yang-song/score_sde_pytorch

Implements a unified SDE framework for score-based generative modeling with support for multiple architectures (NCSN++, DDPM++, NCSN, NCSNv2, DDPM) and training/evaluation pipelines for image generation tasks. Enables exact likelihood computation, conditional generation (inpainting, colorization, class-conditional), and latent code manipulation through reversible stochastic processes. Integrates with Hugging Face Diffusers library for easy inference and features modular, extensible design for custom SDEs, predictors, and correctors.

2,089 stars. No commits in the last 6 months.