Concept Corrector

This is the official code for the paper "Concept Corrector: Erase concepts on the fly for text-to-image diffusion models" (arXiv).

Timelines

🔥 [2025-08-23] In PRCV 2025, two reviewers recommend "strong accept" and three reviewers recommend "weak accept"!

🔥 [2025-08-23] Our paper has been accepted by PRCV 2025!

🔥 [2025-09-06] Our code has been released!

🔥 [2025-09-21] Our paper has been accepted as Oral in PRCV 2025!

Run it!

1. Download the repo.

git clone https://github.com/RichardSunnyMeng/ConceptCorrector.git
cd ConceptCorrector

2. Install the environments

conda create -n concept_corrector python==3.10
conda activate concept_corrector
pip install --upgrade pip
pip install -r requirements.txt

3. Install VLM requirements.

In this paper, we use LLaVa-OneVision-Qwen2-7B as the checker. Of course, you can use other VLMs, e.g., Qwen-VL.

To install requirements for LLaVa-OneVision-Qwen2-7B, please refer to this site.

4. Generate images. Using our official configs configs.yaml, you can directly run the script run.sh.

sh run.sh

Acknowledgments

• Diffusers: Make diffusion models easy to use!
• Stable Diffusion: Outstanding text-to-image performance!
• LLaVa-NeXt: Outstanding multimodal capability!

Citation

@article{meng2025concept,
  title={Concept corrector: Erase concepts on the fly for text-to-image diffusion models},
  author={Meng, Zheling and Peng, Bo and Jin, Xiaochuan and Lyu, Yueming and Wang, Wei and Dong, Jing and Tan, Tieniu},
  journal={arXiv preprint arXiv:2502.16368},
  year={2025}
}