Vittoria De Pellegrini, and Tariq Alkhalifah

 GitHub: https://github.com/DeepWave-KAUST/LAViG-FLOW-pub

LAViG-FLOW is our latent autoregressive video diffusion model that jointly learns CO₂ gas saturation and pressure build-up evolution across diverse CO₂ injection scenarios. It uses paired autoencoders plus a Video Diffusion Transformer to model both fields together, then fine-tunes autoregressively to extend predictions beyond the training horizon while maintaining physical consistency. On an open CO₂ sequestration benchmark, it delivers realistic videos, offering a fast, flexible alternative to traditional simulators.