Meta Ships Muse Image, Its First In-House AI Model

Meta just stopped renting its image generation stack. On July 7, Meta Superintelligence Labs shipped Muse Image — the first image model Meta has built entirely in-house — and previewed a video sibling, Muse Video. For years, Meta AI's image and video features ran on licensed models from Midjourney and Black Forest Labs. That arrangement is over.

What Shipped

Muse Image is live now inside the Meta AI app, meta.ai, Instagram Stories in the US, and WhatsApp in select countries, with Facebook and Messenger coming soon. It's free. Muse Video, built on the same pretraining base, adds native audio support and is described as coming soon to creators and Meta AI. Both ship with Content Seal, Meta's invisible watermarking for verifying AI-generated media.

This is Meta Superintelligence Labs' second major release, following Muse Spark — the large language model that replaced Llama internally back in April. Alexandr Wang, Meta's Chief AI Officer, leads the group.

The Agentic Pitch

Meta isn't describing Muse Image as a straight prompt-to-pixels diffusion model. The company says it operates as an agent: it invokes search and coding tools to improve accuracy, self-refines its own generations, and scales test-time compute to improve output. Layer that on top of Muse Spark for prompt reasoning, and you get a model that plans a composition before it renders anything — supporting multi-reference composition and precision editing rather than one-shot generation.

How It Actually Stacks Up

Meta claims a No. 2 spot on the Arena leaderboard for text-to-image, single-image editing, and multi-image editing, measured by human-preference Elo. Read the fine print, though: independent trackers put Muse Image only 9 Elo points ahead of third place and roughly 105 points behind OpenAI's GPT Image 2. It does beat Google's Nano Banana 2 on single and multi-image editing tasks. Early testers on Hacker News rank it a notch below both Nano Banana Pro and GPT Image in practice. No official benchmarks, parameter counts, or API pricing have been published — every specific performance number in circulation right now is either Meta's own internal claim or community impression.

Why Developers Should Care

The headline isn't the leaderboard position. It's ownership. Meta no longer depends on a licensing partner's roadmap, pricing, or usage policy for its most-used consumer AI surface — it sets all three now. That's a meaningful shift if you're building anything on top of Meta AI outputs, or just tracking where the frontier image-gen market moves next, in the same week Anthropic, OpenAI, and Google are all shipping or defending their own model lineups. There's a privacy wrinkle too: Muse Image can pull from public Instagram photos for social context in @-mentions, and the opt-out setting is buried rather than surfaced by default. If you run a public account, worth checking now.

Sources: CNBC, Meta AI blog, DataCamp, eesel AI.