Zero-shot realistic image deblurring with consistency model

At present, diffusion-based image deblurring methods rely on paired blurry-clear datasets for training, and the types of blur causes in image synthesis cannot yet be determined with sufficient precision to model real-world scene blur datasets. Furthermore, diffusion models are computationally expensive and may generate distort images during deblurring. To solve the above problems, this paper proposes image deblurring with consistency distillation models (ID-CDM) to restore blurred images in real scenes without training for image deblurring tasks. Specifically, this paper proposes a diffusion coefficient-dominated score model (DCSM), which ignores the impact of the drift term on the forward stochastic differential equation (SDE) adding Gaussian noise, thus simplifying the deblurring network architecture and reducing the computational complexity. The score model is integrated as a distillation model into the consistency model, aiming to enhance the quality of image deblurring while significantly reducing image sampling time. Besides, this paper employs the method of invertible linear transformation to highly compress the blurred image into the latent space for zero-shot processing, and iterative sampling is utilized to improve image accuracy and reduce additional artifact details generated by deblurring images. Finally, to verify the effectiveness of the proposed method, comparative analyses were conducted with classical deblurring methods on the synthetic blurring dataset HIDE and the real scene blurring dataset RealBlur. Analysis results indicate that ID-CDM is a highly competitive deblurring scheme.