lucidrains/denoising-diffusion-pytorch

What is it?

Key features

• 01Implements the DDPM training objective and sampling procedure in PyTorch.
• 02Provides a configurable U-Net backbone with optional flash attention.
• 03A Trainer class handles folder-based image datasets, checkpoints and sample logging.
• 04Supports multi-GPU training through Hugging Face Accelerate and mixed precision.
• 05Includes DDIM-style sampling with fewer steps and FID evaluation during training.
• 06Offers a 1D U-Net and diffusion variant for sequence data.

Use cases

• Training a small image generator on a custom folder of images.
• Studying or modifying the core diffusion training and sampling code.
• Prototyping diffusion models for one-dimensional sequence data.
• Running multi-GPU diffusion training experiments with Accelerate.

Quick start

1. 1

   Install the package from PyPI

   pip install denoising_diffusion_pytorch
2. 2

   Build a model and diffusion object, then train on a folder of images with the Trainer class

3. 3

   Start training

   trainer.train()

Things to know

• !Training on realistic image sizes requires a GPU and a lot of time (the example uses 700000 steps).

Topics

FAQ

What is denoising-diffusion-pytorch?

A PyTorch implementation of Denoising Diffusion Probabilistic Models for training image (and 1D sequence) generative models. This project is a PyTorch implementation of Denoising Diffusion Probabilistic Models, a generative modeling approach that learns to reverse a gradual noising process. It provides a U-Net backbone along with a GaussianDiffusion class that handles the forward noising, the training loss, and sampling from the trained model.

Is denoising-diffusion-pytorch free to use? What license does it have?

denoising-diffusion-pytorch is open source under the MIT license. Check the license terms before using it in commercial products.

Is denoising-diffusion-pytorch still maintained?

Actively maintained. The last commit to denoising-diffusion-pytorch was on Sep 29, 2026.

What can I use denoising-diffusion-pytorch for?

Training a small image generator on a custom folder of images. Studying or modifying the core diffusion training and sampling code. Prototyping diffusion models for one-dimensional sequence data. Running multi-GPU diffusion training experiments with Accelerate.

How do I install denoising-diffusion-pytorch?

Install the package from PyPI: pip install denoising_diffusion_pytorch Build a model and diffusion object, then train on a folder of images with the Trainer class Start training: trainer.train()

Where can I get denoising-diffusion-pytorch?

The source code and installation instructions are on GitHub: https://github.com/lucidrains/denoising-diffusion-pytorch