This line of work advances GANs with three components: a Parametric Mish (PMish) activation function, a modified MMD-GAN repulsive loss integrated into a neural architecture search strategy, and adaptive rank decomposition (ARD) for compression. The approach improves stability and sample quality on datasets such as CIFAR-10, CIFAR-100, STL-10, and CelebA while reducing model size for practical deployment. 

Highlights

• Introduces PMish activation and a stabilized MMD-GAN repulsive loss. 

• Uses NAS guided by MMD criteria to discover robust generator and discriminator designs. 

• Applies ARD for parameter-efficient models suitable for edge hardware.

MMD-AdversarialNAS combines Maximum Mean Discrepancy objectives with neural architecture search to automatically design GAN architectures. It incorporates tensor decomposition to reduce parameters and storage without sacrificing generation quality.

Highlights 

• MMD-guided search for architecture selection. 

• Tensor decomposition to shrink model footprint and improve throughput. 

• Balanced trade-off between sample fidelity and efficiency

A faithful PyTorch implementation of Progressive Growing of GANs. Training starts at low spatial resolution and grows layers progressively to improve convergence and image quality. 

Highlights 

• Progressive resolution schedule for stable training. 

• Modular codebase for experimenting with blocks and losses. 

We study how GAN-based synthetic video generation can improve human action recognition by augmenting scarce or imbalanced training data. The goal is to boost recognition accuracy and robustness with targeted synthetic samples. 

Highlights 

• Task-driven video synthesis for data augmentation. 

• Measurable gains on action recognition benchmarks. 

• Analysis of when and where synthetic augmentation helps most.