Plug and play acceleration using analytical solution of diffusion model and image statistics.
Mathematical Details
The Gaussian Teleportation Diffusion process leverages analytical solutions of diffusion models combined with image statistics. Here are the key mathematical components: $$$$

If you are curious about the analytical diffusion trajectory of delta mixture and general Gaussian mixture models, we provided the demo:
from gaussian_teleport import demo_delta_gmm_diffusion, demo_gaussian_mixture_diffusion
fig1 = demo_delta_gmm_diffusion(nreps=500, mus=None, sigma=1E-5)
fig1.show()
fig2 = demo_gaussian_mixture_diffusion(nreps=500, mus=None, Us=None, Lambdas=None)
fig2.show()
We host the pre-computed mean and covariance matrices for common image datasets.
- MNIST
- CIFAR10
- FFHQ64
- AFHQ64
- ImagNet64
gaussian_teleport
contains the core libraries for the project, including functions to compute analytical scores and analytical diffusion trajectories.edm
subfolder contains the simplified code to train diffusion models in EDM convention. The code is ported and adapted from yuanzhi-zhu/mini_edm repo. The dataset loading code is ported and adapted from NVlabs/edm
Tables
contains most of pre-computed tables, easy to reproduce figures or analysis in the paper.notebooks
contains the demo for using our code and reproducing our analysis.analysis
contains the main experimental analysis code and Figure plotting code.