Abstract
We propose a novel unpaired image-to-image translation method that usesdenoising diffusion probabilistic models without requiring adversarialtraining. Our method, UNpaired Image Translation with Denoising DiffusionProbabilistic Models (UNIT-DDPM), trains a generative model to infer the jointdistribution of images over both domains as a Markov chain by minimising adenoising score matching objective conditioned on the other domain. Inparticular, we update both domain translation models simultaneously, and wegenerate target domain images by a denoising Markov Chain Monte Carlo approachthat is conditioned on the input source domain images, based on Langevindynamics. Our approach provides stable model training for image-to-imagetranslation and generates high-quality image outputs. This enablesstate-of-the-art Fr\'echet Inception Distance (FID) performance on severalpublic datasets, including both colour and multispectral imagery, significantlyoutperforming the contemporary adversarial image-to-image translation methods.