Human Body Restoration with
One-Step Diffusion Model and A New Benchmark

🚩 Accepted by ICML 2025

Jue Gong, Jingkai Wang, Zheng Chen, Xing Liu, Hong Gu, Yulun Zhang, Xiaokang Yang

"A new benchmark and the first one-step diffusion model for human body restoration.", 2025

🔥🔥🔥 News

• 2025-02-05: The OSDHuman repository was released.
• 2025-05-01: OSDHuman was accepted to ICML 2025.
• 2025-05-04: Released the PERSONA test and validation sets.
• 2026-04-18: Released the public inference code and inference instructions.

---

Abstract: Human body restoration, as a specific application of image restoration, is widely applied in practice and plays a vital role across diverse fields. However, thorough research remains difficult, particularly due to the lack of benchmark datasets. In this study, we propose a high-quality dataset automated cropping and filtering (HQ-ACF) pipeline. This pipeline leverages existing object detection datasets and other unlabeled images to automatically crop and filter high-quality human images. Using this pipeline, we constructed a person-based restoration with sophisticated objects and natural activities (PERSONA) dataset, which includes training, validation, and test sets. The dataset significantly surpasses other human-related datasets in both quality and content richness. Finally, we propose OSDHuman, a novel one-step diffusion model for human body restoration. Specifically, we propose a high-fidelity image embedder (HFIE) as the prompt generator to better guide the model with low-quality human image information, effectively avoiding misleading prompts. Experimental results show that OSDHuman outperforms existing methods in both visual quality and quantitative metrics.

---

---

⚒️ TODO

• Release inference code
• Release PERSONA test and validation sets
• Release training code
• Release PERSONA training set

🔗 Contents

• Inference
• PERSONA Test and Validation Set
• Results
• Citation
• Acknowledgements

🚀 Inference

This repository provides the public OSDHuman inference code. The entry point is run_inference.py, and sample inputs are included under test_set/.

Environment Installation

conda create -n osdhuman python=3.10 -y
conda activate osdhuman
pip install --upgrade pip
pip install -r requirements.txt

Prepare Checkpoints

Create local runtime folders if needed:

mkdir -p weights preset outputs

Place the following OSDHuman checkpoints under weights/:

• weights/model.pkl
• weights/model_embedding_change.pkl

Download source:

• Google Drive folder: https://drive.google.com/drive/folders/1sNoSNcJC7cdf0RR5dK87PVTEupjOZ-3B?usp=drive_link

Place the following external base models locally:

• preset/stable-diffusion-2-1-base Source: https://huggingface.co/stabilityai/stable-diffusion-2-1-base
• preset/ram_swin_large_14m.pth Source: https://huggingface.co/spaces/xinyu1205/recognize-anything/blob/main/ram_swin_large_14m.pth

Quick Start

Run inference on the bundled sample set:

python run_inference.py \
  -i test_set \
  -o outputs \
  --osediff_path weights/model.pkl \
  --pretrained_model_name_or_path preset/stable-diffusion-2-1-base \
  --ram_path preset/ram_swin_large_14m.pth \
  --ram_img_only

run_inference.py accepts either a single image path or a directory and writes restored images into outputs/ by default.

🎭 PERSONA Test and Validation Set

We provide three sets for evaluation, including both high-quality and degraded validation images and a real-world test set.

Dataset	Description	Download Link
PERSONA-Val HQ	High-quality validation set	Google Drive
PERSONA-Val LQ	Low-quality validation set	Google Drive
PERSONA-Test	Real-world test set	Google Drive

🔎 Results

The model OSDHuman achieved state-of-the-art performance on both PERSONA-Val and PERSONA-Test. Detailed comparisons are available in the paper.

 Quantitative Comparisons (click to expand)

Results in Table 2 on synthetic PERSONA-Val and real-world PERSONA-Test datasets from the main paper.

 Visual Comparisons (click to expand)

Results in Figure 8 on the real-world PERSONA-Test dataset from the main paper.

Results in Figure 9 on the synthetic PERSONA-Val dataset from the main paper.

 More Comparisons on real-world PERSONA-Test dataset...

Results in Figures 3 and 4 from the supplementary material.

 More Comparisons on synthetic PERSONA-Val dataset...

Results in Figures 5 and 6 from the supplementary material.

📎 Citation

If you find the code helpful in your research or work, please cite:

@inproceedings{gong2025osdhuman,
    title={Human Body Restoration with One-Step Diffusion Model and A New Benchmark},
    author={Gong, Jue and Wang, Jingkai and Chen, Zheng and Liu, Xing and Gu, Hong and Zhang, Yulun and Yang, Xiaokang},
    booktitle={ICML},
    year={2025}
}

💡 Acknowledgements

OSEDiff