文献类型：期刊文献

英文题名：Few-Shot Adaptive Diffusion with Semantic Injection and Parameter Smoothing

作者：Cai, Yunjie[1]; Xiao, Ting[1]; Zhang, Yanbing[1]; Wang, Zhe[1]

机构：[1] Key Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, Department of Computer Science and Engineering, East China University of Science and Technology, Shanghai, China

年份：2025

起止页码：52

外文期刊名：ICMR 2025 - Proceedings of the 2025 International Conference on Multimedia Retrieval

收录：EI(收录号：20253018868633)

语种：英文

外文关键词：Diffusion - Image texture - Quality control - Semantic Web - Semantics - Textures

摘要：Although significant progress has been made in the field of image generation, training generative models with a limited number of samples remains challenging and is under-explored. Existing approaches primarily rely on few-shot model adaptation for network training. However, in situations where the data is extremely scarce (less than 10 samples), the generative network is prone to over-fitting and suffers from a decline in fine texture details and poor generalization. To address these issues, we propose a new few-shot generative adaptation diffusion method based on semantic injection and parameter smoothing. Specifically, we introduce a semantic injection module within the diffusion model, which effectively integrates external structure information with internal semantic information through a two-stage injection mechanism, enhances the detailed texture of the generated images, and aligns the spatial structure. Furthermore, we introduce a cross-domain parameter smoothing strategy, which integrates style loss and diffusion loss and updates parameters in an exponential moving average manner. By gradually incorporating knowledge from the source domain, we can improve both the fidelity of the generated images and the stability of the model. Finally, extensive qualitative and quantitative experiments on multiple few-shot generative adaptation tasks demonstrate the superiority of our approach in generating images with fine texture details and improving the generalization ability of the model. ? 2025 ACM.