A solvable generative model with a linear, one-step denoiser

This repository contains the code necessary to reproduce the results obtained by Indranil Halder in his recent paper - available here.

Experiments

The experiments on the linear diffusion model are run using LDM.py. The U-Net-based non-linear diffusion model with attention layers is defined in model.py, and modules.py. Simulations with it can be performed using simulation.py.

Dependencies

We use Python 3.10.12, and dependencies with their exact version numbers listed in environment.yml.

Citing this work

@misc{halder2025solvablegenerativemodellinear,
      title={A solvable generative model with a linear, one-step denoiser}, 
      author={Indranil Halder},
      year={2025},
      eprint={2411.17807},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2411.17807}, 
}