LAViG-FLOW: Latent Autoregressive Video Generation for Fluid Flow Simulations

Reproducible material for LAViG-FLOW: Latent Autoregressive Video Generation for Fluid Flow Simulations - De Pellegrini V., Alkhalifah T.

📊 Project structure

This repository is organized as follows:

• 📂 asset: folder containing logo;
• 📂 gas_saturation_vqvae: set of python scripts to build the VQ-VAE latent space for the CO₂ Gas Saturation field;
• 📂 pressure_buildup_vae: set of python scripts to build the VAE latent space for the Pressure Build-Up field;
• 📂 gas_saturation_pressure_buildup_ditv: set of python scripts to train latent autoregressive video diffusion transformer (DiTV);
• 📂 benchmarking: baseline model folders and evaluation/timing scripts for benchmark comparison (Conv-FNO, FNO, U-FNO, vanilla MIONet, Fourier-MIONet);
• 📂 stylegan-v-main: StyleGAN-V metric utilities to evaluate video-generation quality (FVD/LPIPS, etc.). Source: https://github.com/universome/stylegan-v;
• 📂 dockers: Dockerfile for Docker users who need to build the docker image.

👾 🤖 Getting started

Important

To ensure reproducibility of the results, we suggest using the environment.yaml file when creating an environment.

If you download the repository as a ZIP from GitHub, your local folder may be named LAViG-FLOW-pub-main.
Please rename the folder to LAViG-FLOW before running scripts, since project paths assume this repository name.

Example:

mv LAViG-FLOW-pub-main LAViG-FLOW
cd LAViG-FLOW

Simply run:

./install_env.sh

It will take some time, if at the end you see the word Done! on your terminal you are ready to go.

Remember to always activate the environment by typing:

conda activate lavig-flow

🐳 Docker support

Prebuilt images live on Docker Hub: https://hub.docker.com/r/vittodepe98/lavig-flow.

docker pull vittodepe98/lavig-flow:latest

You can execute the docker pull command from any directory, but running it while your shell is already inside the cloned LAViG-FLOW/ folder keeps you in the right context for the follow-up docker run step.

The published Docker Hub image is generated from LAViG-FLOW/dockers/lavig_flow.Dockerfile. Use it whenever you want to work on the repository without installing the environment locally—the container behaves like a ready-to-go virtual environment with CUDA, PyTorch, and all dependencies baked in.

From a terminal opened in your cloned LAViG-FLOW/ folder, mount the repo checkout and drop into the container (Linux/macOS shells or Windows PowerShell):

docker run --rm -it --gpus all --name lavig-flow-local -v ${PWD}:/mycode -w /mycode vittodepe98/lavig-flow:latest bash

This command assumes you already cloned the repository locally (git clone ... LAViG-FLOW) and pulled the Docker image. It then:

• mounts your current working tree into /mycode inside the container (-v ${PWD}:/mycode) and sets that as the working directory (-w /mycode), so any file edits you do on the host (e.g., in VS Code) show up instantly inside the container and vice versa;
• enables GPU passthrough with --gpus all;
• starts an interactive shell (-it) that is automatically cleaned up afterwards (--rm) and is easy to find/attach to (--name lavig-flow-local).

Once the shell comes up you can run scripts normally (e.g., python gas_saturation_vqvae/train.py ...) and everything executes with the container’s pre-installed toolchain. Editors like VS Code can also “Attach to Running Container” (Remote-Containers extension) to provide a virtual dev environment directly inside lavig-flow-local.

Note

Disclaimer: All experiments have been run on KAUST’s IBEX cluster (dual Intel® Xeon® compute nodes paired with NVIDIA Tesla V100 GPUs). Different environment configurations may be required for different combinations of workstation and GPU.

Acknowledgements

We deeply appreciate the upstream projects and datasets that make LAViG-FLOW possible across its many submodules. You will find references embedded throughout the individual scripts (pointing to original papers, datasets, and codebases); we extend our thanks to every one of those researchers and maintainers for their foundational contributions, and to the KAUST IBEX support team for providing the NVIDIA V100 infrastructure used in our experiments.

👽 🛸 🐄 Cite us

@misc{depellegrini2026lavigflowlatentautoregressivevideo,
      title={LAViG-FLOW: Latent Autoregressive Video Generation for Fluid Flow Simulations}, 
      author={Vittoria De Pellegrini and Tariq Alkhalifah},
      year={2026},
      eprint={2601.13190},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2601.13190}, 
}