PosterCopilot: Toward Layout Reasoning and Controllable Editing for Professional Graphic Design

Jiazhe Wei^1,*, Ken Li^1,*, Tianyu Lao², Haofan Wang², Liang Wang^1,3, Caifeng Shan¹, Chenyang Si^1,†

¹PRLab, Nanjing University, ²LibLib.ai, ³Institute of Automation, Chinese Academy of Sciences

^*Equal Contribution, ^†Corresponding Author

📄 Paper | 🌐 Project Page | ▶️ Video | 🤗 Model Weights (Coming Soon) | 🤗 Datasets (Coming Soon)

---

🔥 News

• [2025-12-04] Our paper is now available on arXiv!

---

🌟 Highlights

PosterCopilot is a cutting-edge framework that advances layout reasoning and controllable editing for professional graphic design using Large Multimodal Models (LMMs).

✨ Core Features

• 🎯 Geometrically Accurate Layouts
  Achieves precise spatial positioning through a progressive three-stage training strategy that moves beyond simple regression to distribution-based learning

• 🎨 Aesthetic Reasoning
  Instills human-like design principles and aesthetics through reinforcement learning from aesthetic feedback

• ✂️ Layer-level Control
  Enables precise, fine-grained editing of individual layers while maintaining global visual consistency

• 🔄 Multi-round Iterative Editing
  Supports professional iterative design workflows with multiple refinement rounds on specific elements

• 🎭 Versatile Applications
  Handles complete layout generation, insufficient assets synthesis, theme switching, and canvas reframing

📈 Three-Stage Training Paradigm

1. Perturbed Supervised Fine-Tuning (PSFT)
   Reformulates coordinate regression into distribution-based learning for continuous spatial reasoning

2. Reinforcement Learning for Visual-Reality Alignment (RL-VRA)
   Introduces geometric reward signals to ensure visual-reality alignment and spatial accuracy

3. Reinforcement Learning from Aesthetic Feedback (RLAF)
   Employs learned aesthetic rewards to generate coherent and diverse compositions

📊 PosterCopilot Dataset

One of the largest-scale, most thematically diverse, and highest-quality multi-layer poster datasets.

• 160K posters with 2.6M layers (1.2M text + 1.4M image/decorative elements)
• Spans 40+ distinct domains from commercial promotions to public announcements
• Novel OCR-based pipeline addresses over-segmentation challenges in multi-layer datasets

---

📋 To-Do List

We are committed to making outstanding contributions to both academia and the graphic design industry with PosterCopilot. Our open-source plan includes:

✅ Released

• Project page and documentation
• Demo video

🚧 Coming Soon

• Data Pipeline
• Test Dataset
• Training Code
• Model Weights

---

📝 Citation

If you find PosterCopilot useful for your research, please consider citing:

@misc{wei2025postercopilot,
        title={PosterCopilot: Toward Layout Reasoning and Controllable Editing for Professional Graphic Design}, 
        author={Jiazhe Wei and Ken Li and Tianyu Lao and Haofan Wang and Liang Wang and Caifeng Shan and Chenyang Si},
        year={2025},
        eprint={2512.04082},
        archivePrefix={arXiv},
        primaryClass={cs.CV}
  }

---

📧 Contact

For questions and collaborations, please contact:

• Jiazhe Wei: [email protected]
• Ken Li: [email protected]
• Chenyang Si: [email protected]

---

🙏 Acknowledgments

We thank all contributors and the research community for their valuable feedback and support.

© 2025 PosterCopilot project. Released under the Apache 2.0 License.