Facial image inpainting using condition guided diffusion models

Although modern image inpainting has gradually become more reasonable and realistic in quality, there are still significant challenges for large-holed inpainting and diversity inpainting. Models trained adversarially generate realistic but not diverse results, and models optimizing the lower bound of likelihood fit the original data distribution, resulting in blurred reconstructed images. To overcome these problems, we propose a new image inpainting method that adopts a strategy based on a denoising diffusion probability model. The diffusion model directly optimizes the computation of data loglikelihood, thus achieving more accurate inference. We introduce additional condition during network training and sampling to guide image generation. By adjusting the model structure, we introduce an attention mechanism into the network predicting noise to improve the quality of image generation. We conducted experiments on CelebA-HQ and FFHQ to evaluate the effectiveness of our model. Compared with the current most advanced inpainting methods, our model can generate higher quality and more diverse images.