摘要
预训练扩散先验图像复原依赖预训练的扩散模型,无须微调即可处理各种经典图像复原任务。然而,目前的预训练扩散先验图像复原方法在处理高分辨率图像时效率低下,并且存在分布外问题(out of distribution,OOD)。针对以上问题,提出了一种基于预训练扩散模型的两阶段高分辨率图像复原方法,命名为由粗到细(coarse-to-fine,C2F)的方法。首先在预训练模型固定尺寸的coarse阶段得到粗糙的复原结果以保证输出一致性。然后在原尺寸的fine阶段上以coarse阶段结果为起点,使用更短的扩散过程来大幅度提升复原速度与获取一致性结果。在人脸与自然环境等多种场景下,以修复、上色、去模糊三种经典复原任务为目标,两阶段方法在任何尺寸下皆可获得最高水平的输出结果。对于1024尺寸的图像复原,采样次数需求仅需要同类方法的22%,速度达到了同类方法的4.5倍,避免了OOD问题,并且在PSNR与FID指标上达到最高水平。实验表明,所提方法对高分辨率图像的复原速度远高于其他方法,并且避免了OOD问题,具有良好的复原效果。
Pretrained diffusion priors for image restoration rely on pretrained diffusion models to handle various classic image restoration tasks without fine-tuning.However,current methods are inefficient for high-resolution images and suffer from OOD issues.To address these problems,this paper proposed a two-stage high-resolution image restoration method based on pretrained diffusion models,named the C2F method.In the first stage,the method obtained a coarse restoration result at the fixed size of the pretrained model,ensuring output consistency.In the second stage,the method restored the original resolution using the coarse result as a starting point.A shorter diffusion process significantly enhanced restoration speed and ensured consistency.In experiments with facial restoration,natural environments,and three classic tasks—repair,colorization,and deblurring—the two-stage method achieves top-level results at any resolution.For 1024 resolution,the proposed method requires only 22%of the sampling steps of similar methods,achieving 4.5 times faster speed while avoiding OOD issues.It also reaches the highest levels of PSNR and FID scores.This paper demonstrates that the proposed method restores high-resolution images much faster than other methods,avoids OOD issues,and produces high-quality results.
作者
谢源远
周非
周志远
张宇曈
Xie Yuanyuan;Zhou Fei;Zhou Zhiyuan;Zhang Yutong(School of Communication&Information Engineering,Chongqing University of Posts&Telecommunications,Chongqing 400065,China)
出处
《计算机应用研究》
北大核心
2025年第8期2545-2551,共7页
Application Research of Computers
基金
国家自然科学基金资助项目(62271096)。
关键词
图像复原
扩散模型
预训练模型
image restoration
diffusion model
pretrained model
