GAN Compression: Efficient Architectures for Interactive Conditional GANs.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Conditional Generative Adversarial Networks (cGANs) excel at controllable image synthesis but are computationally intensive.
• High computational costs hinder the interactive deployment of advanced cGANs like GauGAN.

Purpose of the Study:

• To propose a general-purpose framework for compressing cGAN generators.
• To reduce inference time and model size of cGANs for practical applications.

Main Methods:

• A novel compression framework addressing GAN training instability via knowledge transfer and unified learning.
• Efficient architecture search for cGANs, accelerated by decoupled training and weight sharing.
• Leveraging neural architecture search to find efficient generator designs.

Main Results:

• Significant computational reduction for various cGANs: CycleGAN (21×), Pix2pix (12×), MUNIT (29×), and GauGAN (9×).
• Maintained high image quality post-compression.
• Demonstrated effectiveness across diverse GAN architectures and training settings.

Conclusions:

• The proposed framework enables practical, interactive image synthesis with cGANs.
• Offers a viable solution for deploying compute-intensive generative models.
• Paves the way for real-time applications of controllable image generation.