xrenaa/DisCo

Leverages contrastive learning to discover disentangled directions in the latent spaces of pretrained generative models (StyleGAN2, SNGAN, VAE, Flow) without supervision or paired data. The method operates model-agnostically across different architectures and evaluates disentanglement using standard metrics (MIG, DCI) on benchmark datasets like Shapes3D, Cars3D, and MPI3D. Implemented in PyTorch with support for both training custom direction vectors and evaluating representation quality across multiple generative model families.

136 stars. No commits in the last 6 months.