Prototyping repo for generative models to be integrated into MONAI core.
- Network architectures: Diffusion Model, Autoencoder-KL, VQ-VAE, (Multi-scale) Patch-GAN discriminator.
- Diffusion Model Schedulers: DDPM, DDIM, and PNDM.
- Losses: Adversarial losses, Spectral losses, and Perceptual losses (for 2D and 3D data using LPIPS, RadImageNet, and 3DMedicalNet pre-trained models).
- Metrics: Multi-Scale Structural Similarity Index Measure (MS-SSIM) and Maximum Mean Discrepancy (MMD).
- Diffusion Models and Latent Diffusion Models Inferers classes (compatible with MONAI style) containing methods to train, sample synthetic images, and obtain the likelihood of inputted data.
- MONAI-compatible trainer engine (based on Ignite) to train models with reconstruction and adversarial components.
- Tutorials including:
- How to train VQ-VAEs, VQ-GANs, AutoencoderKLs, Diffusion Models and Latent Diffusion Models on 2D and 3D data.
- Train diffusion model to perform conditional image generation with classifier-free guidance.
- Comparison of different diffusion model schedulers.
- Diffusion models with different parameterisation (e.g. v prediction and epsilon parameterisation).
- Inpainting with diffusion model (using Repaint method)
- Super-resolution with Latent Diffusion Models (using Noise Conditioning Augmentation)
Our short-term goals are available in the Milestones section of the repository and this document.
In the longer term, we aim to integrate the generative models into the MONAI core library (supporting tasks such as, image synthesis, anomaly detection, MRI reconstruction, domain transfer)
To install MONAI Generative Models, it is recommended to clone the codebase directly:
git clone https://github.com/Project-MONAI/GenerativeModels.git
This command will create a GenerativeModels/ folder in your current directory. You can install it by running:
cd GenerativeModels/
python setup.py install
For guidance on making a contribution to MONAI, see the contributing guidelines.
