A Fast Multi-Scale Textile Pattern Generation Method Combining Layered Loss and Convolutional Attention

Abstract

Existing image generation models frequently encounter challenges such as line discontinuity, structural loss, and poor style adaptability when dealing with pattern designs that possess complex geometric structures. To address these issues, this paper proposes a novel deep neural network-based method for the digital design of traditional patterns, constructing a three-stage human–machine collaborative workflow of “structure generation—sketch translation—style transfer.” By injecting adaptive noise and incorporating a multi-scale discrimination mechanism into the StyleGAN generation algorithm, and by integrating edge consistency and structure-aware losses, the continuity and completeness of line drawing generation are significantly improved. A conditional generative adversarial network (Pix2PixHD) cross-domain mapping model combined with a self-attention mechanism is employed to accurately achieve the automatic conversion of irregular hand-drawn sketches into standardized line drawings. Furthermore, a neural style transfer strategy based on multi-scale feature disentanglement is designed, jointly utilizing the Gram matrix and the Wasserstein distance, and supplemented by a convolutional attention module, to realize high-fidelity fusion between traditional structures and modern styles. Experiments verify that our method delivers superior visual quality and structural fidelity compared with state-of-the-art models. It realizes cloud motif expansion, feature extraction, sketch-to-standard line drawing translation and pattern style transfer, and offers theoretical and practical references for aesthetic and intelligent computer-aided design.