Meta Launches Muse Image and Muse Video: Agentic AI That Self-Corrects Its Own Mistakes

Meta Superintelligence Labs (MSL), the research division of Meta Platforms, has today unveiled two new generative AI models: Muse Image for image generation and precise editing, and Muse Video for video content. Muse Image is publicly available immediately, while Muse Video enters a preview phase for selected creators. Both models introduce a technical distinction that sets them apart from conventional generative tools — an agentic architecture through which the model internally corrects its own mistakes without any user intervention.

Arena Leaderboard Ranking

The Arena leaderboard is a reference system for comparing AI models based on direct user voting, where evaluators do not know which model produced which output. As of July 5, 2026, Muse Image holds the #2 position across three categories simultaneously: text-to-image (generating an image from text), single-image editing (precise editing of a single image), and multi-image editing (editing and combining multiple images). Muse Video simultaneously holds the #3 position on the Arena leaderboard for the text-to-video category.

Achieving high rankings across three distinct evaluation categories simultaneously is a rare result. While many models are optimized for a specific task, Muse Image demonstrates strong competency in both pure text-to-image generation and precise editing — tasks that technically demand different underlying capabilities.

What Agentic Architecture Means for the End User

The key distinction of Muse Image is not only the visual quality of its outputs, but the architecture through which those outputs are produced. The model operates as an agent that internally, without explicit user instructions, invokes external tools:

Code tool — Muse Image writes and executes code where that is more precise than direct visual generation. Concrete examples include data visualizations, QR codes, and animated GIFs. Rather than “hallucinating” the appearance of graphic elements, the model generates them algorithmically through actual code execution.

Web search tool — when a user requests generation that requires current or factually accurate visual information, the model searches the web and grounds its output in real data rather than trained assumptions.

It is particularly noteworthy that the self-correction capability emerged spontaneously during training — MSL did not explicitly program or request it. The model identifies anomalies in its own generation and corrects them through an internal iterative loop, resulting in consistently higher average quality without requiring feedback instructions from the user.

Meta has also documented a log-linear relationship between inference-time compute and output quality — the so-called test-time compute scaling. Investing more compute resources at generation time, without retraining the model, directly improves the visual result. This property is shared by leading language models, but Muse Image applies it to visual generation and editing tasks.

Multi-Reference Composition and Integration with Muse Spark

Muse Image supports multi-reference composition: the user provides multiple input images as visual references, and the model intelligently combines them into a new generation. For Instagram users, the system can draw on visual references from public Instagram profiles — aesthetics, style, or a person — and integrate them into a new image. The platform also offers personalized presets that allow users to quickly apply their favorite styles.

The model is integrated with Muse Spark — Meta’s system for collaborative agentic planning. The combination of Muse Spark (planning) and Muse Image (execution) opens up the possibility of automating multi-step visual workflows. Meta specifically highlights the use case for generating marketing materials for small business owners who do not have graphic designers.

Content Seal: Robust Watermarking for AI-Generated Visuals

Every image generated with Muse Image automatically carries a Content Seal — an invisible watermark embedded directly into the image’s pixels, not its metadata. The key advantage: Content Seal survives cropping, compression, and screenshots, unlike EXIF tags that are easily stripped by standard image-editing tools.

Meta also offers a publicly available Content Seal verification tool that can confirm whether an image was generated by Muse Image. This directly addresses regulatory requirements for labeling AI-generated content that are emerging from the EU regulatory framework and global initiatives for digital media transparency.

Muse Video: Native Audio and Preview Phase

Muse Video is currently in preview and coming soon to the Creators platform and the Meta AI app. Its key technical feature is native audio support — sound is generated in parallel with the video, not added in post-production. The model already holds the #3 position on the Arena leaderboard for text-to-video.

Meta is transparent about current limitations: audio-video synchronization and accurate rendering of fast motion are areas under active development and improvement.

Platforms and Availability Starting Today

Muse Image is available from today on:

• Meta AI app and meta.ai — globally
• Instagram Stories — users in the US
• WhatsApp — a limited number of markets
• Facebook — integration in preparation

With this launch, Meta Superintelligence Labs enters direct competition with the leading visual generation models, combining high Arena rankings with an agentic architecture and a robust watermark that meets upcoming regulatory requirements. Integration across Meta’s platforms with more than three billion active users gives Muse Image unprecedented reach potential in the visual generative AI segment.