Meta has released Muse Image, its first image generation model from Superintelligence Labs. The model works as an AI agent, using external tools like web search to refine outputs and improve accuracy. It ranks second on Image Arena behind OpenAI's GPT Image 2, with Muse Video placing third for text-to-video. The model is available in the Meta AI app, on meta.ai, in Instagram Stories in the US, and in WhatsApp. Facebook and advertisers are set to follow. Source: thedecoder
Muse Image is designed to edit images precisely, changing only what users request while maintaining consistency across multiple steps. It can also combine elements from multiple reference images, including people, objects, and environments. On Image Arena, it scores second in human preference for text-to-image and both single- and multi-image editing. It outperforms models like Nano Banana and Grok Imagine but lags behind OpenAI's GPT Image 2. The model's self-refinement behavior emerged during reinforcement learning, leading to better images and higher reward scores. Meta claims quality improves with compute used at inference time, while reasoning scales better than generating multiple images and selecting the best one. Source: thedecoder
The model includes a feature allowing users to @-mention a public Instagram account, prompting Meta AI to use the account's publicly visible photos to generate images of that person. No consent is required, and the feature is enabled by default for public accounts. Users must opt out by adjusting Instagram settings. Critics argue this raises privacy concerns, especially in Europe, where data protection regulations like GDPR and the EU AI Act may scrutinize the feature. The AI Act requires AI-generated content resembling real people to be labeled, but whether Meta's invisible watermark meets these standards remains unclear. Source: thedecoder
