ImageCritic

The Consistency Critic: Correcting Inconsistencies in Generated Images via Reference-Guided Attentive Alignment

This repository hosts ImageCritic, a reference-guided post-editing approach designed to correct inconsistencies in generated images. It aims to solve the inconsistency problem in generated images by applying attention alignment and a detail encoder, providing significant improvements over existing methods in various customized generation scenarios.

The model was presented in the paper The Consistency Critic: Correcting Inconsistencies in Generated Images via Reference-Guided Attentive Alignment.

• 📚 Paper (arXiv)
• 🌐 Project Page
• 💻 Code (GitHub)
• 🤗 Hugging Face Space Demo
• 📦 Hugging Face Dataset

<img src='https://raw.githubusercontent.com/HVision-NKU/ImageCritic/refs/heads/master/figure/teaser.png' width='100%' />

🖼️ Visual Results

ImageCritic can effectively resolve detail-related issues in various customized generation scenarios, providing significant improvements over existing methods.

<img src='https://github.com/HVision-NKU/ImageCritic/blob/master/figure/compare.png?raw=true' width='100%' />

Online HuggingFace Demo

You can try ImageCritic demo on HuggingFace.

Citation

If you find this project useful for your research, please consider citing our paper:

@article{ouyang2025consistency,
  title={The Consistency Critic: Correcting Inconsistencies in Generated Images via Reference-Guided Attentive Alignment},
  author={Ouyang, Ziheng and Song, Yiren and Liu, Yaoli and Zhu, Shihao and Hou, Qibin and Cheng, Ming-Ming and Shou, Mike Zheng},
  journal={arXiv preprint arXiv:2511.20614},
  year={2025}
}

📧 Contact

If you have any comments or questions, please open a new issue or contact Ziheng Ouyang

License

Licensed under a Creative Commons Attribution-NonCommercial 4.0 International for Non-commercial use only. Any commercial use should get formal permission first.