A Low-Cost Skip Connection Weight Tuning Strategy for Lace Image Generation Based on Diffusion Models

The production of lace patterns faces significant challenges due to the creative limitations and high costs associated with manual design. Diffusion models offer a promising solution for lace image generation, with U-Net-based diffusion architectures being widely adopted for their strong multi-scale feature extraction capabilities. In these architectures, the design of skip connections plays a particularly critical role in preserving fine-grained details. However, existing skip connection weight tuning methods perform poorly when applied to lace images, which are characterized by intricate textures and complex patterns. These methods often fail to effectively maintain the detailed structure and overall integrity of lace designs. In this work, we adopt the EDM2 diffusion model for lace image generation and introduce a training-free skip connection weight tuning strategy. During the inference phase, the skip connection weights in the U-Net structure are selectively adjusted: shallow layers retain their original weights, while the weights of deeper skip connections are progressively increased up to 1.1. This targeted adjustment enhances the model's ability to preserve fine texture information from deeper features. Experimental results and analysis demonstrate that the proposed method consistently reduces both FID and FDDinova scores across multiple training epochs. The approach is simple to implement, incurs no additional training cost, and exhibits strong practicality and scalability. It offers a new perspective for generating high-quality lace texture images.