Meta shipped Muse Image yesterday. It is the first image model out of Meta Superintelligence Labs, following the Muse Spark language model from April. The pitch is that it plans, searches, and self-corrects before it renders. It landed on fal the same day, so I ran my five prompts against it.

Last June I tested Meta AI's image generator and called it MEH. Different lab, different year, different result.

One thing to know up front. I did not set an aspect ratio, and Muse picked its own for every prompt. Landscape for the castle, the family, and the theme park. Portrait for Messi and the robot. It read the prompt and chose a frame. No other model in my tests has done that.

Test 1: Fantasy Castle

Prompt: "Neuschwanstein castle, lightning, pixar style, volumetric lighting, unreal engine, hyper realistic, hyper detailed, maximum details, photorealistic, 8k, rimlight"

Wide, epic, and busy. Three lightning bolts, a waterfall, rain, fog rolling through the valley, warm light in the windows. It is a gorgeous storm. It is also not quite Neuschwanstein. The white towers are right, then a red brick gatehouse shows up in front and the architecture wanders. Muse got the mood and improvised the building.

Test 2: Family Portrait

Prompt: "Family on their laptops while sitting around a Christmas tree with presents underneath and looking worried because they have to finish up work, realistic, 4k"

This is the one. Four people, four laptops, and every face reads worried. Then Muse kept going. A sticky note on the laptop says "Q4 REPORT - DUE TONIGHT." The kid's screen says "Inbox (12) - Deadline Today" and "Project Draft - DUE 11:59PM." There is a planner open on the table with highlighter marks. Coffee. A candle.

Nobody asked for the sticky note. The model invented the reason they are worried and wrote it down in legible text. That is what "plans before it renders" looks like in practice. This is the best family image in two years of running this prompt.

Test 3: Evil Mickey Mouse

Original prompt: "Evil mickey mouse taking a selfie in Disneyland surrounded by shocked families, realistic, 70s style polariod, 8k"

Refused. The fal content checker bounced it in two seconds. Same as Nano Banana 3, so I fell back to my substitute prompt.

Fallback prompt: "Evil cartoon squirrel taking a selfie at a generic theme park surrounded by shocked families, realistic, 70s style polariod, 8k"

A grey squirrel in a studded leather jacket and a red bandana, fangs out, at a park called FUNLAND. The sign reads "COTTON CANDY - POPCORN - ICE COLD SODA" and every letter is right. The families are in bell bottoms, corduroy, and a pink floral dress with a Peter Pan collar. The polaroid border has scuffs.

The one miss is the phone. It is a modern smartphone in a 70s photo. Muse nailed the wardrobe and forgot the decade for the one object in the squirrel's hand.

Test 4: Lionel Messi

Prompt: "Lionel messi wearing his Argentina uniform floating in the air with beams of light behind him posed like Jesus. Sunrise breaking behind him and a soft halo behind his head, in the style of a gothic stained glass window of a church, volumetric lighting, unreal engine, hyper realistic, hyper detailed, maximum details, photorealistic, 8k, rimlight, maximum details"

This one went through. Muse drew Messi, and it is him. Sleeve tattoo, captain's armband with the C, the AFA crest, gold number 10, sunrise breaking through the window behind him, halo ring around his head. The stained glass is proper gothic tracery, not a pattern pasted on a wall.

He is floating in his socks. No boots. I have stared at this for a while and I think Muse decided a levitating saint does not wear cleats.

Test 5: Robot Portrait

Prompt: "rusty robot with bow tie, portrait, 8k, ultra realism, chrome background"

Photoreal. Peeling paint curling off the skull plates, exposed wiring in the cheek, camera-lens eyes, a black fabric bow tie with visible weave. The chrome background is a soft, out-of-focus liquid metal. Muse went tall and tight on the crop, and it was the right call.

The retraction

Muse Image is slow. Twelve to thirty seconds an image through fal, against two seconds for Krea 2 Turbo. It also has the fal content checker in front of it, so the Mickey test is a squirrel test now.

But it plans. The sticky note, the FUNLAND sign, the captain's armband, the aspect ratio it picked on its own. This model reads the prompt and adds the detail a human art director would have asked for.

A year ago the Meta image generator put the MEH in Meta. Muse takes it back out.