Walter Hugo Lopez Pinaya
Walter Hugo Lopez Pinaya
Add alpha-WGAN from https://arxiv.org/abs/1908.02498
Add the GAN from ["Hierarchical Amortized GAN for 3D High Resolution Medical Image Synthesis"](https://ieeexplore.ieee.org/abstract/document/9770375) (official code available in this [link](https://github.com/batmanlab/HA-GAN))
For now, we added just the inference code for some models in the model zoo. It would be great to have the reference to train these models as a bundle.
Add tutorial on how to obtain measures of privacy and resemblance on synthetic datasets.
In the 3D examples, we had problems with the original LDM's adversarial losses. This does not mean it will not work well for the 2D scenarios.
Create the components necessary to perform the training and inference of super-resolution task based on diffusion models. As reference, we could use approaches used in Imagen (https://imagen.research.google/), Cascated Diffusion (https://cascaded-diffusion.github.io/),...
Add tutorial with approach similar to eDiff-I (https://arxiv.org/pdf/2211.01324.pdf) which relies on an ensemble of expert denoisers.
Add in [CONTRIBUTING.md](https://github.com/Project-MONAI/GenerativeModels/blob/main/CONTRIBUTING.md) instructions about how to create and run a tuorial using jupytext.
A possible new feature to support described in https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/5521 and https://github.com/huggingface/diffusers/pull/2059 . Instead upscaling atthe image level, we can perform upscaling at the latent space level. This might help to...
With the transformer and the VQ-VAE, I believe we have most of the elements to implement MUSE (https://muse-model.github.io/) from "Muse: Text-To-Image Generation via Masked Generative Transformers".