LoongX

LoongX: Neural-Driven Image Editing

🌐 Homepage ｜ 📖 Paper ｜ 🤗 Data ｜ 🚀 Quickstart ｜ 🖊️ Citation

This repository is the official implementation of Neural-Driven Image Editing (NeurIPS 2025).

💡 News

• 2026/02/09: The L-Mind dataset is released at Hugging Face.
• 2025/07/08: The homepage is launched and will be continuously updated at LoongX.
• 2025/07/07: The homepage, dataset and demo will be released one by one. Stay tuned!
• 2025/07/06: The paper is uploaded to arXiv.

📖 Introduction

We introduce LoongX, which effectively integrates multimodal neural signals to guide image editing through novel Cross-Scale State Space (CS3) encoder and Dynamic Gated Fusion (DGF) modules. Experimental results demonstrate that editing performance using only multimodal neural signals is comparable to text-driven baselines (CLIP-I: 0.6605 vs. 0.6558), and the combination of neural signals with speech instructions surpasses text prompts alone (CLIP-T: 0.2588 vs. 0.2549), proving the method's effectiveness and its significant potential in intuitive and inclusive human-AI interaction.

🚀 Quick Start

Setup (Optional)

1. Environment setup

conda create -n loongx python=3.10
conda activate loongx

2. Requirements installation

pip install -r requirements.txt

Training

Training instructions can be found in this guide.

Inference

For inference, please refer to inference.sh. You need to modify the input and output paths in the script.

Evaluation

For evaluation, please refer to test.sh. You need to modify the input and output paths in the script.

🖊️ Citation

@inproceedings{zhouneural,
  title={Neural-Driven Image Editing},
  author={Zhou, Pengfei and Xia, Jie and Peng, Xiaopeng and Zhao, Wangbo and Ye, Zilong and Li, Zekai and Yang, Suorong and Pan, Jiadong and Chen, Yuanxiang and Wang, Ziqiao and others},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems}
}