Hugging Face Diffusers

Hugging Face's open-source diffusion-model toolkit, with ready-made schedulers, networks, and generation pipelines.

Diffusers is Hugging Face's open-source library for diffusion models, built on PyTorch. It splits a diffusion model into three pieces: a scheduler (which decides the noising and denoising steps, such as DDPM or DDIM), a model (such as a U-Net or a diffusion transformer), and a pipeline that strings the two together for inference — and it can load many publicly available image and video generation models directly. In robotics, people often only borrow its scheduler: Diffusion Policy swaps “denoising an image” for “denoising a sequence of actions,” and the sampling steps are the same as in image diffusion, so there's no need to rewrite them from scratch. It's also commonly used to load a pretrained video model and fine-tune it when building a world model or doing video generation.

ExampleThe official Diffusion Policy code calls Diffusers' DDPMScheduler and DDIMScheduler directly to train and sample actions.

Also called

diffusers

Related

Diffusion Model · Diffusion Policy · Denoising Diffusion Probabilistic Model · Denoising Diffusion Implicit Model · Noise Schedule · Hugging Face

Sources

Diffusers 官方文档 (Chinese)
Diffusion Policy GitHub