Evaluating Diffusion-Based Super-Resolution for Trustworthy Quantitative Metallography

Super-resolution (SR) holds promise for improving metallographic analysis, but diffusion-based methods raise concerns about hallucinated structures that could bias quantitative results. We present the first systematic study of diffusion SR in quantitative metallography. Using OSEDiff with a fixed domain prompt, we generate a fourfold super-resolved version of the Texture Boundary in Metallography (TBM) dataset (SR-TBM) and train uncertainty-aware edge detectors on both original and SR images. Expert audit confirmed that SR-TBM introduces no spurious grain boundaries, establishing that diffusion SR can be trusted under domain-guided prompting. At the same time, models trained on SR-TBM achieve a 47\% reduction in grain-size error (Heyn intercept metric) compared to models trained on original TBM, surpassing prior baselines including MLOgraphy and AutoSAM. These results demonstrate that diffusion SR, when guided appropriately, both preserves scientific validity and enhances performance in grain-size estimation. We release SR-TBM and code (https://github.com/Scientific-Computing-Lab/SR-TBM) to encourage reproducible, physics-aware evaluation of generative enhancement methods in materials science.