Authors: Priyadharshini P, M.Gayathri
Abstract: This research paper presents a comprehensive implementation and evaluation of the Single Image Denoising Diffusion Model (SinDDM) for sophisticated image editing and synthesis tasks using only a single training image. Unlike conventional diffusion-based generative models that rely on extensive datasets, SinDDM employs an innovative multi-scale training strategy to learn hierarchical priors from a single input image. The model supports a wide range of image manipulation tasks, including artistic style transfer, semantic image harmonization, region-of-interest (ROI) guided editing, and CLIP-based text-guided content generation. Experimental results demonstrate that SinDDM consistently produces coherent, high-quality, and semantically aligned outputs without requiring extensive training data or pre-trained encoders, making it particularly suitable for personalized applications and data-efficient computational scenarios. This paper provides detailed architectural insights, implementation methodologies, comparative analysis, and potential applications of the proposed framework
DOI: http://doi.org/10.5281/zenodo.17862616
