I've been trying out alternative versions of this that pass images through to e.g. the Claude 3 vision models, but they're harder to share with people because they need an API key!