<p>We present a probabilistic diffusion-based framework for reconstructing scientific microstructure images with missing or corrupted regions. Motivated by challenges in characterizing porous media, our method employs denoising diffusion probabilistic models to learn a conditional distribution over image completions given partial observations. Trained on grayscale images of porous structures, the model generalizes well across samples with varying morphology and entropy. We evaluate its performance on two distinct datasets using a range of masking strategies, including irregular occlusions, large missing regions, and structured patterns such as stripes and cutouts. The proposed model reconstructs high-fidelity completions that are both visually plausible and physically consistent. Quantitative evaluations based on pore size distribution, two-point correlation functions, and pixel-level error metrics show that the generated outputs preserve critical features and statistical descriptors of the original media. Additional analyses of pixel intensity profiles and latent activation patterns reveal that the model can infer fine-scale details while maintaining global structure. These results explain the potential of latent diffusion-based inpainting as a robust tool for digital reconstruction and scientific imaging in complex material systems.</p>