| dc.description.abstract | In the image deblurring task, most work uses pixel-level loss to reduce the distortion between the restored result and ground truth. However, these kind of methods overlook the human perception of image quality, leading to insufficient details in the restored results. Recently, diffusion models, which have achieved impressive success in image synthesis, have also been applied to the image deblurring task. Although the existing diffusion-based image deblurring methods can address the perception issue, they require more computational consumption or processing time during inference. In this paper, we propose a method that employs a pre-trained latent diffusion model to enchance the existing image deblurring model. This approach only utilizes latent diffusion model to improve perceptual quality of the result of the original image deblurring model (e.g., FFTformer) during training. And the pre-trained latent diffusion model will be adjusted to make it suitable for aiding the image deblurring network by new prompt tuning methods, as proposed in this paper. Compared with fine-tuning, the proposed method requires fewer training parameters and maintains the prior knowledge obtained during pre-training of the latent diffusion model. In experiments, our proposed method shows a 0.64 dB decrease in PSNR. However, it improves perceptual metrics, with LPIPS decreasing by 0.012, NIQE decreasing by 0.51, FID decreasing by 0.63, CLIP-IQA increasing by 0.002, and CLIP-IQA^+ increasing by 0.01. | en_US |