konpatp/diffae

Combines diffusion models with autoencoder architecture to learn interpretable latent representations that support both reconstruction and generative sampling. The framework uses a two-stage approach: first training a diffusion model as an encoder, then optionally training a latent diffusion probabilistic model (DPM) in the learned latent space for generation. Supports semantic image manipulation via learned classifiers, latent interpolation, and unconditional synthesis on face and object datasets (FFHQ, CelebA, LSUN) with pre-trained checkpoints provided.

959 stars. No commits in the last 6 months.