Detail-rich pore-scale reconstruction with conditional latent diffusion

Accurate characterization of pore-scale structures is essential for modeling flow and transport in porous media, yet high-resolution imaging is often limited by acquisition cost and resolution constraints. In this paper, we propose a generative framework that leverages a Variational Autoencoder (VAE) and a Latent Diffusion Model (LDM) to reconstruct fine-scale porous media images from their coarse counterparts. The VAE is trained to encode high-resolution porous structures into a compact latent space, enabling efficient and physically meaningful representations. A conditional diffusion model is then trained in this latent space to iteratively denoise coarse-resolution inputs and generate realistic structures consistent with the encoded statistics of the training data. This two-stage architecture decouples image complexity from generative sampling and allows for computationally efficient learning of multiscale features. We demonstrate the effectiveness of this approach on a dataset of porous media images, both qualitatively and quantitatively, and show that the generated samples preserve key structural and statistical properties such as pore connectivity and phase distribution compared to the original high-resolution version. • Latent diffusion enhances resolution of complex porous media images • Combines VAE compression with UNet-based denoising in latent space • Preserves pore geometry and porosity in super-resolved reconstructions • Outperforms traditional interpolation methods