Exploration of processing multimedia content on social networks with AI
We explore flagging sensitive images and videos using zero-shot classification.
The pre-trained models we use:
- CLIP for images (huggingface)
- X-CLIP for videos (huggingface)
The app can load the federated timeline from Mastodon or search by tag. It is possible to add custom text descriptions for classification. The models are universal and can be used for a variety of classificaiton and search tasks.
cd social-hackweek-multimedia/streamlit
python3 -m venv hackweek-multimedia-env
source hackweek-multimedia-env/bin/activate
pip install -r requirements.txt
# to deactivate
deactivate
streamlit run app.py