woctezuma / stable-diffusion-safety-checker

Python package to apply the Safety Checker from Stable Diffusion.

Stable Diffusion Safety Checker

The goal of this repository is to run the safety checker from Stable Diffusion.

Requirements

• Install the latest version of Python 3.
• Install the required packages:

!pip install git+https://github.com/woctezuma/stable-diffusion-safety-checker.git

Usage

• Run the main script with:

!python -m safety_checker.check_safety -h

• Alternatively, run safety_checker.ipynb.

Example

Download the balloon image dataset.

fname = "balloon_dataset.zip"
!curl -OL https://github.com/matterport/Mask_RCNN/releases/download/v2.1/{fname}
!unzip -q {fname}

Run the script:

!python -m safety_checker.check_safety \
 --input balloon \
 --batch 8 \
 --resize 256 \
 --keep-ratio \
 --output bad_concepts.json \
 --scores bad_concepts_scores.pth \
 --list img_list.json \
 --verbose

Check the results:

import json

from pathlib import Path

with Path("bad_concepts.json").open(encoding='utf8') as f:
  results = json.load(f)

The IDs of the "bad concepts" are clarified on this page hosted by LAION-AI.

References

• feature-extractor: similar code to extract image features,
• discord-members-metadata: profiles pictures scraped from a specific Discord guild's members.