A Likely Geometry of Generative Models.

Experimential results for the paper "A Likely Geometry of Generative Models" (https://arxiv.org/abs/2510.26266).

Installation and Requirements

The implementations in the GitHub is Python 3.10.12 and has been implemented in both JAX and PyTorch.

Reproducing Experiments

All experiments can be re-produced by running the notebooks and the runtime.py and run_gen2d_geometry for the given data, method and hyper-parameters. The ControlNet interpolation can be found at https://github.com/FrederikMR/controlnet_interpolation/tree/main. The Energy-based interpolation can be found at https://github.com/FrederikMR/ebm_interpolation.

Reference

If you want to use the algorithm or the method proposed in the paper for scientific purposes, please cite:

@misc{rygaard2026likelygeometrygenerativemodels,
      title={A Likely Geometry of Generative Models}, 
      author={Frederik Möbius Rygaard and Shen Zhu and Yinzhu Jin and Søren Hauberg and Tom Fletcher},
      year={2026},
      eprint={2510.26266},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2510.26266}, 
}