Skip to content

CryoFM1: Overview

CryoFM: A Flow-based Foundation Model for Cryo-EM Densities
The Thirteenth International Conference on Learning Representations (ICLR), 2025.

What is cryoFM?

CryoFM can solve inverse problems without fine-tuning.

CryoFM is the first 3D density map foundation model in the cryo-EM field. It is a base model pretrained on high-resolution density maps using a Hierarchical Diffusion Transformer (HDiT) architecture. Through posterior sampling, cryoFM can be applied to various cryo-EM downstream inverse problems without fine-tuning. We have open-sourced implementations for denoising, anisotropy denoising, and missing wedge in-painting.

CryoFM1 is available in two variants:

  • CryoFM-S: A smaller model optimized for 64×64×64 voxel volumes at 1.5 Å/pixel resolution

CryoFM-S arch.

CryoFM-S architecture.
  • CryoFM-L: A larger model designed for 128×128×128 voxel volumes at 3.0 Å/pixel resolution

CryoFM-L arch.

CryoFM-L architecture.

Get Started

  • Sampling: Learn how to generate density maps using CryoFM1.
  • Downstream Tasks: Explore denoising, anisotropy denoising, and missing wedge in-painting.

Resources

  • Model Weights: Available on Hugging Face.
  • Source Code: Available on GitHub.
  • Dataset (EMDB ID Lists): Available on Zenodo.

Citation

If you use CryoFM1 in your research, please cite:

@inproceedings{
zhou2025cryofm,
title={Cryo{FM}: A Flow-based Foundation Model for Cryo-{EM} Densities},
author={Yi Zhou and Yilai Li and Jing Yuan and Quanquan Gu},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=T4sMzjy7fO}
}