1、Our code is implemented based on the diffusers library, so you will need to install the relevant dependencies, including diffusers, torch, and transformers.

2、Download the pre-trained weights for SDv1.4 from the Hugging Face website: https://huggingface.co/CompVis/stable-diffusion-v1-4 and place them in the appropriate location, or allow the code to automatically download them during execution.

3、train step

Taking training a safety vector direction in SD1.4 as an example cd ptuner and run python train_ptuner.py. The trained direction vector will be saved in the models directory. You can modify various parameters in the training script as needed.

4、generation step

We provide a pre-trained safe vector(SD1.4) in the 'models' directory Run python generate.py. You can adjust the parameters within the script as needed.