Diffusion-Based Restoration of Corroded Bronzeware Images under a Minimal-Intervention Framework: Spatial-Compliance Indices and Over-Restoration Risk Diagnostics

Corroded bronzeware photographs preserve evidential textures such as corrosion and patina. Digital restoration should repair losses without rewriting these traces. This study evaluates diffusion-based inpainting for bronzeware restoration under a minimal-intervention principle. LaMa, Stable Diffusion Inpainting, and ControlNet are compared using a three-part evidence chain: (i) synthetic-mask proxy tests for image quality, (ii) spatial-compliance indices (IIR/BSR/OIR) and \(\:\varvec{\varDelta\:}{\varvec{E}}_{00\:}\--{\tau\:}\) risk maps computed with a \(\:\text{r}\) = 10 px boundary band, and (iii) blinded expert ranking and acceptability judgments. In proxy tests, diffusion models typically improved perceptual fidelity and global coherence. In real-damage cases, however, they produced higher boundary spillover and larger out-of-mask change footprints under \(\:{\tau\:}_{-}case\), indicating over-restoration risks to evidential textures. LaMa achieved comparable in-mask completion with markedly lower spillover. Experts treated most outputs as draft-level aids rather than publishable conservation surrogates. The proposed framework enables transparent, auditable decisions that balance quality, risk, and authenticity.