Improving GANs using optimal transport

Researchers introduced OT-GAN, a new GAN variant that uses optimal transport to better match generated and real data.

The model minimizes an adversarially learned mini-batch energy distance in feature space, enabling unbiased mini-batch gradients and stable training.

OT-GAN performs well with large mini-batches and reports strong, state-of-the-art results on image generation benchmarks.