Meta Muse Image and Muse Video

The first image and video models from Meta Superintelligence Labs, nearly two years after Movie Gen was announced and never shipped. Muse Image works less like a generator than an agent: it can search the web to ground a picture in real facts, write and run code for exact charts, QR codes and animated GIFs, and check its own drafts and fix them, a habit Meta said emerged in training ("We didn't design this behavior"). It edits only what it is asked to across many turns and composes people, clothing, objects and settings from several reference images at once. It launched in the Meta AI app, on meta.ai, in Instagram Stories in the US and in WhatsApp, every output carrying an invisible Content Seal watermark. Within three days Meta pulled an Instagram feature that let users @-mention public accounts as visual references, saying it "missed the mark."

Muse Video, built on the same base, came as an early preview: four ten-second samples with native audio, and Meta named its own weak spots, audio-video sync and physically accurate fast motion. On Arena's human-preference leaderboards, Muse Image ranked second for text-to-image and image editing and Muse Video third for text-to-video. Meta was back in the race, with the largest distribution of any model maker behind it.