Efficient Diffusion Model for Image Restoration by Residual Shifting.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Diffusion models excel in image restoration (IR) but suffer from slow inference due to numerous sampling steps.
• Existing acceleration techniques often compromise restoration quality, leading to blurry results.

Purpose of the Study:

• To develop a novel and efficient diffusion model for image restoration that significantly reduces sampling steps.
• To achieve high-quality image restoration with accelerated inference speeds, avoiding performance degradation.

Main Methods:

• Proposes a new diffusion model for IR that minimizes required diffusion steps without post-acceleration.
• Establishes a Markov chain for efficient transitions between low-quality and high-quality images by shifting residuals.
• Devises a tailored noise schedule to control shifting speed and noise strength during diffusion.

Main Results:

• The proposed method achieves superior or comparable performance to state-of-the-art methods on four IR tasks.
• Demonstrates effectiveness with as few as four sampling steps, drastically improving inference speed.
• Successfully applied to image super-resolution, inpainting, blind face restoration, and deblurring.

Conclusions:

• The novel diffusion model offers a significant advancement in efficient and high-performance image restoration.
• The method overcomes the speed-quality trade-off inherent in previous accelerated diffusion techniques.
• Enables rapid, high-fidelity image restoration across diverse classical IR applications.