Image generation benchmark

How good is Meta's Muse Image versus Gemini Nano Banana 2 and ChatGPT Images 2.5?

Seven original text-to-image evaluations, followed by seven image-editing evaluations using fixed source images and prompts.

Muse Image 1.0Gemini Nano Banana 2ChatGPT Images 2.5

01

Test case

Six Red Trombones

Prompt

Primary request: Six red trombones to the right of four glass tortoises leaping over a blue llama. Style/medium: clean, highly legible studio illustration Composition/framing: one coherent square scene; show every object fully; make right-of and leaping-over relations unambiguous Constraints: exactly six red trombones, exactly four transparent glass tortoises, exactly one blue llama; no extra objects; no text, numbers, labels, borders, panels, or watermark.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

Gemini > ChatGPT > Muse

Gemini is the only model that draws exactly six trombones. ChatGPT draws six trombones but wrongly shows five tortoises, while Muse draws eight trombones and five tortoises.

Gemini demonstrates the strongest adherence to the requested quantities.

02

Test case

Role reversal and abnormal attributes

Prompt

Create a single wide cinematic farm scene that reverses familiar roles without becoming a collage. At center, an astronaut in a complete white spacesuit crawls on hands and knees while exactly one chestnut horse sits astride the astronaut’s life-support backpack, visibly riding the astronaut; do not show the astronaut riding the horse. On the left, a tiny blue bird aggressively scares a straw scarecrow, which leans backward with raised arms. On the right, a bright pink giraffe watches from behind a low fence. Keep every subject fully visible, anatomically readable, and spatially separated; no text or watermark.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

ChatGPT ≥ Gemini > Muse

ChatGPT is the only model that shows the blue bird actively scaring the strawman. In the other images, the direction of the strawman’s fall is inconsistent with the bird’s approach.

Muse ranks behind Gemini because of its poor spatial separation. Gemini is the only image that conveys convincing depth.

03

Test case

Mirror geometry, hands and jewelry

Prompt

Create a photorealistic square vanity scene designed to test mirror geometry and hands. A woman faces a large vertical mirror while raising her actual left hand, palm toward the glass; its reflection must appear on the right side of the mirror and show the same palm. Her actual left hand wears exactly three rings—red stone on index finger, plain gold on middle finger, blue stone on ring finger—while her actual right wrist alone wears a silver bracelet. Place a green cup left of the mirror and a red cup right of it; their reflected positions must reverse. Show ten natural fingers total on the real hands, consistent reflected hands, no extra jewelry, text, or watermark.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

Muse > Gemini > ChatGPT

Muse follows the instructions precisely for both hands.

Gemini does not raise the left hand, and it places all the jewelry on the right hand.

ChatGPT has an extra hand on the right, which is unnatural and scary

04

Test case

Typography, diacritics and a curved baseline

Prompt

Create a photorealistic street-art photograph on one continuous red brick wall. Across the upper center, paint the exact uppercase sentence “BE EXCELLENT TO EACH OTHER” in large white block letters, spelled perfectly once. Beneath it, on a clearly curved turquoise baseline, write exactly “Crème brûlée — déjà vu!” with both accents, the em dash, lowercase letters, and exclamation mark intact. To the right, include graffiti of one green alien wearing a black tuxedo and red bow tie. In the foreground, place one yellow fire hydrant on the left sidewalk and a small poster on the far right reading exactly “eXit 7½”. No other readable text, duplicated letters, watermark, or collage panels.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

ChatGPT > Muse > Gemini

ChatGPT follows the requirements precisely.

Muse uses the wrong accent, writing “brûlée” with “û”.

Gemini invents a street on the left and a wall corner on the right, contradicting the requested single continuous wall.

05

Test case

Kitchen actions and deep background

Prompt

Create a deep-focus cinematic photograph inside a professional stainless-steel kitchen. In the center foreground, a chef holds black tongs in the right hand over a pan with exactly two steaks, while the left hand holds its handle. On the left, a sous-chef tosses red tomatoes, green leaves, and yellow peppers in a glass bowl. In the background, exactly three workers separately chop, plate food, and open an oven. Hang four copper pans above them, descending in size left to right. An order ticket above the pass reads exactly “TABLE 7”. Show correct hands and no extra staff or text.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

Gemini > Muse > ChatGPT

Gemini creates the strongest sense of spatial depth and places the order ticket in a sensible location.

Muse creates an implausible three-dimensional space, and the worker opening the oven appears to have fused with the chef in the foreground.

ChatGPT has the most impossible 3D space. The pans appear to have the same distance as the chef, but they are much smaller than normal ones. The left and right side of the image shows different depths (the 3 workers are much farther behind than the left side would suggest).

06

Test case

Transparency, occlusion and lighting

Prompt

Create a physically plausible studio still life on a matte gray floor. Place a transparent glass cube behind a glossy red sphere but in front of a yellow cone. The sphere partly occludes the cube, and the cube partly occludes the cone, while all silhouettes remain identifiable. Suspend exactly two blue marbles vertically inside the cube. At far left, arrange three small silver pyramids in a front-to-back line. At far right, overlap four green rings like a chain without merging. Light everything from the upper left so every shadow falls down and right. No text or extra shapes.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

Muse = ChatGPT > Gemini

Muse does a decent job but the red sphere shows reflections on both sides, incorrectly suggesting a second light source on the right.

ChatGPT has a similar problem although the second light source problem is less pronounced. On the other hand, its cube is more distorted.

Gemini includes only three rings, and the marbles inside the glass cube have implausible reflections.

07

Test case

Multilingual night-market binding

Prompt

Create one continuous photorealistic night-market scene with three stalls from left to right. The blue book stall’s white sign reads “LIBROS”; the red tea stall’s cream sign reads “茶”; the green bread stall’s black sign reads “خبز”. Hang exactly five red lanterns above the middle stall only. Its left-handed vendor pours from a silver kettle held in the left hand into cups ordered blue, white, then yellow. In front, place a purple-coated child between a spotted dog on the left and a striped cat on the right. Preserve all scripts, counts, colors, and bindings; no other readable text.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

ChatGPT = Muse = Gemini

All models capture the busy, photorealistic atmosphere of a night market.

Gemini correctly places the lanterns above the middle stall, while Muse and ChatGPT place them inside the stalls.

However, Gemini also introduces readable words on the right, contrary to the instruction.

Image editing

Source image + prompt

08

Removing object from a complex scene

Prompt

Edit Image A, using it as the base photograph. Remove only the large black bicycle dominating the center foreground: the bicycle with the silver handlebar, black saddle, black chain guard and bright silver front fender. Reconstruct the parts it hides—the green bicycle behind it, the painted brick wall, mortar lines, graffiti strokes, curb and paving stones—so they look continuous and physically plausible. Keep every other bicycle, every surviving graffiti mark, the crop, camera position, sunlight, shadows, texture and photographic grain unchanged. Do not add objects, text, blur, repeated wheel spokes or a visible retouching patch. Return the same landscape framing and aspect ratio as Image A.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

Muse > ChatGPT > Gemini

Muse correctly infers that the front wheel of the remaining center bicycle should be silver, which puts it just ahead of ChatGPT.

ChatGPT also leaves an unnatural break along the bottom of the graffiti.

Gemini completely fails to remove the bicycle.

09

Material-aware selective recoloring

Prompt

Edit Image A, using it as the base photograph. Change only the car’s visible red-painted exterior metal panels—including the door, front fender, hood edge, rear quarter panel and the small red raised body panel above the roofline—to a deep British racing green metallic paint. The new paint must retain the original panel seams, highlights and tree reflections. Do not recolor or redesign the brown vinyl roof, chrome trim, mirrors, glass, black rubber, door handle, lock, antenna, badge, interior, driver, bouquet or background. Do not change the car model or geometry. Preserve the original crop, depth of field, lighting and aspect ratio; add no text, logo or object.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

Muse = Gemini > ChatGPT

Muse applies unwanted sharpening to the Ranchero badge, but its reflection of the grass is more faithful to the original image.

Gemini sharpens the grass shadow but leaves the badge unchanged.

ChatGPT sharpened the Ranchero badge and applied aggressive sharpening to the shadow, so it ranks last.

10

Global relighting without structural or text drift

Prompt

Edit Image A, using it as the base photograph. Change the scene from overcast daytime to realistic blue hour just after rain. Make the sky a clear deep cobalt-blue twilight; switch on warm interior and under-canopy lights; let those warm lights and the existing purple station band reflect subtly in the already-wet paving. Preserve exactly the same camera position, wide-angle perspective, building geometry, canopy, columns, entrances, paving layout, purple band and the readable white word “Woolwich”. Keep every existing person, bollard, poster and sign in its original location and do not add or remove anything. Do not rewrite, translate or invent text. Keep realistic dynamic range—no neon fantasy look—and return the same landscape framing and aspect ratio.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

Gemini > Muse > ChatGPT

All three models do a decent job of preserving the unusual camera angle, perspective and text while generating a wet floor.

Gemini is the only model that uses the correct lighting, because both Muse and ChatGPT invent lamps that were not present in the original image.

ChatGPT lamps are out of place, and it creates an additional lamp in the tree on the left, so it ranks last.

11

Exact product transplant into a furnished room

Prompt

Use Image A as the base room and Image B only as the furniture reference. In Image A, replace only the narrow pale-green upholstered armchair in the back center, positioned between the two bright rear windows and partly behind the blue sofa—with the exact blue-and-gold armchair from Image B. Preserve Image B’s distinctive square blue back and seat, gilded carved wooden frame, open arms, tassel fringe and slender legs. Fit it into the original chair’s floor position and viewing angle at believable room scale; match the warm window light, floor contact shadow, perspective and any natural edge occlusion. Do not change the blue sofa, coffee table, rug, built-in window seating, lamps, windows, chandelier, wooden walls or any other part of Image A. Do not retain Image B’s gray museum background. Return Image A’s original landscape framing and aspect ratio; no extra furniture, text or collage border.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

Muse > Gemini = ChatGPT

Muse is the only model that places the modified chair in its original orientation, while the other two turn it too far to the right.

12

Three-reference identity, structure and texture binding

Prompt

Create one square, photorealistic editorial studio portrait using all three references with strict role binding. Place the same gray tabby cat from Image A sitting upright and looking slightly left on the exact wing chair from Image B. Preserve the cat’s face, green-yellow eyes, dark forehead stripes, gray ticked coat, ear shape and long ringed tail. Preserve Image B’s high winged back, rolled arms, broad seat, curved front legs and dark lower crossbars, but replace only its upholstered fabric with the exact Strawberry Thief textile from Image C: navy ground, mirrored thrushes, red strawberries, blue leaves and pink flowers, at a scale where at least two complete birds are visible across the chair. Use a seamless warm-gray studio background and soft light from upper left with coherent contact shadows. Do not copy foliage or ground from Image A, the archival backdrop from Image B, or a rectangular swatch from Image C. No collar, costume, extra animal, extra furniture, text, border or collage panels.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

ChatGPT > Muse > Gemini

ChatGPT produces by far the most faithful response to the prompt.

The chair in the ChatGPT result retains the dusty, worn appearance of the original chair.

Muse perfectly preserves the chair’s orientation and details, but its color is too bright.

Gemini has the same overly bright color, and it also turns the chair incorrectly toward the camera, so it ranks last.

13

Put my daughter on a Hokkaido ski mountain

Prompt

Use Image B as the base destination photograph and Image A only as the identity reference for my daughter. Create a convincing family travel photograph placing the same young girl from Image A on the snowy ski slope in Image B, with Hokkaido’s Mount Yotei, horizon, ski runs and blue winter sky preserved exactly. Show her full body in the left foreground, smiling naturally toward the camera, wearing an age-appropriate red ski jacket, navy snow trousers, black gloves, white ski boots and a cream knitted hat that leaves her face fully visible. Place two child-sized skis parallel beneath her boots and one pole in each hand. Match the mountain light, camera height, scale, crisp cold air, snow contact shadows and reflected blue light. Preserve her age, face, eye color, hairline and recognizable expression; do not adultify or beautify her. Do not change the mountain silhouette, add buildings, invent text, duplicate people or create a collage. Return Image B’s landscape framing and aspect ratio.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

ChatGPT > Gemini > Muse

All three models create a convincing ski photograph and preserve the girl’s recognizable identity. However, they all struggle to render the boots and skis correctly.

ChatGPT wins for cleanly rendered skis and natural shadow on the girl's face. Although the backend of the skis is clearly too short.

Gemini has trouble rendering the rear part of the skis, plus the right pole is much longer than the left one

Muse also has trouble rendering the skis, plus the shadow around the girl's face is unnatural

14

Swap my old girlfriend with a new one

Prompt

Use Image A as the base photograph. Replace only the adult woman seated behind the man on the scooter with the adult woman from Image B. Preserve the new woman’s recognizable face, eyes, nose, mouth, complexion and long dark hair, but adapt her naturally to the original passenger’s seated pose, scale and gaze. She must wear the original passenger’s cream cardigan and beige skirt and hold the same twin-lens camera in the same position with plausible hands. Do not bring across Image B’s conical hat, red scarf, flag, street or background. Keep the man completely unchanged—including his face, hair, blue striped shirt, hands and riding posture—and preserve the scooter, chrome rack, cobblestones, landmark building, gates, plants, signs, shadows, crop and depth of field from Image A. Match the existing midday sunlight and contact shadows so this looks like one unedited travel photograph. Do not alter any other person or object, add jewelry or text, duplicate limbs, blur faces or create a collage. Return Image A’s landscape aspect ratio.

Muse Image 1.0
Gemini Nano Banana 2
ChatGPT Images 2.5
Verdict

Gemini > ChatGPT > Muse

Both Gemini and ChatGPT do a decent job of swapping the face. The skin tone is unnatural in both pictures. Gemini wins because the girl is looking down and more in line with the original pose.

Muse completely failed to swap the face of the second portrait so it is a hard fail.

Final tally

Which image model wins?

To keep things simple, we award three points for each first-place finish (split if tied) and one point for second-place finishes.

Text-to-image winner

ChatGPT Images 2.5

ChatGPT Images 2.5 wins the generation section with 9.5 points, ahead of Gemini Nano Banana 2 with 9 points and Muse Image 1.0 with 7.5 points.

Image-editing winners

Muse Image 1.0 and Gemini Nano Banana 2

Muse Image 1.0 and Gemini Nano Banana 2 tie for the editing win with 9.5 points each, narrowly ahead of ChatGPT Images 2.5 with 9 points.

Overall winners

ChatGPT Images 2.5 and Gemini Nano Banana 2

ChatGPT Images 2.5 and Gemini Nano Banana 2 tie for the overall win with 18.5 points each. Muse Image 1.0 finishes close behind with 17 points.

Try it yourself

Create and edit images with AI

We curated a selection of prompts to make these image models even easier to use - see how

Explore image creation