Muse Video is a video generation model developed by Meta Superintelligence Labs (MSL), previewed alongside the release of Muse Image on July 7, 2026. Built on the same pretraining base as its image-generating counterpart, the model is designed to produce high-fidelity video content directly from text descriptions. It marks a key step in Meta's shift toward in-house generative media models, moving away from relying on third-party foundational systems.
The core strength of Muse Video lies in its multimodal design, combining prompt adherence, visual fidelity, and temporal consistency. Notably, the model features native audio support, allowing it to generate synchronized sound effects or voiceovers in a single pass rather than relying on separate audio post-processing. At the time of its initial preview, Meta acknowledged ongoing work to close performance gaps in specialized areas such as precise audio-video synchronization and physically accurate fast-paced motion.
While architectural specifics and parameter counts have not been publicly disclosed, the model is designed to integrate into Meta's broader AI ecosystem. In early text-to-video benchmarks, the model achieved a top-three ranking on human preference leaderboards.
Prompting and Capability Focus
Meta's shared prompts emphasize combining visual aesthetics, camera directions, and detailed audio characteristics to leverage the model's native multi-sensory capabilities. For example, an official benchmark prompt includes describing both the visual medium—such as 'a handmade paper-cutout / construction-paper collage stop-motion animation'—and the speaker's tone, pacing, and voice quality: 'calm warm male narrator with a clear natural voice and relaxed, well-enunciated pacing, 16:9'.