openai/guided-diffusion
Guided Diffusion
This is a classifier-guided diffusion framework for high-fidelity image generation. It implements a cascaded diffusion pipeline that chains a base diffusion model with a dedicated upsampler to progressively increase image resolution in stages, and uses classifier-guided diffusion sampling to steer the reverse diffusion process toward higher-quality outputs.
The framework provides tools for training diffusion models from scratch using distributed processes with gradient accumulation, as well as training classifier models that provide gradient-based guidance during sampling. It supports both unconditional image generation and classifier-guided synthesis, and includes a dedicated upsampling module for increasing image resolution through a diffusion-based pipeline.
The system is built around a noise-prediction denoising objective with a timestep-embedded U-Net backbone, modeling the diffusion process as a discrete-time Markov chain of Gaussian transitions. Documentation covers model training, classifier training, and sampling from both unconditional and guided models.
Features
- Classifier-Guided Variants - Provides a classifier-guided diffusion framework that steers sampling using a classifier for higher fidelity and controlled attributes.
- Image Diffusion Models - Generates high-fidelity images by sampling from a diffusion model, optionally guided by a classifier for improved quality.
- Cascaded Pipelines - Implements a cascaded pipeline that chains a base diffusion model with a dedicated upsampler for progressive resolution increase.
- Classifier-Guided Methods - Provides classifier-guided diffusion sampling that injects classifier gradients to steer sample quality toward higher fidelity.
- Diffusion Model Training - Trains a diffusion model on a dataset using distributed processes and adjustable settings.
- Resolution Upscaling - Increases image resolution by passing low-resolution inputs through a dedicated diffusion upsampler.
- Classifier-Guided Generation - Steers a diffusion model's sampling process with a classifier to produce higher-fidelity images.
- Timestep-Embedded Variants - Uses a U-Net backbone with sinusoidal timestep embeddings to condition the denoising process on the current noise level.
- Noise Prediction Objectives - Trains the model to predict the noise added at each timestep, enabling iterative denoising from pure noise to a clean image.
- Guided Synthesis Classifiers - Uses a classifier to steer a diffusion model's sampling process for higher-fidelity image generation.
- Diffusion-Based Upsamplers - Increases image resolution by passing low-resolution inputs through a dedicated diffusion-based upsampling pipeline.
- Diffusion-Based Upsampling - Increases image resolution by passing low-resolution inputs through a dedicated diffusion upsampler.
- Classifier Training for Guidance - Trains classifiers that provide gradient-based guidance during diffusion sampling to enhance output quality.
- Distributed Training - Trains a diffusion model from scratch on a dataset using distributed processes and adjustable settings.
- Guidance Classifier Training - Trains a classifier that steers diffusion sampling toward higher-quality outputs.
- Unconditional Generation - Generates images from a diffusion model that does not require class labels or a classifier.
- Markov Chain Monte Carlo Sampling - Models the diffusion process as a fixed-length Markov chain of Gaussian transitions, reversing it step by step during generation.
- Gradient Accumulation Strategies - Scales model training across multiple GPUs using data parallelism and gradient accumulation for large-batch convergence.
- Generative Models - Diffusion models for high-quality image generation.
- Generation - Listed in the “Generation” section of the Awesome Diffusion Models awesome list.
Star history
How this analysis was created: This summary and feature list were written by an AI model that read the project's README and public documentation pages. Each feature links to the documentation it came from; stars, license and language come straight from the GitHub API. The model does not read the source code, and the analysis is refreshed when the project is re-analysed. Learn more on our About page.
