Motivated by recent advancements in text-to-image diffusion, we study erasure of specific concepts from the model's weights. While Stable Diffusion has shown promise in producing explicit or realistic artwork, it has raised concerns regarding its potential for misuse. We propose a fine-tuning method that can erase a visual concept from a pre-trained diffusion model, given only the name of the style and using negative guidance as a teacher. We benchmark our method against previous approaches that remove sexually explicit content and demonstrate its effectiveness, performing on par with Safe Latent Diffusion and censored training.

To evaluate artistic style removal, we conduct experiments erasing five modern artists from the network and conduct a user study to assess the human perception of the removed styles. Unlike previous methods, our approach can remove concepts from a diffusion model permanently rather than modifying the output at the inference time, so it cannot be circumvented even if a user has access to model weights

Given only a short text description of an undesired visual concept and no additional data, our method fine-tunes model weights to erase the targeted concept. Our method can avoid NSFW content, stop imitation of a specific artist's style, or even erase a whole object class from model output, while preserving the model's behavior and capabilities on other topics.

The finetuned weights for both NSFW and art style erasures are available on our project page.

To run the gradio interactive demo locally, clone the files from demo repository

• Create an environment using the packages included in the requirements.txt file
• Run python app.py
• Open the application in browser at https://127.0.0.1:7860/
• Train, evaluate, and save models using our method

• To get started clone the following repository of Original Stable Diffusion Link
• Then download the files from our repository to stable-diffusion main directory of stable diffusion. This would replace the ldm folder of the original repo with our custom ldm directory
• Download the weights from here and move them to stable-diffusion/models/ldm/ (This will be ckpt_path variable in train-scripts/train-esd.py)
• [Only for training] To convert your trained models to diffusers download the diffusers Unet config from here (This will be diffusers_config_path variable in train-scripts/train-esd.py)

After installation, follow these instructions to train a custom ESD model:

• cd stable-diffusion to the main repository of stable-diffusion
• [IMPORTANT] Edit train-script/train-esd.py and change the default argparser values according to your convenience (especially the config paths)
• To choose train_method, pick from following 'xattn','noxattn', 'selfattn', 'full'
• python train-scripts/train-esd.py --prompt 'your prompt' --train_method 'your choice of training' --devices '0,1'

Note that the default argparser values must be changed!

The optimization process for erasing undesired visual concepts from pre-trained diffusion model weights involves using a short text description of the concept as guidance. The ESD model is fine-tuned with the conditioned and unconditioned scores obtained from frozen SD model to guide the output away from the concept being erased. The model learns from it's own knowledge to steer the diffusion process away from the undesired concept.

To generate images from one of the custom models use the following instructions:

• To use eval-scripts/generate-images.py you would need a csv file with columns prompt, evaluation_seed and case_number. (Sample data in data/)
• To generate multiple images per prompt use the argument num_samples. It is default to 10.
• The path to model can be customised in the script.
• It is to be noted that the current version requires the model to be in saved in stable-diffusion/compvis-<based on hyperparameters>/diffusers-<based on hyperparameters>.pt
• python eval-scripts/generate-images.py --model_name='compvis-word_VanGogh-method_xattn-sg_3-ng_1-iter_1000-lr_1e-05' --prompts_path 'stable-diffusion/art_prompts.csv' --save_path 'evaluation_folder' --num_samples 10

The preprint can be cited as follows

@article{gandikota2023erasing,
  title={Erasing Concepts from Diffusion Models},
  author={Rohit Gandikota and Joanna Materzy\'nska and Jaden Fiotto-Kaufman and David Bau},
  journal={arXiv preprint arXiv:2303.07345},
  year={2023}
}