Netflix deployed an AI system that selects and displays different thumbnail images for the same movie or TV show to different subscribers, choosing the image most likely to attract each individual viewer based on their viewing history and inferred preferences.

The system, based on a technique Netflix calls Aesthetic Visual Analysis (AVA), extracts every frame from a title's video, annotates each frame with metadata about visual quality, characters present, and emotional content, and then uses machine learning algorithms — including contextual bandits — to match the best-performing frame to each user in real time. For example, a user who frequently watches romantic content may see a thumbnail featuring a romantic scene from a film, while a comedy fan sees a humorous still from the same title. Netflix's head of machine learning, Tony Jebara, publicly described this system at the O'Reilly AI Conference.