A spatial compliance framework for evaluating minimal intervention in small sample diffusion inpainting for archaeological image restoration

Image inpainting in high-stakes, low-data settings requires not only plausible reconstruction inside damaged regions but also strict restraint against unintended edits outside them. However, conventional restoration metrics primarily measure global fidelity or perceptual similarity and do not explicitly capture out-of-mask intervention behavior. To address this issue, this study proposes a spatial-compliance evaluation framework for small-sample diffusion inpainting and validates it in the restoration of archaeological pottery images. Using 35 high-resolution archival photographs of eight Western Zhou vessels from the Gumei site, we established a unified small-sample protocol based on a “relative ground truth + mask-simulated damage” strategy and compared three representative models: LaMa, Stable Diffusion Inpainting, and ControlNet Inpainting. In addition to PSNR, SSIM, LPIPS, and FID, we introduced response-based spatial compliance indicators to quantify damage coverage and out-of-mask overflow. The three models achieved similar structural fidelity (PSNR approximately 20 dB; SSIM = 0.63–0.67), but differed substantially in intervention behavior.Under the present exploratory protocol, ControlNet showed the lowest FID among the three models and a more favorable balance between distributional consistency and spatial restraint (FID = 126.95, compared with 138.52 for Stable Diffusion and 144.32 for LaMa). Stable Diffusion exhibited stronger damage coverage but also the greatest overflow risk (miscoverage area ratio = 0.0797), whereas ControlNet reduced this value to 0.0450 while maintaining competitive restoration quality. These results show that restoration quality in small-sample inpainting cannot be adequately characterized by conventional metrics alone. The proposed framework provides an interpretable and reusable basis for behavior-aware model evaluation and selection in minimal-intervention restoration scenarios. Given the limited size of the available archaeological image set, the results are interpreted as exploratory evidence under a controlled small-sample protocol rather than as a large-scale benchmark.