openai/improved-diffusion
Improved Diffusion
This project is a diffusion model framework for training and sampling from denoising probabilistic models to generate images from noise. It functions as a generative image model that creates visual content by iteratively refining random noise into coherent images.
The system includes a distributed GPU trainer designed to scale complex neural network architectures across multiple graphics processing units. It also provides an image dataset preprocessor to prepare, scale, and standardize raw image collections for training.
The framework covers model training and image generation, utilizing noise scheduling and iterative denoising sampling. It incorporates memory management through microbatching and supports image dataset loading and preparation.
Features
- Image Diffusion Models - Provides a system for generative image modeling via diffusion processes that refine random noise into coherent visuals.
- Denoising Probabilistic Training - Implements probabilistic training for denoising models to generate high-quality images through learned noise removal.
- Distributed GPU Training - Provides infrastructure to synchronize model weights across multiple GPUs for accelerating large-scale neural network training.
- Diffusion Model Training - Supports the training of denoising probabilistic models using configurable architectures, noise schedules, and microbatching.
- Generative Image Models - Implements a generative image model that creates visual content by iteratively refining random noise.
- Noise-to-Image Generation - Converts random noise into coherent visual content using trained diffusion checkpoints and custom respacing methods.
- Image Set Sampling - Generates batches of synthetic images from trained checkpoints using specific sampling strides and denoising algorithms.
- Iterative Denoising Pipelines - Employs iterative denoising pipelines to generate coherent images by repeatedly removing noise from a random state.
- Diffusion Model Frameworks - Functions as a research framework for training and sampling from denoising diffusion probabilistic models.
- Training Noise Schedules - Implements training noise schedules using a cosine curve to preserve image detail during the diffusion process.
- Sampling Timestep Respacing - Implements a timestep respacing mechanism to reduce the number of iterations required for image sampling during inference.
- Training Memory Management - Utilizes microbatching for training memory management to fit high-resolution image data into limited GPU VRAM.
- Variational Lower Bound Optimizers - Optimizes the model by minimizing a variational lower bound as a proxy for data log-likelihood.
- Generation - Listed in the “Generation” section of the Awesome Diffusion Models awesome list.
