SceneDecorator: Towards Scene-Oriented Story Generation with Scene Planning and Scene Consistency

Quanjian Song^1,*,   Donghao Zhou^2,*†,   Jingyu Lin^1,*†,   Fei Shen³,  
Jiaze Wang²,   Xiaowei Hu^4,‡,   Cunjian Chen^1,‡,   Pheng-Ann Heng²

¹Monash University   ²The Chinese University of Hong Kong
³National University of Singapore   ⁴South China University of Technology
_{^*Equal contribution.   ^†Project lead.   ^‡Corresponding authors.}

           

🎉 News

• 2025.10: 🔥 Our paper, code, and project page are released.
• 2025.09: 🔥 SceneDecorator has been accepted by NeurIPS 2025.

🎬 Overview

In this work, we design a training-free framework called SceneDecorator, to address two key challenges in story generation: scene planning and scene consistency. SceneDecorator comprises two core techniques: (i) VLM-Guided Scene Planning. Leveraging a powerful Vision-Language Model (VLM) as a director, it decomposes user-provided themes into local scenes and story sub-prompts in a ''global-to-local'' manner. (ii) Long-Term Scene-Sharing Attention. By simultaneously integrating mask-guided scene injection, scene-sharing attention, and extrapolable noise blending, it maintains subject style diversity and long-term scene consistency in story generation. Overall framework is shown below:

🔧 Environment

git clone https://github.com/lulupig12138/SceneDecorator.git
# Installation with the requirement.txt
conda create -n SceneDecorator python=3.10
conda activate SceneDecorator
pip install -r requirements.txt
# Or installation with environment.yaml
conda env create -f environment.yml

🚀 Start

bash start.sh

🎓 Bibtex

🤗 If you find this code helpful for your research, please cite:

@article{song2026scenedecorator,
  title={Scenedecorator: Towards scene-oriented story generation with scene planning and scene consistency},
  author={Song, Quanjian and Zhou, Donghao and Lin, Jingyu and Shen, Fei and Wang, Jiaze and Hu, Xiaowei and Chen, Cunjian and Heng, Pheng-Ann},
  journal={Advances in Neural Information Processing Systems},
  year={2026}
}