---
license: mit
tags:
  - image-generation
  - diffusion
  - flow-matching
  - imagenet
  - sit
datasets:
  - imagenet-1k
pipeline_tag: unconditional-image-generation
library_name: pytorch
---

# Stable Velocity: A Variance Perspective on Flow Matching

[![arXiv](https://img.shields.io/badge/arXiv-2602.05435-b31b1b.svg)](https://arxiv.org/abs/2602.05435)
[![Project Page](https://img.shields.io/badge/Project-Page-blue)](https://linydthu.github.io/StableVelocity/)
[![GitHub](https://img.shields.io/badge/GitHub-Code-black?logo=github)](https://github.com/linYDTHU/StableVelocity)

## Model Description

This repository contains **SiT-XL/2** checkpoints trained with **Stable Velocity Matching (StableVM)** and **Variance-Aware Representation Alignment (VA-REPA)** on **ImageNet 256×256**, as described in the paper.

## Checkpoints

> Class-conditional generation on **ImageNet 256×256** with classifier-free guidance (CFG).

| Checkpoint | Steps | FID ↓ |
|:---|:---:|:---:|
| [`2000000.pt`](checkpoints/2000000.pt) | 2.0 M | 1.34 |
| [`2400000.pt`](checkpoints/2400000.pt) | 2.4 M | 1.33 |
| [`3000000.pt`](checkpoints/3000000.pt) | 3.0 M | **1.30** |


## Download Checkpoints

```python
from huggingface_hub import hf_hub_download

# Download a specific checkpoint
ckpt_path = hf_hub_download(
    repo_id="linYD0718/stable-velocity",
    filename="checkpoints/3000000.pt",
)
```

Or via CLI:

```bash
huggingface-cli download linYD0718/stable-velocity checkpoints/3000000.pt --local-dir ./
```

## Citation

```bibtex
@misc{yang2026stablevelocityvarianceperspective,
      title={Stable Velocity: A Variance Perspective on Flow Matching}, 
      author={Donglin Yang and Yongxing Zhang and Xin Yu and Liang Hou and Xin Tao and Pengfei Wan and Xiaojuan Qi and Renjie Liao},
      year={2026},
      eprint={2602.05435},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.05435}, 
}
```

## License

This project is licensed under the MIT License.