NVIDIA FastGen Overview
FastGen is a PyTorch library for training and distilling diffusion models into fast generative models. Use it to fine-tune a teacher, train a student that generates in fewer steps, and run inference and evaluation.
Start here
- Get started: install FastGen and run the CIFAR-10 training and inference example.
- Run inference: generate outputs from a teacher or student model.
- Prepare your dataset: choose a loader and supply your training data.
- Troubleshoot a run: resolve common configuration, memory, and checkpoint issues.
What you can do
- Distill diffusion models across modalities such as image, video, and audio into students that use fewer generation steps.
- Fine-tune a supported model on your data, including conditional and causal generation tasks.
- Run inference and evaluate teacher and student models.
Core concepts
A configuration combines a method, a network, a dataset, and training settings. Experiment configs are runnable starting points; model and data combinations have different prerequisites. Check the selected recipe before starting a large run.
Learn more
Resources collects external models, research projects, and integrations. See the GitHub repository, issue tracker, and release history for source code, support, and released revisions.