/Product-of-Experts-GAN

PyTorch unoffical implementation of "PoE-GAN : Multimodal Conditional Image Synthesis with Product-of-Experts GANs"

Primary LanguagePython

Product-of-Experts-GAN

This Repository is unoffical implementation of PoE-GAN : Multimodal Conditional Image Synthesis with Product-of-Experts GANs in PyTorch

This implementation is mostly relied on "rosinality's stylegan2-pytorch"

image

Notice

!! Now really unstable and I will reproduce and improve continuously please wait !!
I have tried to follow offical paper experiment settings, but some parts are not perfectyl implemented. ( TBD! )

  • It works when enough batch size
  • Train only non-saturating gan loss first, and from intermediate steps using modality loss ( contrastive loss and KL loss )

Requirements

> CUDA 11.3
> pytorch-hed
> clip-pytorch

Sample

image