Sampling variability ranks reconstruction error in lacquerware image inpainting across models and collections
Year: 2026
Type: preprint
Abstract: Generative inpainting can produce visually persuasive but inaccurate completions of lacquerware photographs. We examined whether variability across three seeded reconstructions predicts the masked error of a fourth, held-out reconstruction under synthetic occlusion. The analysis included 64 original objects, 40 external museum catalogue entries, and two diffusion models. Object-level Pearson correlations between variability and error were 0.7176 and 0.7664 in the original cohort, and 0.9043 and 0.9162 externally. Retaining approximately 40% of the lowest-variability entries reduced the mean masked error of selected outputs in each analysis, although no reconstruction was improved. The incremental advantage over inexpensive image descriptors was uncertain in the original cohort. Subsequent two-expert foreground and material assessment supported the association in restricted external subsets, while anonymous visual ratings showed limited agreement despite adjacent-grade differences. Sampling variability may help prioritize synthetic-loss reconstructions for human review; it does not establish historical authenticity or conservation suitability.
Source: Research Square
Authors: Peng Yue¹, Jia Hou², Xiao Zhang³
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225912583
