Skip to content

Design of noise, interpolation schedules, and sources in generative dynamics, with enhanced numerical performance

Notifications You must be signed in to change notification settings

yifanc96/GenerativeDynamics-NumericalDesign

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

73 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

GenerativeDynamics - NumericalDesign

Design of interpolation schedules, noise, and source terms in flow and diffusion-based generative models, with a focus on data distributions from science and engineering applications.

The Git Repository includes examples based on solutions to stochastic Allen-Cahn and Navier-Stokes equations. These examples involve training neural networks on datasets with wide-ranging Fourier spectra, making them numerically ill-conditioned. The neural network training codes require GPU computation, and the repository comes with provided datasets. Note that these codes will need modification to run on your system (specifically, updating data paths, save locations, and wandb accounts). The directory is not fully cleaned, so there might be multiple files with slightly different setups, but the principle remains the same.

We also present proof-of-concept examples using Gaussian measures and mixtures in high dimensions. These examples are based on explicit formulas and can be run directly on CPUs without modification.

All implementations adopt the stochastic interpolants framework.

Relevant papers

@article{chen2025lipschitz,
  title={Lipschitz-Guided Design of Interpolation Schedules in Generative Models},
  author={Chen, Yifan and Vanden-Eijnden, Eric and Xu, Jiawei},
  journal={arXiv preprint arXiv:2509.01629},
  year={2025}
}

@article{chen2025scale,
  title={Scale-Adaptive Generative Flows for Multiscale Scientific Data},
  author={Chen, Yifan and Vanden-Eijnden, Eric},
  journal={arXiv preprint arXiv:2509.02971},
  year={2025}
}

Dataset

Other helpful repository

For general codes of stochastic interpolants, please refer to interpolant. Specifically for probabilistic forecasting applications, please refer to forecasting. There, the point source is the current state, and the generative model forecasts the next state. In this repository, we also have the case where the point source is always zero, and the generative model forecasts the difference between the next and current states.

I'll thank Mark Goldstein for helping me in the initial phase of the forecasting code, and Mengjian Hua for preparing the stochastic Navier-Stokes dataset.

About

Design of noise, interpolation schedules, and sources in generative dynamics, with enhanced numerical performance

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published