Meta's Movie Gen was a research preview, not a product

The shot

On 4 October 2024, Meta's AI research division published a blog post premiering Movie Gen, describing a '30B parameter transformer' video model able to 'generate videos of up to 16 seconds at a rate of 16 frames per second,' alongside a '13B parameter audio generation model' producing audio 'up to 45 seconds.' The post states the system covers four capabilities: text-to-video generation, personalized video from a user photo, precise text-based video editing, and synchronized audio generation. Thirteen days later, on 17 October 2024, Meta's research division posted the accompanying technical paper, 'Movie Gen: A Cast of Media Foundation Models,' to arXiv, with a revised version following in February 2025.

What the documents show

The blog post states plainly that 'Movie Gen outperforms similar models in the industry across these tasks when evaluated by humans,' a comparison Meta ran and reports itself rather than an outcome any outside body confirmed. The arXiv abstract adds detail the blog omits, describing a training context of '73K video tokens' and stating the models 'set a new state-of-the-art on multiple tasks' by Meta's own account. Meta's blog explicitly separates the announcement from a shipped product, saying only that the company will 'work closely with filmmakers and creators' toward 'a potential future release.'

The workflow

For a production, Movie Gen at this stage offered no account signup or API; the blog names Blumhouse as a partner whose selected filmmakers made videos with the model 'before its public debut,' the same partnered-access pattern seen elsewhere in this research cycle. The four capabilities described, particularly text-based editing that targets 'only the relevant pixels' rather than regenerating a whole frame, describe a workflow step still requiring a human edit decision about which pixels should change and why, not an automated substitute for that judgment.

What the tool does not change

Meta's own post states the company 'acknowledge[s] that our current models have limitations,' naming inference time and quality among the areas needing further work, in the vendor's own words rather than a critic's. The claimed human-evaluation win rates describe preference judgments on selected comparison sets, not a general guarantee that any given generated shot will outperform an alternative in production. Meta's post itself cautions that generative tools are 'not a replacement for the work of artists and animators,' a statement this site treats as the vendor's own framing rather than an independent finding.

• Is a cited win rate a human-preference comparison the vendor ran, or an independent third-party evaluation.
• Does the accompanying paper's technical detail change what the announcement post claimed, or only extend it.
• Was the demonstrated edit selective about which pixels changed, and does the source describe that as a targeted or a global operation.

Movie Gen's own materials describe a research preview with partnered access and self-reported evaluation, dated weeks apart across a blog post and a paper, not a production tool generally available in October 2024.