Functional Textile Pattern Generation Based on Visual Semantics and Aesthetics Using Diffusion Models

Yingying Xu , Minwei Zhao , Hao Nong

Article
2026 / Volume 9 / Pages 1-18
Received 6 August 2025; Accepted 22 August 2025; Published 16 January 2026
https://doi.org/10.31881/TLR.2026.001

Abstract
This paper presents a novel method for generating functional textile patterns by introducing a dual-guidance diffusion model that disentangles functional control from stylistic description. We achieve this by translating abstract functional requirements into quantifiable semantic vectors, which are injected into the generation process via a dedicated encoder and cross-attention mechanism, enabling precise control over functional visual features. To this end, we first built a textile pattern dataset that links functional attributes such as “moisture-wicking” to corresponding visual elements such as channel-like textures. We then developed a Functional-Semantic Guided Diffusion Model (FS-DM). This model incorporates a specialized encoder into the LDM’s U-Net, using a cross-attention mechanism to inject the desired functional semantics into the pattern generation process. This design enables fine-grained control over the final pattern’s structure and appearance. The generated patterns were assessed using a multidimensional evaluation framework. Experimental results show that the method successfully creates patterns with clear visual guidance based on predefined functions. These patterns achieved a functional visual semantic relevance score 35.7% higher than baseline models and also scored exceptionally well on computational aesthetic metrics and in subjective user evaluations. This research offers a new, intelligent, and efficient paradigm for functional textile design with significant industrial application value.

Keywords
diffusion model, functional textile, pattern design, visual semantics, deep learning