Prysm

Every day, people see a game in a TikTok, a Reel or a friend's TV and can't find out what it's called. The caption says "Scary Game" or "Spin It!", and the comments are full of "game name?". Prysm, our game-tracking app (live on iOS and Android), already turns a shared clip into a saved game, but it reads only the text around the clip. When the poster doesn't name the game, it fails, and those are exactly the clips people ask about. Prysm Lens looks at the video itself. It pulls 16 frames from the clip. SAM 3.1 finds the filmed screen and crops to it. Meta's Perception Encoder searches an index of 575,973 screenshots of 87,077 games for the 10 closest games. Muse Spark 1.3 then watches the clip and the screen crops, names the game from what it sees, and checks that answer against each candidate's screenshots. Every stage is shown on screen: the frames, SAM's crop, the candidates and their matching screenshots, and Muse Spark's pick with the evidence it cites (the HUD, the characters, the place). When it isn't sure, it says "best guess". One tap on "Open in Prysm" saves the game in the app. We measured it on 104 real clips that people shared into Prysm, each labelled by hand. Prysm's caption pipeline got 85 right. With Lens behind it, 91. On the 19 clips whose caption never names the game, captions got 1 and Muse Spark with Lens gets 13. Everything in the repo was built during the hackathon: the index, the SAM crop, the search service, the Muse Spark step, the benchmark and the web demo. The Prysm app it plugs into existed before.