DSG-BDF: a design semantic graph-guided bidirectional diffusion framework for controllable clothing pattern generation based on ControlNet and CLIP ViT-L/14
Diffusion models have shown promise in clothing theme pattern generation, but the lack of a stable cross-modal coupling mechanism between structural constraints and textual semantics leads to semantic drift and style fluctuations under complex design conditions. This paper proposes the Design Semantic Graph Guided Bidirectional Diffusion Framework (DSG-BDF). Design theme text is encoded via the CLIP ViT-L/14 encoder and organized into a design semantic graph capturing pattern element relationships and style attributes, which a graph neural network encodes into structured semantic vectors. These vectors are fused with CLIP features and injected into the UNet cross-attention layer as a semantic conditional path, while ControlNet encodes contour and layout features as a structural constraint path. A bidirectional collaborative module enables mutual modulation between the two paths during denoising, and a joint loss covering semantic consistency, structural preservation, and style stability supervises training. Experiments show that DSG-BDF achieves a CLIP-Score of 0.347 ± 0.003, FID of 19.8 ± 0.4, and SCI of 0.912 ± 0.012, outperforming the strongest baseline (FashionDiff: CLIP-Score 0.341, FID 20.9) and improving over the Stable Diffusion baseline (CLIP-Score 0.289, FID 28.6) by 20.1% and 30.8% respectively. These results demonstrate that the proposed framework achieves stable semantic-structural co-generation under complex design conditions.
Authors
- Zhenzhen Xu
- Xuzhi Sun
- Ge Pan
- Lan Wen
- Yingzi Zhang
- Yi Fang
Institutions
- Suzhou University of Technology (CN)
- Anhui Polytechnic University (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-15
- DOI
- https://doi.org/10.1038/s41598-026-80722-8
- Primary Topic
- Generative Adversarial Networks and Image Synthesis
- Type
- article
- Field-Weighted Citation Impact
- 0.00