🌐 Diffusion World 🌐

Welcome to Diffusion World !

This repository is a collection of diffusion models, implemented in PyTorch and PyTorch Lightning.

🚀 Features

• Models You Can Explore:

  • DDPM (denoising diffusion probabilistic model)
  • D3PM (structured denoising diffusion model in discrete state-spaces)

• Datasets Integrated:

  • MNIST
  • CIFAR-10

🌟 Examples

• To train DDPM with MNIST dataset

python run.py --model=ddpm --dataset_name=mnist

• To train D3PM with CIFAR-10 dataset

python run.py --model=d3pm --discrete=True --dataset_name=cifar10

After training, retrieve your models and results from the ./results directory.

• For additional command line arguments, use:

python run.py --help

📊 Training Results

• DDPM with cifar10 dataset

python run.py --model_name=ddpm --dataset_name=cifar10

Click to collapse

epoch 1

epoch 5

epoch 10

• D3PM with mnist dataset

python run.py --model_name=d3pm --discrete=True --dataset_name=mnist --scheduler_mode=jsd

click to collapse

epoch 1

epoch 5

epoch 10

TODO

Click to expand

• Add a diffusion model of Noise Conditional Score Network(https://arxiv.org/abs/1907.05600)

• Add DDIM sampling method

References

Thanks to the following repositories and papers for the inspiration and guidance:

• Denoising Diffusion Probabilistic Models
• Structured Denoising Diffusion Models in Discrete State-Spaces
• lucidrains/denoising-diffusion-pytorch
• google-research/d3pm
• cloneofsimo/d3pm