Improving GAN training with optimal transport methods

AI Impact Summary

Adopting optimal transport techniques is planned to improve GAN training stability and sample quality by using OT-based loss formulations (e.g., Wasserstein/Sinkhorn). This can reduce mode collapse and enable faster convergence, but OT computations may add per-iteration overhead and require integrating OT libraries or custom losses into the training loop. Teams should prototype OT-based objectives in a subset of GANs, validate improvements with metrics like FID/IS, and plan for potential changes to hardware utilization and training budgets.