Muse Image and Muse Video

What it is

Muse Image and Muse Video are media generation models from Meta Superintelligence Labs. Meta announced them on 7 July 2026 as its first media generation models.^[1]

Meta called Muse Image its most advanced image generation model. It follows instructions, edits precisely, composes images from multiple reference images, and can use agentic tools such as search and code.^[1] Agentic tools let the model take steps beyond generating pixels, such as looking up information before it draws.

Muse Video is built on the same pretraining base as Muse Image and generates native audio, meaning the sound is produced together with the video rather than added afterwards.^[1]

Origins

Muse Image and Muse Video follow Meta's earlier research on media generation. The predecessor was Movie Gen, presented in October 2024, which was never released publicly.^[2]

Meta Superintelligence Labs described the launch as the start of its media work. Muse Image integrates with the Muse Spark model. That integration supports coherent multi-turn editing and inline prompts that mix text and images.^[1]

Main features

Multi-turn editing. Muse Image supports coherent editing across several turns of conversation. A user can ask for one change, then another, and the model keeps the image consistent with the earlier edits.^[1]

Mixed prompts. Inline text-and-image prompts let a user combine words and pictures in one request.^[1]

Quality and compute. Meta reported that self-refinement emerged during reinforcement-learning training. Quality rises roughly log-linearly in Elo with more test-time compute. Test-time compute is the extra work a model does when it generates an answer, and a log-linear rise means each doubling of that work adds a similar amount to the score.^[1]

Human preference rankings. On Arena human-preference rankings as of 5 July 2026, Muse Image ranked No. 2 for text-to-image, single-image editing and multi-image editing. Muse Video ranked No. 3 for text-to-video.^[1] Arena rankings are based on votes from people comparing outputs.

Known gaps. Meta named audio-video synchronisation and physically accurate fast motion as current gaps for Muse Video.^[1]

Watermarking. Images from Muse Image in Meta AI carry an invisible "Content Seal" watermark. A detection tool is in preview.^[1] An invisible watermark is embedded in the image itself, so it is not visible to viewers but can be checked by a detection tool.

Its place in development

Muse Image was available at launch in the Meta AI app and on meta.ai. It also appeared in Instagram Stories in the United States and in WhatsApp in limited countries. Facebook was described as coming soon.^[1]

Muse Video was announced for creators and Meta AI, and entered early preview at launch.^[1]

Compared with Movie Gen, the Muse models are the first in this line to reach users, at least in preview. Movie Gen was a research set that Meta did not release publicly.^[2] For a learner, the change is the difference between a research demonstration and a product, even though the underlying work on image and video generation is related.^[1][2]