Enterprise AI Analysis

A high-quality dataset of 494 Cantonese embroidery artworks was constructed, overcoming limitations of online image quality. It includes 45 thematic categories, with precise semantic labeling using WD14-tagger and manual correction. This meticulous data preparation is crucial for training robust AI models for specialized artistic styles.

The LoRA (Low-Rank Adaptation) fine-tuning method was applied to the Stable Diffusion 1.5 base model using the curated dataset. This plug-and-play approach efficiently adapts the large model to Cantonese embroidery textures with limited training data, achieving high-fidelity texture simulation while maintaining visual harmony. Optimal performance was observed at the 8th epoch with a weight of 0.9.

ControlNet provides multi-condition guidance (Depth, LineArt, Color) to preserve geometric structure and color consistency. SAM (Segment Anything Model) ensures high-precision spatial semantic constraints, preventing misplacement and refining local details. This combined guidance mechanism is critical for generating visually coherent and semantically accurate embroidery patterns.

The proposed method significantly outperforms baseline models (VGG19, CycleGAN, AdaIN, Cross-Image-Attention) across LPIPS, FID, and PSNR metrics, demonstrating superior texture reconstruction and detail generation. Qualitative analysis confirms its ability to reproduce unique silk textures and stitching patterns with natural color transitions, even at high resolutions.

Expert user evaluation validated the method's superior performance in craftsmanship reproduction, material texture realism, and color fidelity. The high inter-rater consistency (Alpha > 0.70) across all metrics underscores the robustness and practical applicability of the generated embroidery styles for digital preservation and innovative design.