This code is a Fork of the original TabDiff library which we made due to licencing. We used the code to generate synthetic data for experiments in our paper:
@misc{Lautrup2025_dgms,
title={Disjoint Generative Models},
author={Anton Danholt Lautrup and Muhammad Rajabinasab and Tobias Hyrup and Arthur Zimek and Peter Schneider-Kamp},
year={2025},
month=jul,
eprint={2507.19700},
archivePrefix={arXiv},
}
For our purposes we had to modify a few parts of the code. The main changes are to the preprocessing to support our datasets, and to the training loop to cut out unnessessary evaluation steps. We also added a few flags to the CLI to support these changes.
See the codebook for the experimental setup:
| Link | Description |
|---|---|
| TabDiff Baseline | Codebook with the commands used. |
The analysis of these results and others is found in the project repository: Disjoint Generative Models
The remainder of this readme is from the original TabDiff codebase, and is included here for propper attribution. If you use this codebase for anything other than the experiments in our paper, please refer to the original TabDiff codebase and cite the original paper:
This work is licensed undeer the MIT License.
This repo is built upon the previous work TabSyn's [codebase]. Many thanks to Hengrui!
Please consider citing our work if you find it helpful in your research!
@inproceedings{shi2025tabdiff, title={TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation}, author={Juntong Shi and Minkai Xu and Harper Hua and Hengrui Zhang and Stefano Ermon and Jure Leskovec}, booktitle={The Thirteenth International Conference on Learning Representations}, year={2025}, url={https://openreview.net/forum?id=swvURjrt8z} }If you encounter any problem, please file an issue on this GitHub repo.
If you have any question regarding the paper, please contact Minkai at minkai@stanford.edu or Juntong at shisteve@usc.edu.