Corroded bronzeware photographs preserve evidential textures such as corrosion and patina that should not be unnecessarily rewritten during digital restoration. This study evaluates LaMa, Stable Diffusion Inpainting, and ControlNet within a minimal-intervention framework. Synthetic-mask proxy tests were used to assess in-mask reconstruction, while real-damage cases were evaluated using spatial-compliance indices, risk maps, and blinded expert review. Diffusion models improved perceptual fidelity and global coherence in proxy tests, but in real-damage settings they produced greater boundary spillover and out-of-mask rewriting, indicating a higher risk of over-restoration. LaMa achieved comparable in-mask completion with markedly lower spillover. Experts regarded most outputs as draft-level aids rather than publishable conservation surrogates. The proposed framework supports transparent and auditable evaluation of restoration quality, intervention risk, and authenticity.
Sui et al. (Thu,) studied this question.