bryandlee/malnyun_faces

Combines StyleGAN2 and U-GAT-IT with transfer learning and adaptive augmentation techniques (FreezeD, ADA) to train high-resolution face generators on limited cartoon data—achieving 256×256 generation in ~10 hours on a single GPU with just 500 images. Implements latent space exploration via GANSpace and closed-form factorization, with interactive Gradio interfaces for semantic attribute discovery. Demonstrates cross-domain synthesis by chaining pre-trained FFHQ models with cartoon-specific translation modules while preserving facial characteristics.

904 stars. No commits in the last 6 months.