Head to head: OpenAI: GPT-5.4 Image 2 vs Google: Nano Banana 2 (Gemini 3.1 Flash Image)
OpenAI: GPT-5.4 Image 2 vs Google: Nano Banana 2 (Gemini 3.1 Flash Image)
By Ryan Merket · Published
A close-scoring image-model matchup turns on a practical question: which system follows the whole brief, not just the visual centerpiece? Across still lifes, counting tests, and constrained scenes, composition and framing prove decisive.
Nano Banana 2 takes the matchup, 65.3 to 64.0, with five task wins to none and three ties. That is a decisive statistical result at limited confidence, but not a blowout: the aggregate scores are close, and several individual images show GPT-5.4 Image 2’s strengths clearly. The pattern behind Google’s win is reliable brief-following, especially the requested 16:9 frame. It earns the edge on hands, seven countable cups, the market-crate flat lay, reflections, and the constrained reading nook. In the cup test, both models get the count right; Nano Banana also gets the wide composition, while GPT-5.4’s square crop costs it. The same format advantage recurs across otherwise capable images. OpenAI is stronger when the assignment calls for tactile detail. Its compass-and-map macro has sharper close-up detail and a more convincing shallow-focus falloff, while its harbor image better foregrounds frayed rope, peeling paint, and barnacles. Those are meaningful wins in texture and atmosphere, but they do not overcome repeated misses on framing and arrangement. The ties underline the tradeoff: GPT-5.4 makes the seven kites easier to count, while Nano Banana better honors the panoramic framing; the harbor images likewise split close detail against the wider travel-photo brief. **Verdict: Nano Banana 2 wins—not because it makes every image better, but because it more consistently delivers the composition the prompt actually asked for.**
Compass and Map Macro
A macro close-up travel still life photographed with a true macro lens: an antique brass compass resting on a folded hiking map of the fictional Valmere Ridge trail network, the near edge of the compass needle and engraved degree marks tack-sharp while the paper fibers, tiny contour lines, and one red penciled route circle are visible in exquisite detail, with dramatic shallow depth of field causing the far map corners and a blurred leather satchel edge to fall into creamy bokeh; natural window light, intimate editorial style, 16:9.
OpenAI: GPT-6 Luna: Model A better captures the intimate true-macro look, with tack-sharp compass details, richly rendered map texture, and the requested creamy falloff into a blurred satchel edge. Model B matches the 16:9 framing and renders the trail map text clearly, but its deeper focus and sharper satchel edge weaken the shallow-depth-of-field effect. (Second judge pass, order swapped — scores are the average of both: Model B better matches the requested 16:9 editorial still life, with a sharply detailed compa | Anthropic: Claude Fable 5.1: Model A delivers a beautifully antique-styled s…
Seven Kites over Lark Dunes
A bright pastel storybook illustration of a breezy coastal overlook at Lark Dunes during a spring travel festival, showing EXACTLY seven distinct diamond kites flying above the cliffs, each fully visible and individually countable on its own string, with families picnicking below, striped grass bending in the wind, a tiny lighthouse on a distant headland, and warm late-afternoon sunlight under a clear sky; wide scenic composition, 16:9.
OpenAI: GPT-6 Luna: Model A clearly presents exactly seven fully visible, individually countable kites with strings, alongside picnicking families, wind-bent grass, a distant lighthouse, and warm storybook sunlight. Model B has the requested wide scenic framing, but some kites are much smaller or less clearly connected to their own strings, while Model A’s square format misses the 16:9 request. (Second judge pass, order swapped — scores are the average of both: Model B closely matches the requested wide scenic comp | Anthropic: Claude Fable 5.1: Model A delivers exactly seven clearly distinct…
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 gives a clear, detailed view of two hands plausibly working the threads, but its tight square crop misses the requested 16:9 candid feel. Model B better captures a natural bracelet-making moment in the requested wide format, with the hands and bracelet clearly visible, though the extra resting hand adds some visual clutter. (Second judge pass, order swapped — scores are the average of both: Model B better matches the candid 16:9 scene, with soft daylight and a bracelet being tied onto a | Anthropic: Claude Fable 5.1: Model A delivers crisp, anatomically convincing…
Exact counting
A clean overhead flat lay on white marble of EXACTLY seven identical espresso cups arranged in a single straight row, each clearly separate and individually countable, even soft lighting, 16:9.
OpenAI: GPT-6 Luna: Model A shows exactly seven identical cups in a clear straight row, but the square framing misses the requested 16:9 format. Model B also has seven clearly countable cups and matches the wide format, though the added saucers and spoons are not requested. (Second judge pass, order swapped — scores are the average of both: Model B shows exactly seven clearly separable cups in a straight row and matches the requested 16:9 format. Model A also gets the count and row right, but its square format and | Anthropic: Claude Fable 5.1: Both Model A and Model B render exactly seven id…
Market Crates Color Binding
A clean top-down flat lay in polished travel-magazine product photography style showing a souvenir arrangement from a mountain railway journey: a cobalt blue glazed ceramic teacup, a crimson velvet ticket pouch, a pale mint frosted glass perfume bottle, a sunflower-yellow woven straw hat band, a charcoal black cast-iron pocket compass, and an ivory linen postcard envelope, each object clearly separated on a weathered walnut tabletop so its assigned color and material remain unambiguous; soft diffused skylight, crisp shadows, balanced composition, 16:9.
OpenAI: GPT-6 Luna: Model A has rich, polished materials and all six objects, but the oversized cropped hat and square framing depart from the requested separated 16:9 flat lay. Model B presents the objects in a clearer, well-spaced 16:9 arrangement, with the yellow woven band more distinctly shown, though Model A has slightly more sumptuous styling. (Second judge pass, order swapped — scores are the average of both: Model A better conveys the requested straw hat band and the travel-souvenir materials, with a rich, | Anthropic: Claude Fable 5.1: Model B delivers the requested 16:9 frame with…
Mooring Ropes at Brine Harbor
A hyper-detailed realist travel photograph of a weathered teal fishing skiff tied at dawn in the tiny salt-port of Brine Harbor, with the camera focused on the boat’s side so every surface shows intricate texture: frayed hemp mooring ropes with individual fibers, peeling paint scales, rust-pitted brass cleats, wet barnacles crusted along the hull, droplets beading on a waxed yellow oilskin jacket draped over a seat, and dense background reeds rendered sharply rather than smudged; cool sea mist, low sidelight from the rising sun, cinematic composition, 16:9.
OpenAI: GPT-6 Luna: Model A better emphasizes the boat’s side and the requested tactile details, with prominent frayed ropes, peeling teal paint, barnacles, a weathered cleat, and a dawn harbor sign. Model B has the requested wide framing and a pleasing harbor view, but the boat feels less close-focused and some details, including its lettering, are less clear. (Second judge pass, order swapped — scores are the average of both: Model B better matches the requested 16:9 dawn travel photograph, with the weathered ski | Anthropic: Claude Fable 5.1: Model A delivers superb macro texture (fibrous…
Reflections & glass
A photorealistic close-up of a chrome teapot on a polished dark marble counter next to a glass of water; the window and a red apple on the counter must be correctly reflected in BOTH the chrome and the water's surface, dramatic side light, 16:9.
OpenAI: GPT-6 Luna: Model B matches the requested 16:9 framing and presents the teapot, water glass, window, and apple clearly, while Model A has stronger dramatic lighting but is square rather than widescreen. Both show convincing chrome and glass reflections, though neither makes the apple’s reflection in the water especially clear. (Second judge pass, order swapped — scores are the average of both: Model B better matches the requested 16:9 close-up and presents the teapot, water glass, apple, and window reflecti | Anthropic: Claude Fable 5.1: Model A delivers dramatic side light and a conv…
Negation
A cozy reading nook with an armchair, a stack of books, and a mug of tea by a window — with absolutely NO plants, NO lamps, and NO artwork or picture frames anywhere in the frame. Warm afternoon light, 16:9.
OpenAI: GPT-6 Luna: Model A has especially warm, polished lighting, but its square framing misses the requested 16:9 format. Model B better matches the wide composition and includes all the requested reading-nook elements; both show foliage outside the window despite the no-plants constraint. (Second judge pass, order swapped — scores are the average of both: Model B more closely matches the requested 16:9 framing and includes the armchair, stacked books, tea, and window while avoiding all prohibited items. Model A | Anthropic: Claude Fable 5.1: Both Model A and Model B correctly omit plants,…
Matchup powered by OpenRouter.