基于扩散模型的印花图案生成方法设计

Design of a Printed Pattern Generation System Based on Diffusion Models

摘要:

印花图案的设计是服装生产领域非常关键的一环，但目前人工设计的印花图案存在内容相似化、设计效率低下的问题，因此设计并实现了一种基于扩散模型的印花图案生成方法。采用深度学习技术提取、扩充现有印花图案数据集，并从颜色和类别维度生成印花图案的文本描述，完成印花图案数据集制作。使用已制作的数据集微调扩散模型，并对图像特征空间进行平铺处理，使得生成的印花图案在边界过渡处满足纺织行业四方连续的要求。对局部扩散特性进行了分析，结合文本引导，实现细节可变的图文生图效果。实验结果表明，所设计的印花图案生成方法具备生成高质量印花图案的能力，并且其特征空间平铺方法使得印花图案边界过渡处较为自然。

Abstract:

The design of printed patterns is a crucial aspect of garment manufacturing. However, manually designed patterns often suffer from content similarity and low design efficiency. Therefore, a printed pattern generation method based on a diffusion model was designed and implemented. This method employs deep learning technology to extract and expand an existing dataset of printed patterns, then generated textual descriptions of these patterns from the dimensions of color and category to complete the dataset. Fine-tune diffusion model using pre-made dataset, and flatten processing was applied to the feature space of the images, as required in the textile industry. Studies were analyzed on the characteristics of local diffusion, achieving a image generation effect with variable detail from images and text. Experimental results demonstrate that our designed pattern generation method is capable of producing high-quality patterns, and its feature space tiling method provides smooth transitions at pattern edges.