关键词: 数码印花图案, 跨模态生成, 色彩语义一致性, 扩散模型, 印花图案个性化定制, 人工智能生成内容技术

Abstract:

Objective Textile digital printing technology is popular in the textile dyeing and printing industry because of its high printing precision, fast production speed, and environmentally friendly processes. However, current digital printing pattern design still relies heavily on manual creation by designers, leading to long cycles, low efficiency, and poor market responsiveness. Leveraging Artificial Intelligence Generated Content (AIGC) technology to achieve rapid text-to-digital printing pattern generation can significantly improve the development efficiency of digital textile products and support user personalized customization, opening up new methods for digital pattern design. Nevertheless, mainstream text-to-image generation models (e.g., Stable Diffusion (SD) trained on generic datasets) have two key limitations, i.e. monotonous generated patterns lacking artistic diversity, and mismatch between the generated colors and the text-described color semantics due to inaccurate color attribute binding. Therefore, it is crucial to develop a specialized generative model for textile digital printing patterns and propose new methods to improve color generation accuracy.

Method A dataset containing 3 090 pairs of digital textile patterns and detailed textual descriptions was first constructed. The SD-V1.5 model was fine-tuned using the Low-Rank Adaptation (LoRA) technology. Subsequently, for the fine-tuned SD model, this study introduced a cross-attention-based color excitation module into the first 70% of the denoising process, without requiring any additional training. This module comprises two loss components, namely the target object color excitation loss and the background color excitation loss. By minimizing these two losses through gradient descent, the model learns to increase the attention weights of color-related keywords in the text prompt, thereby focusing more on the target and background colors of the pattern during generation. Finally, the color semantic consistency was evaluated using the CLIP Score, Text-Text Similarity (T2T-Sim), and a newly proposed Color-Text Accuracy (C2T-Acc) metric.

Results The effectiveness of the proposed method was evaluated through both qualitative and quantitative experiments. Qualitatively, the generated patterns under various text prompts and random seeds showed consistent style, pattern color, and background color with the textual descriptions. Benefiting from the diverse pattern types and richly annotated descriptions in the constructed dataset, the generated results also exhibited strong artistic creativity. Quantitatively, although the introduction of attention excitation caused slight perturbations to noise prediction, resulting in a minor decrease in the CLIP Score metric, the T2T-Sim and C2T-Acc indicators were improved by 2.96% and 8.94%, respectively, compared to the more advanced Structured Diffusion and Attend-and-Excite models. These results indicate that the proposed method not only significantly enhances the color semantic consistency of the patterns but also demonstrates substantial improvements in overall pattern quality and visual fidelity. Furthermore, ablation studies show that both the target color excitation and background color excitation contribute positively to the overall improvement, leading to better color generation, detailed pattern formation, and accurate background rendering.

Conclusion This study investigates cross-modal generation of textile digital printing patterns from textual descriptions. The powerful generative model SD-V1.5 is adopted as the baseline and fine-tuned on a custom-built dataset of textile printing patterns. In order to enhance the consistency between generated pattern colors and color semantics in the text, a cross-attention-based color excitation method is introduced, which increases the attention weights of color-related keywords during generation. Extensive qualitative and quantitative evaluations confirm the effectiveness of the proposed method. With only a textual input, users can efficiently obtain semantically aligned and high-quality textile digital printing patterns, enabling convenient, personalized customization. By leveraging advanced AIGC technologies, the proposed approach establishes a domain-specific generation model for textile digital printing, contributing to the development of new productive forces and facilitating the digital, intelligent, and sustainable transformation of the next-generation textile printing industry.

Key words: digital printing pattern, cross-modal generation, color-semantic consistency, diffusion model, personalized pattern customization, artificial intelligence generated content