Hugging Face's logo

Datasets:
WithAnyone
/
MultiID-Bench

Tasks:
Text-to-Image
Modalities:
Image
Text
Formats:
parquet
Languages:
English
Size:
< 1K
ArXiv:
Tags:
face-generation
identity-consistent
multi-person
benchmark
Libraries:
Datasets
pandas
License:
Dataset card Data Studio Files
xet
Community
2
MultiID-Bench / README.md
nielsr's picture
nielsr HF Staff
Improve dataset card for MultiID-Bench: Add task categories, tags, intro, usage, and full license
f6e5bab verified
preview code
|
raw
history blame
8.73 kB
metadata
license: other
license_name: multiid-2m
license_link: LICENSE.md
language:
  - en
task_categories:
  - text-to-image
tags:
  - face-generation
  - identity-consistent
  - multi-person
  - benchmark
dataset_info:
  features:
    - name: ID
      dtype: string
    - name: GT
      dtype: image
    - name: input_images
      sequence: image
    - name: prompt
      dtype: string
    - name: task_type
      dtype: string
    - name: bboxes
      dtype: string
    - name: subset
      dtype: string
    - name: num_persons
      dtype: int32
  splits:
    - name: train
      num_bytes: 194376894
      num_examples: 433
  download_size: 183022320
  dataset_size: 194376894
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

MultiID-Bench in WithAnyone

arXiv Project Page Code HuggingFace MultiID-Bench MultiID-2M Demo

The MultiID-Bench dataset is a benchmark introduced in the paper WithAnyone: Towards Controllable and ID Consistent Image Generation. It is specifically tailored for multi-person scenarios in text-to-image research, providing diverse references for each identity. This dataset aims to quantify "copy-paste" artifacts and evaluate the trade-off between identity fidelity and variation, enabling models like WithAnyone to achieve controllable and identity-consistent image generation.

Links

Sample Usage

This section provides instructions for downloading the MultiID-Bench dataset and preparing it for evaluation.

Download the Dataset

You can download the MultiID-Bench dataset using the Hugging Face CLI:

huggingface-cli download WithAnyone/MultiID-Bench --repo-type dataset --local-dir <path to MultiID-Bench directory>

Prepare Data for Evaluation

After downloading, if your dataset is in a parquet file format, you can convert it into a structured directory of images and JSON metadata using the parquet2bench.py script provided in the GitHub repository.

First, ensure you have cloned the GitHub repository:

git clone https://github.com/Doby-Xu/WithAnyone
cd WithAnyone

Then, convert the downloaded parquet file:

python MultiID_Bench/parquet2bench.py --parquet <path to downloaded parquet file> --output_dir <root directory to save the processed data>

The output directory will contain a structure like this, with subfolders for each ID and meta.json files containing prompts:

root/
├── id1/
│   ├── out.jpg
│   ├── ori.jpg
│   ├── ref_1.jpg
│   ├── ref_2.jpg
│   ├── ref_3.jpg
│   ├── ref_4.jpg
│   └── meta.json
│
├── id2/
│   ├── out.jpg
│   ├── ori.jpg
│   ├── ref_1.jpg
│   ├── ref_2.jpg
│   ├── ref_3.jpg
│   ├── ref_4.jpg
│   └── meta.json
│
└── ...

The meta.json file should contain the prompt used to generate the image, in the following format:

{
    "prompt": "a photo of a person with blue hair and glasses"
}

Environment Setup for Evaluation

To run the evaluation scripts, you need to install several packages. Besides the requirements.txt from the GitHub repo, install the following:

pip install aesthetic-predictor-v2-5 
pip install facexlib
pip install colorama
pip install pytorch_lightning
git clone https://github.com/timesler/facenet-pytorch.git facenet_pytorch

# in MultiID_Bench/
mkdir pretrained

You will also need the following models to run the evaluation: CLIP, arcface, aesthetic-v2.5, adaface, and facenet. The first three will be automatically downloaded. For adaface, download adaface_ir50_ms1mv2.ckpt from this link and place it in the pretrained directory.

Run Evaluation

You can run the evaluation script as follows, using the prepared data:

from eval import BenchEval_Geo

def run():
    evaler = BenchEval_Geo(
        target_dir="<root directory mentioned above>",
        output_dir="<output directory to save the evaluation results>",
        ori_file_name="ori.jpg", # the name of the ground truth image file
        output_file_name="out.jpg", # the name of the generated image file
        ref_1_file_name="ref_1.jpg", # the name of the first reference image file
        ref_2_file_name="ref_2.jpg", # the name of the second reference image file
        # ref_2_file_name=None, # if you only have one reference image, set ref_2_file_name to None
        # ref_3_file_name="ref_3.jpg", # the name of the third reference
        # ref_4_file_name="ref_4.jpg", # the name of the fourth reference,
        caption_keyword="prompt", # the keyword to extract the prompt from meta.json
        names_keyword=None
    )
    evaler()
if __name__ == "__main__":
    run()

License and Disclaimer

The code of WithAnyone is released under the Apache License 2.0, while the WithAnyone model and associated datasets are made available solely for non-commercial academic research purposes.

  • License Terms:
    The WithAnyone model is distributed under the FLUX.1 [dev] Non-Commercial License v1.1.1. All underlying base models remain governed by their respective original licenses and terms, which shall continue to apply in full. Users must comply with all such applicable licenses when using this project.

  • Permitted Use:
    This project may be used for lawful academic research, analysis, and non-commercial experimentation only. Any form of commercial use, redistribution for profit, or application that violates applicable laws, regulations, or ethical standards is strictly prohibited.

  • User Obligations:
    Users are solely responsible for ensuring that their use of the model and dataset complies with all relevant laws, regulations, institutional review policies, and third-party license terms.

  • Disclaimer of Liability:
    The authors, developers, and contributors make no warranties, express or implied, regarding the accuracy, reliability, or fitness of this project for any particular purpose. They shall not be held liable for any damages, losses, or legal claims arising from the use or misuse of this project, including but not limited to violations of law or ethical standards by end users.

  • Acceptance of Terms:
    By downloading, accessing, or using this project, you acknowledge and agree to be bound by the applicable license terms and legal requirements, and you assume full responsibility for all consequences resulting from your use.

Acknowledgement

We thank the following prior art for their excellent open source work:

Citation

If you find this project useful in your research, please consider citing:

@article{xu2025withanyone,
  title={WithAnyone: Towards Controllable and ID-Consistent Image Generation}, 
  author={Hengyuan Xu and Wei Cheng and Peng Xing and Yixiao Fang and Shuhan Wu and Rui Wang and Xianfang Zeng and Gang Yu and Xinjun Ma and Yu-Gang Jiang},
  journal={arXiv preprint arxiv:2510.14975},
  year={2025}
}