Generation with Unified Diffusion

Diffusion utilities for consistent basis transforms, schedules, and sampling. Built around whitened and unwhitened spaces with data, score and epsilon conversions.

This code implements the methods described in the paper: GUD: Generation with Unified Diffusion

Tutorial Dependencies

First, install the package with pip install gud (depending on your setup, probably in a new virtual environment, or instead with uv).

Alternatively, it can be installed directly from the github repository via pip install git+https://github.com/mathisgerdes/unified-diffusion.git.

The tutorial notebooks require additional packages beyond the core dependencies. Install them with:

pip install optax matplotlib torchvision

Or individually:

• optax - Gradient processing and optimization library for JAX
• matplotlib - Plotting library
• torchvision - Computer vision datasets and transforms
• ipython - Interactive Python shell (for notebook display utilities)