Meta Unveils Muse Video, Its First AI Video Generation Model

Meta Superintelligence Labs unveiled Muse Video on July 7, 2026, its first in-house text-to-video generation model. It's an early preview for now, not a finished product, but results from an independent comparison ranking already place it near the top of a market dominated by Google, ByteDance and Runway.

What Meta showed

Muse Video was built on the same training foundation as the earlier Muse Image and inherits part of its architecture. The company highlights native audio support, meaning the audio track is generated together with the visuals in a single model pass rather than as a separate step bolted on afterward, which Meta says improves consistency between what's seen and what's heard.

In comparative tests on the Arena platform, where users rate generated content in blind comparisons, Muse Video ranked third in human preference for the text-to-video category, just behind the leading competing models. Meta touts strong prompt adherence, image quality, and consistency across successive frames.

Limitations the company admits to

The company itself points out where the model still falls short. Weak spots include audio-visual synchronization during faster scenes and physical plausibility of rapid motion, such as sports or action sequences. Meta says these two areas are now the team's priority before the model gets a full release.

For now, Muse Video hasn't launched in any public app. Meta says the product will arrive soon for creators and in the Meta AI app, without giving a specific date. That sets it apart from Muse Image, which went straight into use across several of the company's products.

A shadow of controversy over Muse Image

The Muse Video launch comes a week after the loud and controversial debut of Muse Image, an image model that lets users generate photos based on likenesses pulled from other people's public Instagram profiles without asking the account owner for consent. That feature drew criticism over privacy and the use of other people's likenesses without explicit authorization.

Muse Video doesn't currently offer an equivalent feature based on other people's profiles, but since it shares its architecture with Muse Image and is set to reach the same consumer products, questions about control over user data and likenesses will likely resurface once the model exits preview.

The generative video market gets crowded

Meta's entry into video generation coincides with a wave of competing launches. In recent weeks ByteDance showed off Seedance 2.5 with native thirty-second generation and local editing, while Google keeps expanding its Veo line, including the latest variant integrated with the Gemini Omni Flash model. The common thread among the new models is generating visuals and synchronized audio together in a single pass, something that required separate tools as recently as a year ago.

For content creators and marketers, that means an increasingly short path from idea to a finished video with sound, but also growing competition between platforms over which model best handles character consistency and realistic motion, exactly the elements Meta itself flags as weaknesses in Muse Video.

Sources: official Meta AI blog (ai.meta.com), ForkLog (forklog.com), Pexo (pexo.ai)