denoising-diffusion-pytorch

Why we included this project

This is the reference implementation most practitioners start from when they want to train a diffusion model on their own images rather than call a hosted API. It turns the Denoising Diffusion Probabilistic Model paper into clean, composable PyTorch building blocks: a UNet backbone and a GaussianDiffusion wrapper that handle forward noising, the training loss, and the sampling loop. You can feed in a folder of images or raw tensors, and the included Trainer class takes care of the rest, while the lower-level API lets you swap in your own architecture for more control. Beyond the core DDPM, the repo tracks newer research ideas like cold diffusion, so it also works as a practical tour of how the technique has evolved. For teams learning generative modeling or needing a dependable baseline to extend, this code is easier to digest than the original TensorFlow release and has influenced many derivatives across the ecosystem.