EditIDv2: Editable ID Customization with Data-Lubricated ID Feature Integration for Text-to-Image Generation

Guandong Li, Zhaobin Chu
iFlyTek Typemovie Research

EditID: Training-Free Editable ID Customization for Text-to-Image Generation

Guandong Li, Zhaobin Chu
iFlyTek Typemovie Research

🚩 Updates

• 2025.12.29 🎉 EditIDv2 paper accepted by Multimedia Systems, exciting news!
• 2025.09.09 🔥 EditIDv2 technical report officially published on arXiv.
• 2025.09.09 🔥 EditIDv2 project code officially open-sourced, GitHub repository now available, welcoming community feedback and engagement!
• 2025.08.21 🎉 EditID paper accepted by EMNLP 2025, exciting news!
• 2025.03.16 🔥 EditID paper published on arXiv.

We will continue to open-source more resources, including training code, model weights, and datasets. Stay tuned! 🌟

📖 Abstract

We propose EditIDv2, a tuning-free solution specifically designed for high-complexity narrative scenes and long text inputs. Existing character editing methods perform well under simple prompts, but often suffer from degraded editing capabilities, semantic understanding biases, and identity consistency breakdowns when faced with long text narratives containing multiple semantic layers, temporal logic, and complex contextual relationships. In EditID, we analyzed the impact of the ID integration module on editability. In EditIDv2, we further explore and address the influence of the ID feature integration module. The core of EditIDv2 is to discuss the issue of editability injection under minimal data lubrication. Through a sophisticated decomposition of PerceiverAttention, the introduction of ID loss and joint dynamic training with the diffusion model, as well as an offline fusion strategy for the integration module, we achieve deep, multi-level semantic editing while maintaining identity consistency in complex narrative environments using only a small amount of data lubrication. This meets the demands of long prompts and high-quality image generation, and achieves excellent results in the IBench evaluation.

⚡️ Quick Start

🔧 Training

Train EditIDv2 weights using the following command, with the configuration file located at ./train_configs/editid_insert.yaml:

python train_editid_loss.py --config ./train_configs/editid_insert.yaml

After training, you will obtain the EditIDv2 model weights, which can be used for subsequent inference.

🚀 Inference

The inference process for EditIDv2 is fully compatible with PuLID. You can directly reuse PuLID's inference code or ComfyUI workflow by simply replacing it with EditIDv2's trained weights. Here are the quick start steps:

python infer.py --ckpt_path /path/to/editidv2_weights.pth

Alternatively, use PuLID's ComfyUI workflow by loading the EditIDv2 weights. Refer to the PuLID ComfyUI tutorial for guidance.

🌈 More Examples

We provide sample prompts and results to showcase EditIDv2’s capabilities. For additional visualizations, check our paper.

Complex Narrative Scenes

EditIDv2 excels at placing a subject into complex narrative scenes while maintaining identity consistency.

📄 Disclaimer

This project is open-sourced for academic research. Images used are either generated or sourced from public datasets like MyStyle. If you have concerns, please contact us, and we will promptly address any issues. EditIDv2 is released under the Apache 2.0 License. When using other base models, ensure compliance with their licensing terms. This research advances personalized text-to-image generation. Users must comply with local laws and use the tool responsibly. The developers are not liable for misuse.

🚀 Updates

We aim to fully open-source EditIDv2, including training, inference, weights, and dataset, to support the research community. Thank you for your patience! 🌟

• Release technical report.
• Release GitHub repository.
• Release inference code.
• Release model checkpoints.
• Release Hugging Face Space demo.
• Release training code.
• Release dataset.

📜 Citation

If EditIDv2 is helpful, please consider starring the repo. 🌟

For research purposes, cite our paper:

bibtex

@misc{li2025editidv2editableidcustomization,
      title={EditIDv2: Editable ID Customization with Data-Lubricated ID Feature Integration for Text-to-Image Generation}, 
      author={Guandong Li and Zhaobin Chu},
      year={2025},
      eprint={2509.05659},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.05659}, 
}

@article{li2025editid,
  title={EditID: Training-Free Editable ID Customization for Text-to-Image Generation},
  author={Li, Guandong and Chu, Zhaobin},
  journal={arXiv preprint arXiv:2503.12526},
  year={2025}
}