Model Card for EAR

This is an EAR (Erasing Autoregressive Models) model trained to erase specific concepts.

Model Details

Model Description

• Developed by: IMMC
• Model type: AR model
• License: MIT
• Finetuned from model : Janus-Pro

Model Sources

• Repository: [link]
• Paper: [link]

Installation Guide

EAR Environment

git clone https://github.com/immc-lab/ear.git
cd ear
conda create -n ear python=3.12
conda activate ear
pip install -r requirements.txt

Janus-Pro Environment

Ensure that your environment can run Janus-Pro, refer to its official Quick Start for details.

Training Guide

After installation, follow these instructions to train EAR model for Janus-Pro.

Please run the script in train/ after checking the file path:

python train/ear_train_church.py 

Generating Images with EAR

Image generation using the custom EAR model is a straightforward process. Please run the script in infer/.

For automated batch generation of evaluation images, utilize the following script:

python infer/infer_church.py

Evaluation

You can execute the following command to evaluate the generated data. Please run the script in eval/.

The specific evaluation method can be found in our paper.

python eval/eval_object.py  --folder_path {args.output_dir} --topk 10 --batch_size 250

References

This repo is the code for the paper EAR: Erasing Concepts from Unified Autoregressive Models.

Thanks for the creative ideas of the pioneer researches:

• https://github.com/rohitgandikota/erasing: Erasing Concepts from Diffusion Models
• https://github.com/Con6924/SPM: One-dimentional Adapter to Rule Them All: Concepts, Diffusion Models and Erasing Applications
• https://github.com/koushiksrivats/robust-concept-erasing: STEREO: A Two-Stage Framework for Adversarially Robust Concept Erasing from Text-to-Image Diffusion Models
• https://github.com/OPTML-Group/Diffusion-MU-Attack: To Generate or Not? Safety-Driven Unlearned Diffusion Models Are Still Easy To Generate Unsafe Images ... For Now
• https://github.com/deepseek-ai/Janus: Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation
• https://github.com/deepseek-ai/Janus: Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling

Citing our work

The preprint can be cited as follows

@misc{fan2025earerasingconceptsunified,
      title={EAR: Erasing Concepts from Unified Autoregressive Models}, 
      author={Haipeng Fan and Shiyuan Zhang and Baohunesitu and Zihang Guo and Huaiwen Zhang},
      year={2025},
      eprint={2506.20151},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.20151}, 
}