As the world’s best-preserved and largest repository of cave art, the Dunhuang murals hold immense research value in aesthetics, history, and religion. However, due to the harsh geographical environment, the extensive time span of mural creation, and the impact of human activities, the murals have sustained various forms of damage. The diffusion model, a new generation of generative models, has achieved remarkable success in image inpainting. Building on the denoising diffusion probability model, this paper enhances the forward denoising method, optimizes the training process using the exponential moving average method, and proposes a Dunhuang mural inpainting technique based on a classifier-guided conditional diffusion model. This approach not only incorporates the image classification and generation principles of the conditional diffusion model but also integrates the classifier’s control over the generation results. During the restoration phase, the pre-trained diffusion model is employed as a generative prior, with the backward diffusion iteration adjusted specifically for sampling the damaged areas of the murals using the given image information. Experimental results demonstrate that our method can generate semantically rich and diverse output images for various restoration scenarios. This innovation offers a novel research direction in the field of mural image inpainting.