EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

A Prompt-Optimized Vision-Language Model System for Fine-Grained Fashion Style Recognition

Rui Mai, Ting Chen, Xiao Li, Xiaoqun Dai

Journal of Design Intelligence · 05.11.2026 · 1 (4) · 1-13

DOI: 10.64509/jdi.14.118 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Design Intelligence, online erschienen am 05.11.2026. Verfasst von Rui Mai und 3 weiteren Autor:innen. Die Arbeit nennt 25 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

With the accelerated digital transformation of the fashion industry, accurate clothing style recognition has become essential for intelligent e-commerce retrieval, personalized outfit recommendation, and fashion trend analysis. Vision-language models (VLMs), with their cross-modal semantic alignment and zero-shot inference capabilities, offer an efficient technical pathway for this task. However, when applied to the fashion domain, general-purpose VLMs face several challenges, including vague prompt semantics, insufficient fine-grained discrimination, difficulty in capturing composite style characteristics, and high deployment costs. To address these issues, this paper presents a comprehensive clothing style recognition system based on Qwen3-VL-Plus. A four-level structured prompt template—comprising style definitions, visual feature guidance, and confidence distribution constraints—is introduced to direct the model's attention toward fine-grained attributes such as silhouette, fabric, color, and decoration. A confidence distribution modeling mechanism based on Softmax normalization is proposed to convert cross-modal similarity scores into interpretable confidence distributions, enabling quantitative representation of composite styles. Furthermore, INT8 quantization and inference pipeline optimization are employed to achieve lightweight deployment, allowing the system to operate stably in an ordinary CPU environment. On the FashionStyle14 dataset, the proposed system achieves a Top-1 accuracy of 69.21% and a macro-average F1 score of 66.38%, outperforming CLIP by 17.98 and 18.21 percentage points, respectively. INT8 quantization yields a 36.77% reduction in inference time and a 75.12% reduction in model size. In a user experience evaluation involving 30 participants, the system attains an overall satisfaction score of 8.60 out of 10, with 96.67% of users recognizing the interpretability of the recognition results. Experimental results demonstrate that the proposed approach offers significant advantages in recognition accuracy, interpretability, and deployment practicality, providing a complete solution with both technical feasibility and application value for intelligent clothing style recognition in the fashion industry.

Zitieren

APA

Mai, R., Chen, T., Li, X., & Dai, X. (2026). A Prompt-Optimized Vision-Language Model System for Fine-Grained Fashion Style Recognition. Journal of Design Intelligence, 1(4), 1-13. https://doi.org/10.64509/jdi.14.118

Vancouver

Mai R, Chen T, Li X, Dai X. A Prompt-Optimized Vision-Language Model System for Fine-Grained Fashion Style Recognition. Journal of Design Intelligence. 2026;1(4):1-13. doi:10.64509/jdi.14.118

BibTeX
@article{Mai2026,
  author = {Mai, Rui and Chen, Ting and Li, Xiao and Dai, Xiaoqun},
  title = {A Prompt-Optimized Vision-Language Model System for Fine-Grained Fashion Style Recognition},
  journal = {Journal of Design Intelligence},
  year = {2026},
  volume = {1},
  number = {4},
  pages = {1--13},
  doi = {10.64509/jdi.14.118}
}
Kostenlos starten →