Guided Diffusion

Guided Diffusion is OpenAI’s official open-source codebase for the landmark “Diffusion Models Beat GANs on Image Synthesis” research, the work that introduced classifier guidance and showed diffusion models could outperform GANs at image generation. It has around 7,400 stars on GitHub and remains a reference implementation for anyone studying or building diffusion-based image models.

Overview

The repository builds on OpenAI’s improved-diffusion codebase and adds classifier conditioning plus architecture improvements (often called ADM, the ablated diffusion model). OpenAI released pretrained checkpoints for the main models from the paper, covering ImageNet generation at 64×64, 128×128, 256×256 and 512×512 resolutions, dedicated upsampler models that convert low-resolution samples into high-resolution ones, and unconditional LSUN models for bedrooms, cats and horses. Sampling scripts are provided for classifier-guided generation, plain image sampling and super-resolution, and the codebase also includes the tooling to train your own classifiers for guidance. The project was last updated in 2024 and is released under the MIT license.

Key Features

• Official implementation of classifier-guided diffusion (ADM) from OpenAI
• Pretrained checkpoints for ImageNet models from 64×64 up to 512×512
• Upsampler models that turn 64×64 samples into 256×256 and 128×128 into 512×512
• Ready-made sampling scripts for guided, unguided and super-resolution generation
• DDIM sampling support for much faster generation with fewer timesteps
• Training scripts for building the classifiers used to steer generation
• Published FID, precision and recall benchmarks for every released model

💳 Pricing

Guided Diffusion is completely free and open source under the MIT license. The code and all pretrained model checkpoints can be downloaded from GitHub at no cost; you only need your own GPU compute to run them.

Who Is It For

Guided Diffusion is for researchers, students and engineers who want to understand how modern image generation works under the hood, reproduce a foundational result, or use proven diffusion code and checkpoints as a starting point for their own generative image projects.