Meta has introduced the Muse Image image generation model on its Meta AI blog and provided a first look at Muse Video.
https://twitter.com/AIatMeta/status/2074577662840832382
The AI solutions for media content creation were developed by Meta's Superintelligence Labs. Muse Image is now available in the Meta AI app, on meta.ai, in Instagram Stories for users in the U.S., and in WhatsApp in select countries.
Meta claims that Muse Image is its most advanced image generation model. According to the company, the model more accurately follows prompts, can edit images, assemble scenes from multiple references, and utilizes the social context of Instagram.
At Meta Superintelligence Labs, it was stated that Muse Image operates as an agent-based model: it uses internet search, coding tools, and can refine images during the generation process.
Source: Meta.Muse Image has been integrated with Muse Spark. According to Meta, this combination of models leverages code and media generation to create animated GIFs, websites with embedded images, and interactive visual games.
Simultaneously, Meta showcased an early preview of Muse Video. The company claims that this model is built on the same pre-training foundation, delivers high visual quality, and natively supports audio. Meta added that they are still refining audio-video synchronization and the accurate representation of fast motion.
Meta also announced that images created with Muse Image in the Meta AI app receive an invisible watermark called Content Seal. The company asserts that this watermark persists even after cropping, compression, resizing, or screenshotting. In the future, Meta plans to extend this watermark to videos as well.
Additionally, Meta provided metrics for Arena. According to the company, Muse Image ranks second in the categories of text-to-image, single-image editing, and multi-image editing, second only to GPT Image 2.
Source: Meta.As a reminder, at the end of June, Meta unveiled an AI system that restores typed text from brain activity recordings without surgical intervention. The company claims that the average word-level accuracy reached 61%.