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"
!! 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 )
> CUDA 11.3
> pytorch-hed
> clip-pytorch