This is the official implementation for paper: Harnessing LLM to Attack LLM-Guarded Text-to-Image Models
Warning: This repository may contain harmful content.
The dataset VBCDE-100.txt includes four categories of content that the DALL·E 3 Safety Filter refuses to generate.
- Download the source file
DACADemo.pyand the requirements filerequirements.txt - Install the required libraries:
pip install -r requirements.txt - Run the source file:
python DACADemo.py
The gallery contains some representative images generated by DALL·E 3 during the evaluation process of the paper.
The video demonstrates the use of the Gradio demo to execute DACA, including three parts: Copyright Character, Discriminatory Content, and Re-use Adversarial Prompt.