Using DeepSeek-VL Model to Build Multi-Style Interior Decoration Design Generation and Semantic Consistency Enhancement Mechanisms
X. H. Zhang · Advanced Electromagnetics · 2026
This paper proposes a multi-style interior decoration design generation framework based on SDXL, integrating the DeepSeek-VL model, a local cross-modal attention mechanism, and an example-driven semantic consistency enhancement mechanism to effectively address feature confusion and text-image semantic inconsistency during multistyle switching. Beyond intelligent interior visualization, the proposed framework provides reliable semantic modeling capabilities for digital environments that can support engineering-oriented applications such as electromagnetic-aware building planning and intelligent spatial design. Style vector embedding is employed for accurate style representation, while the SDXL conditional decoder enables multi-style feature fusion. DeepSeek-VL achieves global semantic alignment, the local cross-modal attention module verifies the spatial correspondence of key design elements, and the example-driven retrieval mechanism injects features from similar real samples into the UNet layer to improve detail generation. Experimental results demonstrate outstanding performance in semantic consistency (similarity 0.942), style discriminability (distribution entropy 0.88), and visual quality (average MOS 4.65). The proposed framework provides an effective solution for automated multi-style interior design while offering a reliable semantic generation paradigm for intelligent spatial modeling and engineering visualization in complex built environments.