# Differ

**Repository Path**: code_godtao/differ

## Basic Information

- **Project Name**: Differ
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Apache-2.0
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-06-28
- **Last Updated**: 2026-06-28

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# DIFFER: Disentangling Identity Features via Semantic Cues for Clothes-Changing Person Re-ID

This repository contains PyTorch implementation for CVPR 2025 paper: **DIFFER: Disentangling Identity Features via Semantic Cues for Clothes-Changing Person Re-ID**.  

[🔗 Paper Link (arXiv)](https://arxiv.org/abs/2503.22912)

![Model Architecture](figure/differ_method.png)


## Method Overview

DIFFER leverages text prompts to disentangle biometric and non-biometric factors (e.g., clothing, pose, hairstyle) in person re-identification under clothing changes. See our paper for full details.



## Installation

```bash
# Install Python 3.11 (if not already installed)
# Then install dependencies
pip install torch>=2.0
pip install -r requirements.txt
````


## Training & Evaluation

1. Generate or download the **textual features** ([link](https://drive.google.com/drive/folders/171pES67flGW-DCIXyQ2VB7iU3BTUqhj5?usp=sharing)), and place them in a directory with the following structure:

```
TextCaptionDirectory/
├── Dataset1/
│   ├── TextEncoderVersion/
│   │   └── train.npz
│   └── train_caption_summary_biometric.json
├── Dataset2/
│   └── ...
```

* `train_caption_summary_biometric.json` contains all the summarized biometric information and its corresponding textual encodings.

* `train.npz` contains all the text encodings from different perspectives. The perspectives are encoded with the following numeric mapping:   0 (Biometrics), 1 (Hairstyle), 2(Clothing), 3 (Pose), 4 (Interaction), 5 (Environment).  

2, To train the model:
```bash
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port 1234 train.py  \
--jobId 0 \
--loss "ce,triplet,clipBio,clipBioReverse"  \
--config_file 'configs/ltcc/eva02_l_bio.yml' \
MODEL.DIST_TRAIN True \
OUTPUT_NAME "DIFFER" \
DATA.CAPTION_DIR "/TextCaptionDirectory/LTCC_ReID" \
DATA.ROOT "/DatasetsDirectory"
```
Our trained weight can be found [here](https://drive.google.com/drive/folders/1RzAhSeSOgL2u8130mFAMdHI2k4sX9dQD?usp=sharing) .


> **Reproducibility Note**  
> This project uses adversarial training, which can be inherently unstable.  
> Although we fix the random seed, final results may still vary across runs with the same configuration due to FP16 training and inherent training dynamics.     
> To reproduce the results reported in our paper, we recommend running the training multiple times.

3, To evaluate the model,

```bash
CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --nproc_per_node=1 --master_port 6673 test.py \
--config_file 'MODELDIR/config.yml' \
TEST.WEIGHT 'MODELDIR/eva02_l_bio_best.pth' 
```


## Code Acknowledgment

Our implementation is based on [MADE](https://github.com/moon-wh/MADE.git). We reuse parts of their code and build upon it for our method. We thank the authors for their excellent work.

## Citation

If you find this work helpful, please cite:

```bibtex
@InProceedings{Liang_2025_CVPR,
    author    = {Liang, Xin and Rawat, Yogesh S},
    title     = {DIFFER: Disentangling Identity Features via Semantic Cues for Clothes-Changing Person Re-ID},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {13980-13989}
}
```