GenIA

Stefano Esposito¹, Naama Pearl¹, Polina Karpikova¹, Samuel Rota Bulò², Lorenzo Porzi², Peter Kontschieder², Andreas Geiger¹, Jonathon Luiten^3
1Tübingen AI Center, University of Tübingen, ²Meta, Zürich, ³Meta, USA

GenIA reconstructs complete, detailed and input-aligned 3D objects from a single image, a few views of a static scene, or a monocular video. It keeps the generative prior of SAM 3D Objects frozen and aligns it to the observations at inference time with geometric and photometric constraints: no retraining, no weight updates. The output is a canonical object (Gaussians and a mesh) with per-frame poses and, for video, a per-frame deformation.

Requires conda, a CUDA 12.8 toolkit and an NVIDIA GPU.

git clone https://github.com/facebookresearch/GenIA && cd GenIA
bash install.sh && conda activate genia   # also fetches the submodules
python -m genia.core.download_weights   # all model weights (accept the SAM 3D Objects licence on Hugging Face first)

demo/data/ holds example inputs for the three settings:

python demo/run_demo.py image       # a single image
python demo/run_demo.py multiview   # a few views of a static scene
python demo/run_demo.py dynamic     # a monocular video (shapes from ActionMesh)

Results are written to results/. To run your own data, add it as a new scene folder under demo/data/<setting>/, laid out like the examples, and pass --scene <name>. See docs/configuration.md for the configuration options.

This project is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). See the LICENSE file for details.

Third-party components keep their own licences, several of them non-commercial; see THIRD_PARTY_LICENSES.md before any use.

If you use this work in your research, please consider citing:

@article{esposito2026genia,
  title   = {GenIA: Generative Reconstruction with Test-Time Input Alignment},
  author  = {Esposito, Stefano and Pearl, Naama and Karpikova, Polina and
             Rota Bul{\`o}, Samuel and Porzi, Lorenzo and Kontschieder, Peter and
             Geiger, Andreas and Luiten, Jonathon},
  journal = {arXiv preprint arXiv:2610.12388},
  year    = {2026}
}

Built on SAM 3D Objects. Geometry from MapAnything and MoGe; dynamic shapes from ActionMesh; rendering via gsplat and nvdiffrast.