Muse Image

Text-to-image starts from nothing, while edit mode starts from one to ten images you connect. Use text-to-image for posters, infographics, product shots and QR graphics that need exact text and layout; it can't lock a specific subject or face. Edit mode changes, combines or restyles existing images, but needs at least one connected image and regenerates the whole frame, since there is no mask.

Use edit mode. Connect the photo, phrase one clear instruction, and name what to keep unchanged, such as the face, pose, background and lighting. Edits can drift a little if the region is described loosely, so be specific about what to change and avoid stacking many unrelated edits in one instruction.

Up to ten. Supply a subject, a setting and a style source as separate images, and refer to them by input order in the prompt (for example @Image1 and @Image2) so the model knows which element comes from where.

Mostly yes. Headlines, labels, chart numbers and QR codes come out crisp and correctly spelled most of the time. Wrap exact text in quotation marks and keep strings short. Long paragraphs are weak, multi-word strings can show occasional letter swaps, and text is non-deterministic, so some runs need a re-roll.

Mask-based inpainting or outpainting, long passages of body text, exact pixel dimensions, tricky mirror and reflection physics, and stylized painterly illustration. Without style cues the look is polished, neutral realism, so add style direction when you want more character.

It was created by Meta Superintelligence Labs and released on 2026-07-07. It generates and edits images in one workflow, with output reaching about 1600px on the longest edge.