# CellFlow as a training framework vs pretrained model

**URL:** https://discourse.scverse.org/t/cellflow-as-a-training-framework-vs-pretrained-model/3879
**Category:** Help
**Created:** [November 30, 2025, 10:05am UTC](https://discourse.scverse.org/t/cellflow-as-a-training-framework-vs-pretrained-model/3879 "2025-11-30T10:05:30Z")
**Posts on this page:** 2
**Page:** 1

<div class="post-metadata">

### Author: ![tnnandi](https://yyz1.discourse-cdn.com/flex035/user_avatar/discourse.scverse.org/tnnandi/32/1534_2.png) [@tnnandi](https://discourse.scverse.org/u/tnnandi)
#### Post date: [November 30, 2025, 10:05am UTC](https://discourse.scverse.org/t/cellflow-as-a-training-framework-vs-pretrained-model/3879/1 "2025-11-30T10:05:30Z")

</div>

I’m looking for a pretrained model that I can use to generate the perturbed cell state in response to genetic/chemical perturbations, and I came across CellFlow.

I just wanted to confirm whether CellFlow is i) a training framework that I need to use to train models using my dataset(s), or ii) something that offers pretrained models that I can either use directly (zero-shot) or after minimal fine-tuning for predictions on my datasets. Based on my understanding, it is the former, but I’d appreciate it if you could confirm.

---

<div class="post-metadata">

### Author: ![MUCDK](https://yyz1.discourse-cdn.com/flex035/user_avatar/discourse.scverse.org/mucdk/32/1549_2.png) [@MUCDK](https://discourse.scverse.org/u/MUCDK)
#### Post date: [December 5, 2025, 3:17pm UTC](https://discourse.scverse.org/t/cellflow-as-a-training-framework-vs-pretrained-model/3879/2 "2025-12-05T15:17:31Z")

</div>

Hi,

As of now, CellFlow is meant to be retrained for every dataset, but obviously you can also pretrain on more data, and then fine-tune on the dataset of interest.