File size: 3,727 Bytes
aaced19
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
<div align="center">

# Perceptual Flow Matching for Few-Step Generative Modeling

<p>
  <strong>
    Chuyang Zhao<sup>1</sup>&nbsp; Yifei Song<sup>2</sup>&nbsp; Hongfa Wang<sup>3</sup>&nbsp; Jianlong Yuan<sup>1</sup>&nbsp;
    Yuan Zhang<sup>1</sup>&nbsp; Siming Fu<sup>1</sup>&nbsp; Zhineng Chen<sup>2</sup>&nbsp; Huilin Deng<sup>4</sup>&nbsp; Haoyang Huang<sup>1</sup>&nbsp; Nan Duan<sup>1</sup>†
  </strong>
</p>

<p>
  <sup>1</sup>Joy Future Academy&nbsp;&nbsp;
  <sup>2</sup>Fudan University&nbsp;&nbsp;
  <sup>3</sup>Tsinghua University&nbsp;&nbsp;
  <sup>4</sup>USTC
</p>

<p>
  <a href="https://arxiv.org/abs/2607.03524v1"><img src="https://img.shields.io/badge/arXiv-2607.03524-b31b1b.svg"></a>&nbsp;
  <a href="https://huggingface.co/papers/2607.03524"><img src="https://img.shields.io/badge/🤗-Paper%20In%20HF-red.svg" alt="Hugging Face"></a>&nbsp;
  <a href="https://github.com/ZhaoChuyang/PFM"><img src="https://img.shields.io/badge/GitHub-Code-black?logo=github"></a>
</p>

</div>

## 📰 News
- [2026/07/23] We release the training and inference code and checkpoints of PFM on SD3.

## Introduction

We propose Perceptual Flow Matching (PFM) — a simple framework for few-step generation in flow-matching models. By supervising flow matching in a perceptual feature space instead of the conventional VAE latent space, PFM reduces sampling steps from 35–50 to 4–8 while preserving generation quality.

<p align="center">
  <img src="assets/introduction.png" width="90%" alt="Perceptual Flow Matching Method Overview">
</p>


## Usage

### Environment Setup

```bash
conda create -n pfm python=3.10 -y
conda activate pfm

pip install -r requirements.txt
```

### Training

| Task | Dataset | LoRA | Perceptual Model | Backbone | GFT Scale | Script |
|------|---------|------|-----------------|----------|-----------|--------|
| T2I | COCO | ✗ | VGG+DINO | SD3-Medium | 1.0 | [train_sd3_coco_vgg+dino.sh](scripts/train_sd3_coco_vgg+dino.sh) |
| T2I | COCO | ✓ | VGG+DINO | SD3-Medium | 2.0 | [train_sd3_lora_coco_vgg+dino.sh](scripts/train_sd3_lora_coco_vgg+dino.sh) |
| T2I | COCO | ✗ | ConvNeXt+CLIP+DINO | SD3-Medium | 3.5 | [train_sd3_coco_clip+dino+convnext.sh](scripts/train_sd3_coco_clip+dino+convnext.sh) |


### Inference

```bash
cd /path/to/PFM
export PYTHONPATH=$PYTHONPATH:$(pwd)

torchrun --nproc_per_node=8 pfm/eval_sd3.py \
    --checkpoint /path/to/checkpoint/step_500.pth \
    --val_prompts_file evaluations/PartiPrompts.jsonl \
    --sampling_methods consistency \
    --num_steps 4,8 \
    --output_dir outputs/sd3_pfm/eval \
    --max_samples 64
```

## Acknowledgements

This project is built upon [HunyuanVideo](https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5), [JoyAI-Image](https://github.com/jd-opensource/JoyAI-Image), and [FastVideo](https://github.com/hao-ai-lab/fastvideo). We thank the authors for their great work.

## Contact

For any questions, please contact [Chuyang Zhao](mailto:chuyang.zhao@outlook.com)(chuyang.zhao@outlook.com).

We are hiring interns at [JoyAI Research](https://research.joyai.com/). If you are interested, please send your resume to [zhaochuyang.3@jd.com](mailto:zhaochuyang.3@jd.com).

## Citation

```bibtex
@misc{zhao2026perceptualflowmatchingfewstep,
      title={Perceptual Flow Matching for Few-Step Generative Modeling}, 
      author={Chuyang Zhao and Yifei Song and Hongfa Wang and Jianlong Yuan and Yuan Zhang and Siming Fu and Zhineng Chen and Huilin Deng and Haoyang Huang and Nan Duan},
      year={2026},
      eprint={2607.03524},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.03524}, 
}
```

## License

This project is released under the [Apache-2.0](LICENSE) license.