---
license: mit
tags:
  - gene-expression
  - histopathology
  - rectified-flow
  - spatial-transcriptomics
  - image-generation
---

# GeneFlow Checkpoints

Checkpoints for GeneFlow, a rectified-flow model that generates H&E histology
image patches conditioned on gene expression (Xenium spatial transcriptomics).
Companion to the NeurIPS 2025 paper.

Code: https://github.com/wangmengbo/GeneFlow

## Checkpoints

| Folder | Sample | Model type | Epoch | Val loss | SSIM | FID | UNI2-h FID |
|---|---|---|---|---|---|---|---|
| `P1_single/` | XeniumP1 | single-cell | 80 | 0.0300 | 0.1991 | 12.25 | 25.44 |
| `C1_single/` | XeniumC1 | single-cell | 42 | 0.0238 | 0.2584 | 12.40 | 26.32 |
| `C1_multi/`  | XeniumC1 | multi-cell (patch) | 80 | 0.0330 | 0.2884 | 39.21 | 91.06 |

All three checkpoints match or exceed the metrics reported in the paper for
their corresponding configuration (Table 2 / Table 5).

## Usage

Load with the training/inference code in the GitHub repo
(`rectified/rectified_main.py --only_inference` or `rectified/rectified_evaluate.py`),
pointing `--model_path` at the downloaded `.pt` file.