Meta

Key Points

• Meta's new image model Muse Image works as an agent, using external tools like web search and refining its own outputs automatically.
• On the Image Arena platform, the model ranks second behind OpenAI's GPT Image 2. The video model Muse Video, shown alongside as a preview, comes in third.
• An opt-out feature that lets users generate AI images of other people via their Instagram username without consent is already drawing criticism. Privacy regulators in Europe are expected to scrutinize the feature.

Meta's Superintelligence Labs ships its first image model, and it works like an AI agent.

Meta has released Muse Image, the first image generation model from its Superintelligence Labs. It's the first release of its kind since the company restructured its AI lab under Chief AI Officer Alexandr Wang, backed by billions in new investment.

Muse Image doesn't map prompts directly to images. Instead, it works as an agent, similar to OpenAI's GPT Image 2. The model calls tools to produce more accurate results and refines its own outputs.

Those tools include writing and running code to generate things like correct diagrams, scannable QR codes, animated GIFs, websites, or interactive games. A web search feature lets the model ground images in current facts and real-world references, which Meta says improves accuracy on knowledge-heavy prompts.

Muse Image corrects its own intermediate results through local edits or full regeneration. Meta says this self-refinement behavior emerged on its own during reinforcement learning because it led to better images and higher reward scores.

As with language models, quality scales with the compute the model uses at inference time. Reasoning scales far better than the brute-force approach of generating multiple images and keeping the best one.

Second place behind OpenAI

For image editing, Muse Image is designed to change only what users ask for while keeping everything else consistent across multiple editing steps. It can also combine elements from several reference images, including people, objects, clothing, and environments.

On the Image Arena evaluation platform, Muse Image ranks second in human preference scores for text-to-image and for both single- and multi-image editing. It trails OpenAI's GPT Image 2 in each category but beats models like Nano Banana and Grok Imagine.

Muse Video, shown alongside as a preview, sits at third place for text-to-video. Meta acknowledges weaknesses in audio-video sync and fast motion.

Video: Meta

Other people's faces via @-mention

Muse Image is available now in the Meta AI app, on meta.ai, in Instagram Stories in the US, and in WhatsApp. Facebook and advertisers are next. The new model also brings a feature that's already drawing criticism. Users can @-mention a public Instagram account in a prompt, and Meta AI will pull publicly visible photos from that profile to generate a new image of that person. No consent from the person in the photos is required. A username is enough.

The feature is on by default for public accounts. Anyone who wants out has to actively opt out. That means going into Instagram settings and turning off reuse of posts and Reels. Images already generated won't be deleted.

Privacy regulators expected to scrutinize the feature under GDPR

The feature is likely to face pushback in Europe, where data protection rules are far stricter. If it launches in the EU at all, the opt-out approach probably won't hold up. Because the feature uses publicly available photos of real people, observers expect close scrutiny under the General Data Protection Regulation and potential questions around biometric data protections. Meta hasn't announced any GDPR-specific adjustments for the launch.

The EU AI Act adds another layer. Its transparency rules require that AI-generated or manipulated image, audio, and video content resembling real people and qualifying as deepfakes be clearly labeled as artificially created. These obligations under Article 50 take effect on August 2, 2026, just weeks after Muse Image's launch. Meta's invisible watermark system, Content Seal, which survives cropping, compression, and screenshots, points in that direction. But whether a machine-readable-only watermark satisfies the AI Act's labeling requirements remains an open question. The law requires that labeling be recognizable to the people affected. Critics also argue that a watermark applied after the fact may prove where an image came from, but it doesn't stop the image from being created in the first place.