Iterative generation of Huizhou Three Carvings micro-cultural and creative products based on contrastive learning and latent diffusion models

This is the authors' abstract. We don't add key points for this paper.

To address the problems of pattern blurring, category indistinguishability, and cultural semantic deviation in intelligent generation methods for Huizhou Three Carvings intangible cultural heritage patterns, this study develops a multi-module collaborative optimization iterative generation model for Huizhou Three Carvings micro-cultural and creative products. The proposed model utilizes Contrastive Generative Adversarial Networks to avoid feature confusion among various carving types, designs a UNet backbone architecture based on a Latent Diffusion Model, and integrates a CNN-based discriminator to improve carving texture feature extraction. Finally, a dynamically weighted composite loss function is introduced to improve the performance of the task objective, thereby enhancing the model's capacity to represent Huizhou carving texture features and cultural semantics. Experimental results demonstrate that the Fréchet Inception Distance of the proposed model is only 7.26 when the number of negative samples is 32, the Learned Perceptual Image Patch Similarity is merely 0.083, the Structural Similarity Index Measure reaches 0.941, and the Peak Signal-to-Noise Ratio is 29.67 dB. Meanwhile, the classification accuracies for brick carving, wood carving, and stone carving are 94.27%, 93.61%, and 92.05%, respectively, with a Mean Average Precision of the test set reaching 92.74% and a Recall@5 reaching 98.13%. In summary, the proposed model exhibits minimal perceptual distortion, high pixel fitting accuracy, and reliable texture reconstruction capabilities, thereby enabling the digital generation of Huizhou Three Carvings patterns. This study provides reliable technical support for the intelligent iteration of Huizhou Three Carvings patterns and the innovative design of micro-cultural products.