junwann/CSFM-ImageNet1K-Caption

---
license: mit
task_categories:
- text-to-image
language:
- en
pretty_name: CSFM-ImageNet1K-Caption
size_categories:
- 1M<n<10M
---

### CSFM-ImageNet1K-Caption Dataset

[Project Page](https://junwankimm.github.io/CSFM) | [Paper](https://arxiv.org/abs/2602.05951) | [Code](https://github.com/junwankimm/CSFM)

This repository contains dataset associated with the paper "Better Source Better Flow: Learning Condition-Dependent Source Distribution for Flow Matching".

This dataset is used for training and evaluating **Condition-dependent Source Flow Matching (CSFM)**, a framework that learns condition-dependent source distributions for flow matching. We recaptioned the ImageNet-1K dataset using **Qwen3-VL-8B Instruct**, resulting in detailed and descriptive image captions. We hope this dataset will facilitate more systematic and quantitative evaluation in text-to-image generation, where such evaluations have been relatively limited.

Due to licensing restrictions, the images are not included in this release. Image files can be obtained separately from the [Official ImageNet Website](https://www.image-net.org/download.php).

### Citation

If you find this work useful, please cite:

```
@misc{kim2026bettersourcebetterflow,
title={Better Source, Better Flow: Learning Condition-Dependent Source Distribution for Flow Matching},
author={Junwan Kim and Jiho Park and Seonghu Jeon and Seungryong Kim},
year={2026},
eprint={2602.05951},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2602.05951},
}
```