Pejman Tahmasebi
This work presents a latent diffusion modeling framework for generating and analyzing high-resolution porous media images in a computationally efficient manner. By encoding complex microstructures into a low-dimensional latent space using a variational autoencoder, the generative process is performed in a compressed domain, drastically reducing the cost compared to pixel-space diffusion models. The proposed approach enables realistic synthesis of porous structures, interpolation between samples, and generation of 3D volumes from learned 2D statistics. A comprehensive set of evaluations is conducted to examine the effect of latent spatial resolution, the continuity of interpolation paths, and the structural consistency of generated samples. The trajectories of the interpolations further reveal the sensitivity of latent transitions to variations in porosity and morphology. The reconstructed 3D samples exhibit plausible connectivity and spatial organization, and statistical comparisons based on several morphological and physical properties confirm the realism of the generated volumes.