Muse: media editing and code without the confusion

What Muse actually includes

The first step is to separate three distinct products; otherwise, the Muse name quickly becomes confusing. Muse Image works with images, Muse Video with video, and Muse Code is designed for programming tasks and large codebases.

Meta’s official announcement for Muse Image and Muse Video focuses on generation and editing. Muse Image can build a composition from multiple references, combining people, objects, clothing, styles and environments. Meta also reports strong positions in Arena human-preference rankings for image generation and editing.

Muse Video is described as a text-to-video model that also received a high Arena rating. However, as of September 3, 2026, the official description does not confirm a dedicated object-removal or frame-cleanup mode comparable to specialized video erasers.

Cleaning visual clutter is a plausible practical use case of multimodal editing. But for now, it remains an engineering hypothesis rather than a stated feature with clear limitations. I would look closely at background stability, geometry preservation and flicker between frames: that is usually where such pipelines reveal their character.

Muse Code is a different case. Third-party summaries of Meta materials cite scores of 82.9 on Terminal-Bench 2.1 and 59.3 on DeepSWE 1.1. Yet these figures appear to be tied to Meta’s own evaluation setup rather than independently verified public leaderboard entries, so it is too early to compare them directly with competitors.

Where Muse can genuinely change the workflow

The most interesting Muse Code scenario is not generating a project from scratch, but serving as a second pass over existing code. The model may help simplify overengineered solutions, identify unnecessary abstractions and prepare a more compact patch.

I would assess it not by an impressive demo, but by the quality of the diff in a real repository: do tests still pass, does behavior remain unchanged, and does complexity actually decrease? For large codebases, the real context limits, long-session costs and agent reliability are equally critical, although precise official figures are not available in the accessible materials.

For now, Muse looks less like one universal model and more like an umbrella name for different tools. The main open question is not the score, but whether Muse can consistently remove what is unnecessary without adding a new layer of magic on top of the old one.

We also reviewed Seedance 2.0 in ChatCut: the model shows how generative tools handle motion and physics in a scene. That experience is useful when assessing background processing quality and artifacts in Muse.