ActionMesh: Animated 3D Mesh Generation with Temporal 3D Diffusion

ActionMesh is a generative model that predicts production-ready 3D meshes "in action" in a feed-forward manner. It adapts 3D diffusion to include a temporal axis, allowing the generation of synchronized latents representing time-varying 3D shapes.

[Paper] [Project Page] [GitHub] [Demo]

Installation

ActionMesh requires an NVIDIA GPU with at least 32GB VRAM.

git clone https://github.com/facebookresearch/actionmesh.git
cd actionmesh
git submodule update --init --recursive
pip install -r requirements.txt
pip install -e .

Quick Start

You can generate an animated mesh from an input video using the provided inference script. Model weights will be automatically downloaded on the first run.

Basic Usage

python inference/video_to_animated_mesh.py --input assets/examples/davis_camel

Fast Mode

For faster inference (as used in the Hugging Face demo), use the --fast flag:

python inference/video_to_animated_mesh.py --input assets/examples/davis_camel --fast

Citation

If you find ActionMesh useful in your research, please cite:

@misc{sabathier2026actionmeshanimated3dmesh,
      title={ActionMesh: Animated 3D Mesh Generation with Temporal 3D Diffusion}, 
      author={Remy Sabathier and David Novotny and Niloy J. Mitra and Tom Monnier},
      year={2026},
      eprint={2601.16148},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.16148}, 
}

License

The weights and code are provided under the license terms found in the GitHub repository. Please refer to the LICENSE file there for details.