Masked Face Inpainting Using Denoising Diffusion Probabilistic Models

Image inpainting in computer vision is a challenging task, as most existing approaches are trained for certain mask distributions, which limiting their generalization capability when confronted with unknown mask types. However, rather than producing semantically significant texture, training with pixel-level and perceptual losses frequently results in the extrapolation of basic texture in absent areas. In this paper, we propose an image inpainting approach based on Denoising Diffusion Probabilistic Model (DDPM). We utilize a pretrained unconditional DDPM as a generative prior. We alter the generative process of the model, enabling accelerated resampling for unknown regions based on known information from the given image. Since our approach does not involve any adjustments or modifications to the DDPM network itself, it can handle any form of mask and generate high-quality and diverse outcomes. The CelebA-HQ dataset experimental results show that our approach produces state-of-the-art performance.