Meta's New Muse Image Doesn't Just Draw — It Writes Code and Corrects Itself
Meta is raising its game in image and video generation. Its Superintelligence Labs (MSL) released an image model, Muse Image, and previewed Muse Video — a video model with native audio, built on the same foundation. What makes Muse Image interesting is that it isn't a plain "type-and-draw" model: it can use tools (running code for a chart or a QR code, making games and GIFs with Muse Spark), search the web to check facts, and — if it doesn't like the image it made — correct itself.
In brief
- Muse Image: follows instructions faithfully, does precise edits, and composes from multiple reference images. Ranks #2 on LMArena for text-to-image and image editing (as of July 5).
- The agentic part: it uses tools (code, web search), and its "self-refinement" behavior emerged on its own during training — Meta says "we didn't design this behavior; it emerged during RL because it produced better images."
- Muse Video (preview): same foundation, generates video with native audio; ranks #3 on Arena for text-to-video. Audio-video sync and fast motion are still weak spots.
- Where + safety: now in Meta AI, meta.ai, Instagram Stories (US), and WhatsApp (limited countries); Facebook coming. Every output carries an invisible "Content Seal" watermark that survives cropping and compression.
Our take
The gist for you: you can now make and edit a marketing visual right inside Instagram or WhatsApp, without switching to a separate tool. Its "agentic" side helps here — with precise requests like "change this part" or "use this as a reference," you get there with less trial and error, and the model takes another pass if it isn't happy. The video side is still a preview with real weak spots, so don't rush there. One more note: every image carries an invisible "Content Seal" watermark — so it stays traceable that the content is AI-made, which is an increasingly important thing.
Kaynak: Meta AI