To address morphological distortion and background interference in digital restoration of costume patterns from ancient Chinese paintings, we propose PT-RFLow, a tiling-guided adaptive targeted image editing method built on the Kontext context-aware Rectified Flow Transformer. It enhances pattern morphological feature and arrangement rule learning via targeted low-rank fine-tuning, with a dual-modal guidance enhancement mechanism to boost tiling restoration accuracy. To map locally distorted pattern regions to their complete planar form, we construct a cross-dynasty triplet dataset of typical patterns for model training and testing. Experiments show PT-RFLow significantly outperforms baselines in core metrics including Structural Similarity, Perceptual Fidelity and Tiling Regularity; expert evaluation verifies its advantages in structural accuracy, style fidelity and aesthetic consistency, providing an effective technical approach for digital restoration of traditional Chinese patterns.
Guo et al. (Sat,) studied this question.