Generative Adversarial Network with Quadratic Potential

This is a minimal PyTorch code for GAN-QP without gradient vanishing which has no Lipschitz constraint like W-GANs on critic network. Also it is not trained to minimize Wasserstein divergence!

It's a totally different GAN with stable training even with high resolution data without requiring careful hyper-parameters and network architecture configurations!

Once again thanks Jianlin Su, creator of GAN-QP, for suggesting this code as PyTorch implementation of GAN-QP!

Experiments

I performed my own experiments on couple of datasets:

• CelebFaces
• LSUN Bedrooms

I trained the images on 128, 256, and 512 sized GAN-QP.

Note: Training is not yet complete! And maybe that's why results are not that good. I'll try to update these images when I'm done.

128 x 128 Resolution

CelebFaces

LSUN Bedrooms

256 x 256 Resolution

CelebFaces

512 x 512 Resolution

CelebFaces

References

• GAN-QP: A Novel GAN Framework without Gradient Vanishing and Lipschitz Constraint [arXiv]
• Original GAN-QP implementation