Movie Gen

What it is

Movie Gen is a set of media foundation models that Meta premiered on 4 October 2024. A foundation model is a general-purpose model that can be adapted to many tasks. Meta called Movie Gen its "third wave" of generative AI research.^[1] Text-to-video means the model turns a written description into moving images, and text-to-image does the same for a single picture.

The set covers four capabilities: video generation, personalised video generation, precise video editing and audio generation.^[1] Each one is described below as Meta presented it, so the figures are Meta's own.

Origins

Meta trained the models on a combination of licensed and publicly available datasets.^[1] A research paper describing the work was published at ai.meta.com/research/movie-gen.^[1]

Main features

Video generation. The video model is a 30-billion-parameter transformer. It is optimised for both text-to-image and text-to-video, and it produces videos up to 16 seconds long at 16 frames per second. A frame rate of 16 frames per second is low by film standards.^[1]

Personalised video. This capability takes a person's image together with a text prompt and produces a video that features that person.

Precise editing. The editing capability takes an existing video and a text prompt. It supports localised edits, such as adding, removing or replacing elements, and global changes such as the background or the style.^[1]

Audio generation. A 13-billion-parameter audio model generates up to 45 seconds of ambient sound, sound effects and instrumental music. The audio is synchronised to the video. Ambient sound means background sound that sets the scene, such as wind or traffic. An extension technique allows the audio to cover longer videos.^[1]

Human evaluation. Meta reported positive net win rates against competing industry models in A/B human evaluations across all four capabilities. In an A/B test, people see two outputs and pick the one they prefer, so a positive net win rate means the Meta outputs were picked more often than the competitors' outputs.^[1]

Its place in development

Meta did not release Movie Gen publicly at launch. It said it would work with filmmakers and creators before any potential future release.^[1] The research was therefore a showcase of what the models could do, not a product that people could use.

Movie Gen was never released publicly.^[2] In July 2026, Meta Superintelligence Labs introduced Muse Image and Muse Video, which it described as its first media generation models.^[2] Movie Gen is the predecessor research to the Muse models, which are separate models with their own capabilities.^[2]

For a learner, Movie Gen is a useful reference for what the field looked like in 2024: text-to-video, personalised video and audio generated to match the picture, all developed in one research set.^[1]