In this paper, we propose a learning-free domain adaptation (DA) technique called Domain Noise Alignment (DNA) for diffusion-based dense prediction (DDP) models. Based on the fact that the difference in noise statistics during the diffusion process in the DDP model causes domain shift, we achieve domain adaptation by aligning the noise statistics of the source domain with those of the target domain. In the absence of the source domain, we perform noise statistic adjustment incrementally by utilizing the statistics of high-confidence regions similar to the source domain. We demonstrate the effectiveness of the proposed method through experiments on four common dense prediction tasks.
