DCoAR

Deep Concept Injection into Autoregressive Models for Personalized Text-to-Image Generation

---

🔍 Overview

DCoAR is a simple yet effective framework for injecting subject or style concepts into multi-modal autoregressive (AR) models without modifying any pretrained parameters. By leveraging a small set of layer-wise multi-modal learnable tokens, DCoAR achieves high-quality, efficient customization for both subject-driven personalization and training-free subject-style compositional generation.

---

📝 Citation

@misc{wu2025dcoardeepconceptinjection,
      title={DCoAR: Deep Concept Injection into Unified Autoregressive Models for Personalized Text-to-Image Generation}, 
      author={Fangtai Wu and Mushui Liu and Weijie He and Zhao Wang and Yunlong Yu},
      year={2025},
      eprint={2508.07341},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.07341}, 
}