diffusion_distillation

This is an amateur implementation of "One-step Diffusion with Distribution Matching Distillation" https://arxiv.org/abs/2311.18828.

There are multiple approaches made by: Grishina Ekaterina, Ulyana Klyuchnikova, Maxim Bekoev

You can run the distillation process using main.py. We have published our notebooks in the corresponding folder to demonstrate how to use it. You can see our results and hyperparameter settings in our presentation.

MNIST

We have taken pretrained model from https://github.com/TeaPearce/Conditional_Diffusion_MNIST.

• The code for distillation is in notebooks/celeba_distillation.ipynb, just switch parameter 'model_name' to 'ddpm_conditional_mnist'

Our one-step generation result:

CIFAR 10

For that particular dataset model google/ddpm-cifar10-32 was chosen.

• The code for generating image pairs is located in notebooks/ddpm_cifar10_pair_generation.ipynb
• The code for training a model with distillation is located in notebooks/ddpm_distillation_cifar10.ipynb

CelebA

We trained our custom DDPM on 32x32 CelebA images.

• The code for training is in notebooks/training_celeba.ipynb
• The code for distillation is in notebooks/celeba_distillation.ipynb