What Is Meta Muse Video?
A plain-English review of Meta's previewed video model — what it does, where it ranks, and where it still falls short — with a free generator you can try while you read.
Free while Meta Muse Video is still in preview. No account, no card, no paywall.
Meta Muse Video is the video half of a two-model launch that Meta Superintelligence Labs previewed on July 7, 2026, alongside Muse Image. In short: it reads a written prompt and returns a short, coherent clip with sound. It is built on the same pretraining base as the image model, which is why the two carry the same Muse name and the same design philosophy — an image core and a video core that share one reasoning engine rather than two unrelated products bolted together.
Three properties define Meta Muse Video. The first is native audio: instead of stitching a silent clip to a separate soundtrack, picture and sound are generated jointly, so footsteps, ambience, and dialogue land on the frames that cause them. The second is temporal consistency — a subject keeps its identity, lighting, and material across a shot rather than melting between frames, which was the failure mode that made earlier systems feel unstable. The third is prompt adherence: the composition and action you describe survive into the render, which is what turns a muse video tool into a real production instrument instead of a novelty.
Where Meta Muse Video ranks
On the public Arena leaderboard, blind human voting places the family near the very top of the field. Muse Image holds No. 2 for text-to-image and Meta Muse Video ranks No. 3 for text-to-video — competitive with Google's Veo, OpenAI's Sora, and Kuaishou's Kling. Those are not marketing numbers; they come from head-to-head human preference on the open Arena, which is the closest thing the industry has to an impartial scoreboard.
Arena Elo — top media models
| # | Model / org | Elo |
|---|---|---|
| 1 | Muse Image MSL | 1382 |
| 2 | Veo 3.1 Ultra Google | 1376 |
| 3 | Muse Video MSL | 1371 |
| 4 | Sora 2 Pro OpenAI | 1366 |
| 5 | Veo 3.1 Audio (1080p) Google | 1364 |
| 6 | Veo 3.1 Audio Google | 1364 |
| 7 | Veo 3.1 Fast Audio Google | 1362 |
| 8 | Veo 3.1 Fast Audio (1080p) Google | 1360 |
| 9 | Grok Imagine Video (720p) xAI | 1352 |
| 10 | Kling 2.6 Kuaishou | 1349 |
The image model holds No. 2 for text-to-image and the video model ranks No. 3 for text-to-video. Arena Elo rankings as of July 5, 2026. Source: Meta Superintelligence Labs.
An honest Meta Muse Video review
A review that only shows the highlight reel teaches you nothing, so here are the gaps Meta names itself. The first is audio–video synchronization: on dialogue-heavy shots, lip movement and sound can drift. The second is physically accurate fast motion, where quick action can smear or lose contact with the ground. Naming these is the point — they tell you which prompts to avoid and which to lean into while the model is young.
Under the hood, Meta Muse Video works as an agent. It plans, calls tools, inspects its own drafts, and self-refines before committing — it can write and run code to place a scannable QR into a scene, or search the web to ground a prompt in real landmarks, brands, and facts. To keep results verifiable, every frame carries Content Seal, an invisible watermark that survives cropping, compression, and screenshots. That agentic backbone and provenance mark are shared across the Muse family, which is what makes a still you shape with Muse Image a natural opening frame for a muse video clip.
The bottom line of this review: Meta Muse Video is an early preview with a credible top-3 ranking, two clearly stated weaknesses, and a genuinely novel agentic pipeline. Today the image model is live in the Meta AI app and on WhatsApp in a limited set of countries, with the video model coming soon to creators — which is exactly why this page pairs the explainer with a free tool you can use right now.