Arbitrary-steps Image Super-resolution via Diffusion Inversion
Zongsheng Yue, Kang Liao, Chen Change Loy
2024-12-13
Summary
This paper discusses a new technique called Arbitrary-steps Image Super-resolution via Diffusion Inversion, which improves the quality of low-resolution images by using advanced methods from diffusion models.
What's the problem?
When trying to make low-resolution images clearer and more detailed (a process called super-resolution), existing methods often struggle with accuracy and efficiency. Many techniques require a lot of steps to produce a high-quality image, which can be slow and computationally expensive.
What's the solution?
The authors introduce a method that uses diffusion inversion to enhance image quality. They create a noise predictor that helps generate high-resolution images from low-resolution ones by estimating the noise needed during the image enhancement process. Their approach allows for flexible sampling steps, meaning users can choose how many steps to take (from one to five) based on their needs. Even with just one step, their method performs as well or better than many existing techniques.
Why it matters?
This research is important because it makes it easier and faster to improve image quality, which can benefit many fields like photography, video production, and medical imaging. By allowing for efficient processing of images, this technique can help create clearer visuals without needing extensive computational resources.
Abstract
This study presents a new image super-resolution (SR) technique based on diffusion inversion, aiming at harnessing the rich image priors encapsulated in large pre-trained diffusion models to improve SR performance. We design a Partial noise Prediction strategy to construct an intermediate state of the diffusion model, which serves as the starting sampling point. Central to our approach is a deep noise predictor to estimate the optimal noise maps for the forward diffusion process. Once trained, this noise predictor can be used to initialize the sampling process partially along the diffusion trajectory, generating the desirable high-resolution result. Compared to existing approaches, our method offers a flexible and efficient sampling mechanism that supports an arbitrary number of sampling steps, ranging from one to five. Even with a single sampling step, our method demonstrates superior or comparable performance to recent state-of-the-art approaches. The code and model are publicly available at https://github.com/zsyOAOA/InvSR.