Meta has launched Muse Image, the first image generation model to come out of its Superintelligence Labs. The release marks a milestone since the company reorganized its AI lab under Chief AI Officer Alexandr Wang and poured billions into new investment.
Unlike traditional image models that map prompts directly to pictures, Muse Image operates as an agent. It calls on external tools to produce more accurate results and automatically refines its own outputs. This approach is similar to what OpenAI does with GPT Image 2.
The tools it uses include writing and running code to generate correct diagrams, scannable QR codes, animated GIFs, websites, and interactive games. A web search feature allows the model to ground images in current facts and real-world references, which Meta says improves accuracy for knowledge-heavy prompts.
Self-refinement through reinforcement learning
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.
For image editing, the model 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.
Second place behind OpenAI on Image Arena
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.
Stay updated
Get the day's AI and automation news in your inbox. No spam, unsubscribe anytime.
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.
Controversial Instagram photo feature
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 will get access next.
The new model brings a feature that is 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 will not be deleted.
EU privacy regulators expected to scrutinize the feature
The feature is likely to face pushback in Europe, where data protection rules are stricter. If it launches in the EU at all, the opt out approach probably will not 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 has not 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 must 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 does not stop the image from being created in the first place.
Related on Neura Market: