Single image super-resolution with denoising diffusion GANS.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Single Image Super-Resolution (SISR) reconstructs high-resolution (HR) images from low-resolution (LR) inputs, an ill-posed problem.
• Generative models like GANs, VAEs, and Flows improve SISR but face challenges in training stability, sample quality, and computational cost.
• Denoising diffusion probabilistic models offer high sample quality and diversity but suffer from slow sampling speeds, limiting real-world application.

Purpose of the Study:

• To investigate the cause of slow sampling in diffusion model-based SISR.
• To propose a novel SISR method that achieves fast, diverse, and high-quality image generation.
• To improve the practical applicability of diffusion models for real-world SISR tasks.

Main Methods:

• Proposed a new method, Single Image Super-Resolution with Denoising Diffusion GANs (SRDDGAN).
• Combined denoising diffusion models with GANs for conditional image generation.
• Utilized a multimodal conditional GAN to model each denoising step, enabling large-step denoising.

Main Results:

• SRDDGAN achieves superior performance in Peak Signal-to-Noise Ratio (PSNR) and perceptual quality compared to existing diffusion models.
• The model explores the diversity of likely HR spatial domains through an added latent variable Z.
• SRDDGAN demonstrates a significant speed improvement, inferring nearly 11 times faster than the SR3 model.

Conclusions:

• The Gaussian assumption in traditional diffusion models limits step size, contributing to slow sampling.
• SRDDGAN overcomes the limitations of previous methods by enabling large-step denoising, ensuring sample diversity, and maintaining training stability.
• SRDDGAN offers a practical and efficient solution for real-world SISR applications due to its speed and quality.