[ICCV 2025 Highlight] DiffPS: Leveraging Prior Knowledge of Diffusion Model for Person Search

Giyeol Kim^1*   Sooyoung Yang^2*   Jihyong Oh¹   Myungjoo Kang^2,3   Chanho Eom^{† 1}

¹GSAIM, Chung-Ang University  |  ²IPAI, Seoul National University

³Department of Mathematical Sciences and RIMS, Seoul National University

^*Co-first authors (equal contribution)  |  ^†Corresponding author

Project Page  |  Paper (ICCV)  |  arXiv

---

News

• July 24, 2025: DiffPS is selected as a Highlight Paper at ICCV 2025.
• June 24, 2025: DiffPS is accepted to ICCV 2025.

---

Installation

1. Diffusion feature extraction

Environment setup for diffusion features is required. Follow the installation guide in generic-diffusion-feature (NeurIPS'24) to configure your environment. That repository provides the diffusion model setup, dependency versions, and layer interfaces used in this codebase.

Optional: if you clone it next to this repo as generic-diffusion-feature/, models/diffps.py will pick up generic-diffusion-feature/feature automatically. Otherwise set DIFFUSION_FEATURE_PATH to the absolute path of that feature directory.

2. Install dependencies from requirements.txt

pip install -r requirements.txt

3. Install pytorch_wavelets

This project uses pytorch_wavelets for the wavelet transform. Install from the included submodule or from the official repo:

# Option A: install from the included pytorch_wavelets folder
cd pytorch_wavelets && pip install . && cd ..

# Option B: install from official repo
# git clone https://github.com/fbcotter/pytorch_wavelets
# cd pytorch_wavelets && pip install . && cd ..

---

Dataset

Place the dataset under the path specified in the config:

• CUHK-SYSU: set DATASET.PATH in configs/cuhk_sysu.yaml (or configs/_path_cuhk_sysu.yaml).
• PRW: set DATASET.PATH in configs/prw.yaml (or configs/_path_prw.yaml).

Default path in the provided configs is ../dataset/PRW or ../dataset/CUHK-SYSU. Adjust to your directory structure.

For downloading and preparing CUHK-SYSU and PRW, please refer to the dataset instructions in SeqNet.

---

Training

Run training with the provided script:

./train.sh

Or run directly with a config and options:

# PRW dataset
CUDA_VISIBLE_DEVICES=0 python train.py --cfg configs/prw.yaml \
    --opts OUTPUT_DIR "DiffPS" DATASET.BATCH_SIZE 3

# CUHK-SYSU dataset
CUDA_VISIBLE_DEVICES=0 python train.py --cfg configs/cuhk_sysu.yaml \
    --opts OUTPUT_DIR "DiffPS_CUHK" DATASET.BATCH_SIZE 3

You can override any config key via --opts, e.g. OUTPUT_DIR, DATASET.BATCH_SIZE, DATASET.PATH.

---

⚠️ Note on GPU memory (xformers)

There are known issues with certain xformers versions that can cause high GPU memory usage when the diffusion backbone computes cross-attention maps (e.g., for DGRPN). If you run into OOM (out-of-memory) errors:

• Try reducing DATASET.BATCH_SIZE in the config or --opts.
• Check xformers compatibility with your PyTorch and CUDA versions; some builds are known to use significantly more memory for attention. Updating or switching xformers version may help.
• Consider using a GPU with more VRAM if the issue persists.

Tip: Extracting diffusion features and cross attention maps once and saving them locally, then loading from disk during training, is much faster and more memory-efficient than computing them on the fly. We recommend this approach when GPU memory is limited.

Note: We previously experienced an unexpected code loss and have now almost fully recovered the repository. If you have any questions or find something missing, feel free to ask.

---

Citation

If you find this work useful, please cite:

@inproceedings{kim2025leveraging,
  title={Leveraging prior knowledge of diffusion model for person search},
  author={Kim, Giyeol and Yang, Sooyoung and Oh, Jihyong and Kang, Myungjoo and Eom, Chanho},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={20301--20312},
  year={2025}
}

---

References

• SEAS — Person search method. This codebase is built upon SEAS.
• generic-diffusion-feature — Not All Diffusion Model Activations Have Been Evaluated as Discriminative Features (NeurIPS 2024).
• pytorch_wavelets — 2D discrete / dual-tree complex wavelet transforms in PyTorch.