A conditional latent diffusion framework for dynamic CT image denoising in multiphase flow and reactive transport studies of porous media

In dynamic X-ray CT imaging, there is a trade-off between temporal resolution and image quality. To improve image quality, we present CT-DiffNet, a novel conditional latent diffusion framework that enhances low-quality images when only a limited paired low- and high-quality dataset is available for training. Our method leverages the stable diffusion architecture with pre-trained weights and introduces an innovative adjacent-slice 3-channel input strategy that captures 3D spatial information during 2D training. On the same test set, CT-DiffNet outperforms supervised UNet and classical NLM in PSNR and SSIM and stays closest to the HQ reference in Sobel-based edge statistics. For full 3D volumes, it achieves over 30 dB PSNR and stable SSIM on all axes, reduces stacking-direction stripes relative to a 2D baseline that replicates grayscale into three channels, and keeps porosity, saturations, interfacial geometry, and contact angle distributions near the HQ reference. Qualitative denoising of a later dynamic scan without HQ labels shows less noise and fewer streaks while preserving pore and fluid structure. Using only 20% of the training pairs still yields usable quality with limited degradation, which supports application to data-scarce dynamic CT in porous media, multiphase flow, and reactive transport.