The physical properties of clay and cementitious materials are critical in fields ranging from materials science to geological waste disposal. Conventional property simulation relies on three‑dimensional imaging, which is expensive, often inaccessible, and technically limited for certain substances. In contrast, two‑dimensional slices are much easier to acquire, prompting the search for methods that can reconstruct 3D volumes from 2D images.\ \ Recent advances in deep generative models have made this feasible. Among GAN‑based approaches, SliceGAN achieves strong results on homogeneous isotropic and anisotropic systems, yet it struggles to capture the fine details of more complex heterogeneous microstructures, motivating the exploration of alternative frameworks.\ \ We introduce a hybrid strategy that leverages the stability and high‑quality generation of denoising diffusion models. Because true 3D ground truth is unavailable, we replace the conventional denoising loss with an adversarial loss, yielding a stable training process in our experiments. The model jointly optimizes the diffusion noise predictor and a discriminator, allowing it to learn realistic volume distributions without explicit 3D labels.\ \ Our experiments demonstrate that the proposed model generates 3D microstructures of varying complexity with minimal slice artefacts. The generated phase fractions and structural descriptors closely match those of the ground truth, outperforming SliceGAN in both detail preservation and global statistical consistency.\ \ In summary, this stable diffusion‑adversarial framework offers a cost‑effective alternative to expensive 3D imaging, opening new avenues for material characterization and geological modeling.\ \ Review