Head to head: Google: Nano Banana 2 (Gemini 3.1 Flash Image) vs Krea 3 Edit
Google: Nano Banana 2 (Gemini 3.1 Flash Image) vs Krea 3 Edit
By Ryan Merket · Published
A close image-model matchup across eight prompts tests more than visual polish: it pits exact placement and readable poster text against scene-building, object control, and convincing action.
This is a narrow result, not a blowout: Nano Banana 2 finishes at 61.6 to Krea 3 Edit’s 60.1, with a 77% statistical lean and a 4–2 edge in task wins, plus two ties. The gap says Google was more consistently effective across this set—not that it dominated every kind of image task. Krea’s best results came when the prompt had a crisp, checkable layout. Its room placed the floor lamp correctly in the front-right corner, and its velodrome poster preserved the exact three-line wording. Google’s campsite and tabletop still-life were judged ties: the former looked richer, while the latter leaned more photorealistic, but neither pulled decisively ahead overall. Google’s wins were strongest in more demanding scenes and bindings. Its kayak image better preserved the sunrise team-and-dock context while distinguishing all six gear colors; its fencing hall more convincingly combined the lunge, mirror reflection, rain-streaked windows, and trophy case. It also took the shoe-count and bracelet-hands tasks, though neither was a clean demonstration of perfection: judges disagreed on the shoe totals, and Krea’s grip on the bracelet threads was clearer in one assessment. **Call: Nano Banana 2 wins on breadth, by a lean.** Krea 3 Edit is the sharper pick for exact layout and rendered text; Google edges the matchup by handling a wider range of scene requirements more reliably.
Restricted palette
A flat-vector illustration of a mountain campsite at dusk using ONLY four colors — cream, burnt orange, teal, and charcoal — with no gradients and no stray off-palette colors, balanced composition, 16:9.
OpenAI: GPT-6 Luna: Model A presents a more complete, atmospheric campsite with strong depth and a balanced arrangement, while Model B is clean but comparatively sparse and schematic. Both largely use the requested palette, though each includes some tonal variation within its colors. (Second judge pass, order swapped — scores are the average of both: Model A has a more polished, balanced campsite composition and a clearer dusk atmosphere, while both images largely honor the restricted palette. Model B is more stric | Anthropic: Claude Fable 5.1: Model A is richer and more detailed but breaks…
Spatial layout
A clean isometric illustration of a bedroom: a bed against the LEFT wall, a round rug centered on the floor, a desk under the WINDOW on the back wall, and a floor lamp in the FRONT-RIGHT corner. Flat-vector style, consistent perspective.
OpenAI: GPT-6 Luna: Model A is detailed and appealing, but its floor lamp sits near the center of the rug rather than in the front-right corner. Model B follows the requested layout more closely, with a clean isometric style and the lamp positioned beside the right wall. (Second judge pass, order swapped — scores are the average of both: Model B follows the requested placements closely, especially the floor lamp in the front-right corner and the centered rug. Model A is more detailed and polished, but places the la | Anthropic: Claude Fable 5.1: Model B places every element where the prompt s…
Velodrome poster text
A stylized stop-motion claymation scene of a tiny indoor velodrome concession stand during a race break, warmly lit by tungsten bulbs, 16:9; on the front counter hangs a clean, readable retro poster with the exact text: "LAP 27" on the top line, "RIVERMINT SPRINT" on the second line, and "TONIGHT" on the third line; the lettering must be sharp, correctly spelled, and easy to read from the camera’s front three-quarter view; include miniature cyclists streaking by in the blurred background, colorful paper pennants, and tactile handmade clay textures.
OpenAI: GPT-6 Luna: Model B renders the exact wording in the requested three-line layout, while Model A breaks “RIVERMINT SPRINT” across two lines. Model A has a richer, more lively composition and warm handmade detail, but Model B’s poster better satisfies the key text requirement. (Second judge pass, order swapped — scores are the average of both: Model B presents the requested poster prominently with all three exact lines clearly readable, alongside the warm concession stand, pennants, and blurred cyclists. Mode | Anthropic: Claude Fable 5.1: Model B renders the poster exactly as specified…
Kayak gear color-binding
A crisp editorial sports photo, 16:9, of a whitewater kayak team arranging equipment on a wet concrete dock at sunrise: a matte cobalt-blue carbon-fiber kayak, a glossy saffron-yellow fiberglass helmet, a brushed copper rescue whistle, a forest-green rubber dry bag, a burgundy leather thermos, and a pale lilac nylon throw rope, all clearly separated and resting around a kneeling athlete in a charcoal wetsuit; every object must keep its own exact color and material with no color swapping or bleeding, realistic textures, low-angle composition, cool morning mist, sharp natural light reflecting off puddles.
OpenAI: GPT-6 Luna: Model A captures the full sunrise team-and-dock setting with mist, puddle reflections, a kneeling athlete, and all six distinctly colored pieces of gear. Model B renders the equipment attractively, but the tightly cropped athlete and oversized kayak weaken the scene, and it omits the team context. (Second judge pass, order swapped — scores are the average of both: Model B presents the specified gear in a clean, low-angle arrangement, with the key colors clearly distinct and especially effective | Anthropic: Claude Fable 5.1: Model A renders every listed item in its correct…
Attribute binding
A photorealistic tabletop scene: a matte green cube to the LEFT of a glossy red sphere, a small blue cylinder BEHIND both, and a single yellow rubber duck sitting ON TOP of the cube. Colors must stay on the correct objects. Soft studio light, 16:9.
OpenAI: GPT-6 Luna: Model A presents all four objects with the requested colors and spatial relationships in a convincingly photoreal tabletop setting. Model B is clean and well composed, but looks more like a polished render and less like a photographic scene. (Second judge pass, order swapped — scores are the average of both: Model B presents all four objects clearly in the requested positions, with a clean studio-style setting and correct color and material bindings. Model A is highly photorealistic, but its boo | Anthropic: Claude Fable 5.1: Both outputs nail the attribute binding and spa…
Exact count triathlon shoes
A top-down commercial product photograph, 16:9, of a triathlon transition station on black rubber flooring under bright noon light: exactly 11 distinct running shoes laid out in two neat staggered rows beside a folded race bib and a half-full water bottle, each shoe fully visible and individually countable, with varied neon accents but no overlaps, duplicates merged together, or hidden pairs; clean high-detail sports catalog style, crisp shadows, precise spacing, and strong contrast against the dark floor.
OpenAI: GPT-6 Luna: Model A clearly shows exactly 11 individually visible shoes in two neat rows, alongside the bib and bottle on dark flooring. Model B is vibrant and organized, but appears to show only 10 shoes and presents them at more varied angles rather than in a consistent top-down layout. (Second judge pass, order swapped — scores are the average of both: Model B shows 10 distinct shoes, closer to the required count of 11 than Model A’s 12, while both include two neat rows, a bib, and a bottle. Model A is e | Anthropic: Claude Fable 5.1: Neither hits the exact count (Model A shows 12…
Hands & anatomy
A natural candid photo of a person tying a friendship bracelet, both hands clearly visible with anatomically correct fingers and a plausible grip on the threads, soft daylight, shallow depth of field, 16:9.
OpenAI: GPT-6 Luna: Model A has a polished candid setting and natural daylight, but one hand is mostly extended rather than clearly helping with the threads. Model B makes both hands and their plausible pinching grip on the bracelet threads the clear focus, although its tabletop scene feels less candid. (Second judge pass, order swapped — scores are the average of both: Model B clearly shows both hands working with loose bracelet threads, while Model A presents a more natural candid scene with convincing daylight, | Anthropic: Claude Fable 5.1: Model A delivers a convincing candid scene with…
Fencing hall mirror and glass
A cinematic indoor fencing scene, 16:9, in a polished training hall at dusk: one fencer lunges on a narrow piste beside a full-height wall mirror and a glass trophy case, with a chrome scoring box on the floor and rain-streaked windows in the background; the mirror must show the correct reflected pose and room geometry, the glass case must visibly refract and partially obscure a silver cup and a blue ribbon inside, and the chrome box must carry believable highlights of the piste lights; realistic sports photography style, dramatic sidelighting, shallow haze, precise perspective and reflections.
OpenAI: GPT-6 Luna: Model A more fully captures the requested cinematic hall, with a clear lunge, a convincing mirror reflection, rain-streaked windows, and prominent trophy case and chrome scoring box. Model B includes most of the setting elements, but its fencing pose is less dynamic and the case and scoring box are less legible and convincing. (Second judge pass, order swapped — scores are the average of both: Model A more clearly captures the lunging action, rain-streaked dusk setting, mirror reflection, refrac | Anthropic: Claude Fable 5.1: Model A delivers a far more convincing, cinemat…
Matchup powered by OpenRouter.