Abstract
<title>Abstract</title> <p>Garments in ancient paintings often exhibit broken contours and missing drapery details caused by aging. General-purpose inpainting models lack garment-specific structural priors and often overlook small, narrow, and elongated defects. This study proposes a two-stage framework combining Fashion LoRA guidance with Garment Region-Reweighted Flow Matching (GR-FM). First, Fashion LoRA is trained on modern garment line art and structural text to initialize outlines and folds during color restoration. Second, GR-FM normalizes and reweights latent-token regression errors in damaged regions within FLUX.1-Fill, strengthening local supervision. Hard masking is applied during inference to confine modifications. Experiments using four control configurations and three random seeds show that Fashion LoRA mainly improves edge quality in the modern source domain, whereas GR-FM provides more stable cross-domain gains. The complete framework improves pixel, structural, and perceptual metrics on ancient painting datasets and produces digital restoration candidates with traceable spatial boundaries, supporting structure-constrained and auditable cultural heritage restoration.</p>