Abstract
Diffusion models, a specific type of generative model, have achievedunprecedented performance in recent years and consistently produce high-qualitysynthetic samples. A critical prerequisite for their notable success lies inthe presence of a substantial number of training samples, which can beimpractical in real-world applications due to high collection costs orassociated risks. Consequently, various finetuning and regularizationapproaches have been proposed to transfer knowledge from existing pre-trainedmodels to specific target domains with limited data. This paper introduces theTransfer Guided Diffusion Process (TGDP), a novel approach distinct fromconventional finetuning and regularization methods. We prove that the optimaldiffusion model for the target domain integrates pre-trained diffusion modelson the source domain with additional guidance from a domain classifier. Wefurther extend TGDP to a conditional version for modeling the jointdistribution of data and its corresponding labels, together with two additionalregularization terms to enhance the model performance. We validate theeffectiveness of TGDP on Gaussian mixture simulations and on realelectrocardiogram (ECG) datasets.