Chronology-aware generation of early Chinese black-and-white traditional motifs via an edge-aware self-attention variational autoencoder

Chinese black-and-white traditional motifs are important visual records of cultural symbolism, ritual expression, and stylistic evolution, yet their small and poorly annotated image collections limit computational analysis and generation. Here, we present a chronology-aware generative framework for digital heritage representation using a curated dataset of 300 traditional pattern images from four historical periods. The framework integrates ResNet18 feature extraction, LSTM-based temporal modeling, and an edge-aware self-attention variational autoencoder to capture period-specific motif features while preserving high-contrast contours. A Sobel-based edge loss further strengthens structural fidelity in generated motifs. Compared with Edge-guided GAN, Edge-conditioned diffusion, VAE, and CVAE baselines, the proposed model improves KID, LPIPS, Edge-SSIM, SSIM, and PSNR. The results indicate that chronological conditioning and edge-sensitive reconstruction can improve contour preservation and structural readability in the present small-sample setting.