面向住宅室内装饰设计的可控生成与微调优化方法研究  

Research on a controllable generation and fine-tuning optimization method for residential interior decoration design

作　　者：冯佛雄 卢昱杰 仲涛 Feng Foxiong;Lu Yujie;Zhong Tao

机构地区：[1]同济大学土木工程学院

出　　处：《当代建筑》2025年第2期171-176,共6页Contemporary Architecture

基　　金：国家重点研发计划项目“工程建造云边端数据协同机制与一体化建模关键技术”(2022YFC3801700);中国工程院咨询项目“智能建造发展战略研究”(2024-XZ-37);中央高校基本科研业务费专项项目“基于视觉—语言模型的施工安全隐患感知、评估与预控方法”(2024-1-ZD-02);上海市科技创新行动计划项目“大型公共建筑超低能耗智慧运维技术研究”(22dz1207100);中国高校产学研创新基金项目“基于图像生成式技术的建筑室内装饰智能化设计方法研究”(2024SE027)。

摘　　要：针对传统住宅室内装饰设计方法设计周期长、优化效率低、个性化适应能力不足等问题,本文提出了一种面向住宅室内装饰设计的可控生成与微调优化方法。本文基于稳定扩散模型,结合低秩适应优化与材质控制机制,构建了一套涵盖数据集构建、图像生成、材质控制及局部定向优化的完整方法体系,提高了方案生成的精准度与设计适应性,以期为住宅室内装饰设计的智能化发展提供新的技术范式与实践参考。Traditional residential interior decoration design heavily relies on manual drafting and iterative refinements,resulting in prolonged design cycles,low optimization efficiency,and limited adaptability to personalized requirements.With the growing demand for intelligent and automated design solutions,the integration of generative artificial intelligence(AI)into interior design workflows has become an inevitable trend.Recent advancements in generative models,particularly diffusion models,have provided significant breakthroughs in automated design.However,existing methods still face critical challenges,including limited control over design outputs,inadequate customization to meet specific aesthetic and functional requirements,and a lack of constraints for material selection.To address these challenges,this paper proposes a controllable generation and fine-tuning optimization framework for residential interior decoration design.By integrating the Stable Diffusion model with Low-Rank Adaptation(Lo RA)optimization,the Material Control constraint mechanism,and IP-Adapter for localized style transfer,this paper establishes a comprehensive methodology encompassing dataset construction,image generation,material control,and localized optimization.This paper consists of four core components:dataset construction,image generation,material-aware control,and localized optimization.To improve the applicability of generative models in residential interior design,this paper developed a well-structured and meticulously annotated dataset covering modern,minimalist,and classic interior styles.The dataset underwent rigorous preprocessing,including noise reduction,error correction,and hierarchical labeling based on standardized building information classification protocols.This structured annotation methodology ensures robust model training and enhances the generalization capability of the generated outputs.For model optimization,Lo RA training was employed to facilitate efficient style adaptation of the diffusion model with minimal com

关 键 词：住宅室内设计 可控生成 低秩适应模型 材质控制 智能设计 

分 类 号：TU17[建筑科学—建筑理论]