/BicycleGAN

Addresses the one-to-many translation problem by injecting random latent vectors to generate diverse plausible outputs from a single input, trained with paired data using a combination of conditional GAN and VAE-inspired latent code constraints. Implements encoder-generator-discriminator architecture in PyTorch with latent code reconstruction loss to ensure output diversity while maintaining realism, and includes pre-trained models for edge-to-photo, night-to-day, and semantic segmentation tasks.

1,515 stars. No commits in the last 6 months.