Na Liang, Hui Zhou, Huaxia Xia
College of Humanities and Arts, Hunan International Economics University, Changsha, Hunan, China.
School of Design and Art, Hunan University, Changsha, Hunan, China.
PLoS One. 2025 Sep 10;20(9):e0329487. doi: 10.1371/journal.pone.0329487. eCollection 2025.
This study addresses the limitations of traditional interior space design, particularly the timeliness and uniqueness of solutions, by proposing an optimized design framework that integrates a two-stage deep learning network with a single-sample-driven mechanism. In the first stage, the framework employs a Transformer network to extract multi-dimensional features (such as spatial layout, color distribution, furniture style, etc.) from input space images, generating an initial feature vector. In the second stage, a diffusion model is introduced to iteratively optimize the design results based on user-provided single-sample features. The model's performance is validated using multiple publicly available datasets, including InteriorNet, SUN RGB-D, NYU Depth V2, and ScanNet. Experimental results demonstrate that, compared to traditional methods, the design cycle is reduced by 40%, space utilization is increased by 25%, and proportional and scale coordination is improved by 20%. The single-sample-driven personalized design strategy results in a 30% significant improvement in color matching scores. Through the synergistic effect of feature extraction and generative optimization, the two-stage network enhances both design efficiency and the innovativeness and user adaptability of the solution. This study not only offers an efficient and intelligent solution for interior space design but also presents a new technological paradigm for the advancement of artificial intelligence-driven design fields.
本研究通过提出一种优化设计框架来解决传统室内空间设计的局限性，特别是解决方案的及时性和独特性，该框架将两阶段深度学习网络与单样本驱动机制相结合。在第一阶段，该框架采用Transformer网络从输入的空间图像中提取多维特征（如空间布局、颜色分布、家具风格等），生成初始特征向量。在第二阶段，引入扩散模型根据用户提供的单样本特征对设计结果进行迭代优化。使用包括InteriorNet、SUN RGB-D、NYU Depth V2和ScanNet在内的多个公开可用数据集对该模型的性能进行了验证。实验结果表明，与传统方法相比，设计周期缩短了40%，空间利用率提高了25%，比例和尺度协调性提高了20%。单样本驱动的个性化设计策略使色彩匹配得分显著提高了30%。通过特征提取和生成优化的协同作用，两阶段网络提高了设计效率以及解决方案的创新性和用户适应性。本研究不仅为室内空间设计提供了一种高效、智能的解决方案，还为人工智能驱动的设计领域的发展提出了一种新的技术范式。
