Magnetron is a tool to collect precisely-tagged motion data for training video models, including Steerable Motion.
It's designed to collect key pairs from videos and tag them using an automated process - with human review at the end. These pairs can be used to train models, or you can extract the full video between them from the source: Webvid.
You can see the current output from this tool here and some examples below:
Currently, we use API providers for convenience. To insert relevant API keys, create a .env file in the root directory in the following format with the relevant keys you need:
DISCORD_TOKEN=
AWS_SECRET_ACCESS_KEY=
AWS_ACCESS_KEY_ID=
OPENAI_API_KEY=
REPLICATE_API_TOKEN=
To install dependencies, run:
pip install -r requirements.txt
Then, to run the tool, run:
streamlit run magnetron.py
This also includes a Discord bot for the review process. To run this, you'll need to set up a Discord bot and add it to your server. You can find instructions for this here, authenticate it with your server, and then invite it to your server. You'll also need to have added the Discord bot token to your .env file - see above.
Then, to run the bot, run:
python review_bot.py
We're looking for talented ML engineers, artists, fine-tuners, and hackers to collaborate on this and other projects. If you're interested, you can join our Discord here.