Home

Awesome

<!-- <div align='center'> -->

[ECCV 2024] Receler: Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers

project page | arxiv

Official PyTorch implementation for Receler.

Receler

Receler: Reliable Concept Erasing of Text-to-Image Diffusion Models via Lightweight Erasers

Authors: Chi-Pin Huang*, Kai-Po Chang*, Chung-Ting Tsai, Yung-Hsuan Lai, Fu-En Yang, Yu-Chiang Frank Wang

<!-- </div> -->

Setup

### Install
conda create -n receler python=3.8
conda activate receler
pip install -r requirements.txt

### Get pre-trained weight
wget https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4-full-ema.ckpt -O receler/sd-v1-4-full-ema.ckpt

### (optional) Install GroundingDINO
cd evaluation/
git clone https://github.com/IDEA-Research/GroundingDINO
cd GroundingDINO/  # Then follow their README.md to install.

### (optional) Install nudenet
# Refer to https://github.com/notAI-tech/NudeNet for the installation guideline.

Note: If you cannot import cv2, try uninstalling opencv-python and opencv-python-headless, then reinstall opencv-python-headless again.

Receler

Training

Run the following command to erase concepts from SD v1.4 with train_receler.py:

python train_receler.py --concept $concept

Arguments

Optional Arguments

Arguments for Adversarial Learning

Arguments for Concept-Localized Regularization

Output Structure

The results will be saved in the following structure:

save_root/
  └─ <model_name>/
    ├─ ckpts/              # checkpoints of various iterations
    ├─ visualize/          # visualized images 
    ├─ eraser_config.json
    └─ eraser_weights.pt   # checkpoint of last iteration

Inference & Evaluation

Generating Images

Run the following command to generate images with the pre-trained SD v1.4 or the erased model:

python test_receler.py --model_name_or_path path/to/model/folder/ --prompts_path path/to/prompts

Example:

# Generate images with the pre-erased SD v1.4
python test_receler.py --model_name_or_path CompVis/stable-diffusion-v1-4 --prompts_path evaluation/data/sample_prompts.csv

# Generate images with airplane-erased Receler
python test_receler.py --model_name_or_path models/cifar10/receler-word_airplane/ --prompts_path evaluation/data/sample_prompts.csv

# Can also speed up the generation with Accelerate
accelerate launch --num_processes 2 test_receler.py --model_name_or_path models/cifar10/receler-word_airplane/ --prompts_path evaluation/data/sample_prompts.csv

The generated images will be saved in the test_results/ folder by default. You can set the --save_root argument to change the save location.

Evaluation

For details on how to evaluate the generated images, check the evaluation directory.

Acknowledgments

We thank ESD, as our code is largely built upon their work.

If you find this repo useful, please consider citing our work as follows:

@article{huang2023receler,
  title={Receler: Reliable concept erasing of text-to-image diffusion models via lightweight erasers},
  author={Huang, Chi-Pin and Chang, Kai-Po and Tsai, Chung-Ting and Lai, Yung-Hsuan and Yang, Fu-En and Wang, Yu-Chiang Frank},
  journal={arXiv preprint arXiv:2311.17717},
  year={2023}
}