MotifDiff Synthetic Motif Images for CTPD Long-Tail Augmentation

Synthetic training images generated to augment the CTPD (Chinese Traditional Pattern hierarchical detection) benchmark for the MotifDiff study. Images are produced with SDXL conditioned on intangible-cultural-heritage semantic prompts per motif class, with scale-controllable synthesis, CLIP-based filtering, and a noise curriculum; a YOLOv8-s detector is fine-tuned (80 epochs) on the merged real+synthetic sets. Three archives are provided: 800 raw tail-class images (4 classes), 400 CLIP-filtered tail-class images (4 classes), and 1000 mid-frequency-class images (5 classes), plus the generation and evaluation code. Source benchmark: CTPD (Li et al., 2026, Science Data Bank, DOI 10.57760/sciencedb.34731).