TL;DR

Netflix's AVA system finds still frames that can represent a title. A season of about ten episodes holds nearly nine million frames, which is too many for an editor to scan. The pipeline annotates frames, then surfaces a short list of candidates that creative teams use for artwork.

Why it matters

Artwork is how a title gets chosen, and the useful frame is often a specific shot, not a random one. AVA is an engineering answer to that search: computer vision plus editorial criteria, run across the catalog. It is a concrete example of media understanding built into a production tool, not a demo on a single video.

Key takeaways

• The goal is representative frames, chosen from the source video rather than a separate photo shoot for every asset.
• Frames are annotated so the system can tell what is in the shot and whether it carries the story.
• Netflix used its Archer media platform to split video into chunks and process them in parallel.
• Diversity here means visual variety inside one title: shot size, composition, color, and what would misrepresent the show.
• Editors still choose. AVA proposes candidates. It does not publish the artwork by itself.