Online Multi-Granularity Distillation for GAN Compression

TL;DR

This paper introduces an online multi-granularity distillation method to compress GANs significantly, enabling high-quality image generation with much lower computational and memory requirements for deployment on resource-limited devices.

Contribution

It presents the first single-stage online distillation approach for GAN compression, utilizing a progressively promoted teacher generator and multi-granularity concepts to enhance visual fidelity.

Findings

Achieves 40x MACs and 82.5x parameters reduction on Pix2Pix and CycleGAN.

Maintains image quality despite significant model compression.

Demonstrates feasibility for real-time image translation on resource-constrained devices.

Abstract

Generative Adversarial Networks (GANs) have witnessed prevailing success in yielding outstanding images, however, they are burdensome to deploy on resource-constrained devices due to ponderous computational costs and hulking memory usage. Although recent efforts on compressing GANs have acquired remarkable results, they still exist potential model redundancies and can be further compressed. To solve this issue, we propose a novel online multi-granularity distillation (OMGD) scheme to obtain lightweight GANs, which contributes to generating high-fidelity images with low computational demands. We offer the first attempt to popularize single-stage online distillation for GAN-oriented compression, where the progressively promoted teacher generator helps to refine the discriminator-free based student generator. Complementary teacher generators and network layers provide comprehensive and multi-granularity concepts to enhance visual fidelity from diverse dimensions. Experimental results on four benchmark datasets demonstrate that OMGD successes to compress 40x MACs and 82.5X parameters on Pix2Pix and CycleGAN, without loss of image quality. It reveals that OMGD provides a feasible solution for the deployment of real-time image translation on resource-constrained devices. Our code and models are made public at: https://github.com/bytedance/OMGD.