chuyangzhao/PerceptualFlowMatching

Perceptual Flow Matching for Few-Step Generative Modeling

Chuyang Zhao¹  Yifei Song²  Hongfa Wang³  Jianlong Yuan¹  Yuan Zhang¹  Siming Fu¹  Zhineng Chen²  Huilin Deng⁴  Haoyang Huang¹  Nan Duan¹†

¹Joy Future Academy   ²Fudan University   ³Tsinghua University   ⁴USTC

   

📰 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.

Usage

Environment Setup

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
T2I	COCO	✓	VGG+DINO	SD3-Medium	2.0	train_sd3_lora_coco_vgg+dino.sh
T2I	COCO	✗	ConvNeXt+CLIP+DINO	SD3-Medium	3.5	train_sd3_coco_clip+dino+convnext.sh

Inference

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, JoyAI-Image, and FastVideo. We thank the authors for their great work.

Contact

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

We are hiring interns at JoyAI Research. If you are interested, please send your resume to zhaochuyang.3@jd.com.

Citation

@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.