GLEAN: Generative Latent Bank for Image Super-Resolution and Beyond.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing image super-resolution methods often rely on adversarial loss for realistic outputs.
• Generative Adversarial Networks (GANs) possess rich prior information valuable for image restoration tasks.

Purpose of the Study:

• To introduce a novel method, Generative Latent Bank (GLEAN), for improving image super-resolution by leveraging pre-trained GANs.
• To develop an efficient image restoration technique that avoids computationally expensive, image-specific optimization.

Main Methods:

• Utilized pre-trained Generative Adversarial Networks (GANs) like StyleGAN and BigGAN as a latent bank.
• Implemented an encoder-bank-decoder architecture with multi-resolution skip connections for restoration.
• Introduced LightGLEAN, a parameter-efficient version of GLEAN.

Main Results:

• GLEAN significantly improves image super-resolution performance by utilizing GAN priors.
• The method requires only a single forward pass, making it highly efficient compared to GAN inversion techniques.
• LightGLEAN achieves comparable image quality with substantially fewer parameters and computations.
• Extended applications to image colorization and blind image restoration demonstrated favorable performance.

Conclusions:

• Pre-trained GANs serve as effective latent banks for enhancing image restoration tasks.
• GLEAN offers an efficient and versatile solution for super-resolution and other image restoration problems.
• LightGLEAN provides a compelling trade-off between performance and computational cost.