Embroidery patterns, as a traditional art form, carry profound cultural significance. With the development of digital technologies, how to effectively extract and preserve embroidery patterns has become an urgent issue. This paper proposes a controllable embroidery image generation method based on diffusion models, aimed at enhancing the structural consistency and color consistency of embroidery patterns. The method uses Flux.1-dev as the core framework, combined with Low-Rank Adaptation (LoRA) for efficient fine-tuning to learn high-frequency embroidery textures. It incorporates ControlNet-Canny and ControlNet-Color to impose explicit constraints on structure and color, while Quickshift clustering is used for region segmentation and mask construction to assist in local re-painting optimization. Experimental results show that the proposed method significantly improves the stability of pattern structure and the accuracy of color, providing a feasible solution for the digital embroidery image generation and synthesis.

image generation, diffusion model, embroidery images, LoRA, ControlNet