Compressive diffusion model.
This is a repository for the compressive diffusion model (CDM) for image compression. The models are based on the Variational Diffusion Model (VDM) and the Latent Diffusion Model (LDM). The models are trained on the MNIST and CIFAR10 datasets. An additional model was made using the VDM alongside BB-ANS for the MNIST dataset (this model is lossless). The models are trained using the PyTorch framework.
VDM paper: https://arxiv.org/abs/2107.00630
LDM paper: https://arxiv.org/abs/2112.10752
- [✓] Create the v0.1 of the model for the MNIST dataset
- [✓] Create the autoencoder model for v0.1
- [✓] Create the score model for v0.1
- [✓] Train it on the dataset
- [✓] Test and record findings
- [✓] Create the v0.2 of the model for the CIFAR10 dataset
- [✓] Create the advanced autoencoder model for v0.2
- [✓] Create the advanced score model for v0.2
- [✓] Train it on the dataset
- [✓] Test and record findings
- Check the notebooks in the model folder for examples on how to use the model. You might have to do some changes to the code to make it work for you (but should work out of the box for the MNIST/CIFAR dataset and most systems).
- Checkpoint files are not stored due to the minimal size of the datasets. Of course these models (BB-ANS and CIFAR10) are scalable to larger datasets, but the training time will increase.
BB-ANS is a lossless compression algorithm that is used in the VDM model. This is possible do with a diffusion model as it can act as a latent model. The BB-ANS uses this latent model to compress the data. More details can be found in the report and the sources listed there.