Head to head: Google: Nano Banana 2 (Gemini 3.1 Flash Image) vs OpenAI: GPT-5.4 Image 2
Google: Nano Banana 2 (Gemini 3.1 Flash Image) vs OpenAI: GPT-5.4 Image 2
By Ryan Merket · Published
Eight image-generation tasks put composition, exact details, typography and prompt compliance under the microscope. The results turn on a recurring trade-off: matching the brief’s format versus getting particular visual elements exactly right.
Nano Banana 2 takes five tasks to GPT-5.4 Image 2’s three, with no ties. Its strongest advantage is prompt compliance: it delivered the requested 16:9 framing in the cup-count, courier, perspective, watchmaker and Quantum Lane tests, while GPT-5.4 repeatedly defaulted to square images where the brief called for widescreen. That consistency mattered most when framing was part of the task. Nano Banana’s empty, converging aisle and watchmaker-at-work scene better matched their specified compositions; its Quantum Lane poster also kept the title together and the supporting copy visible. These are not just prettier-image wins—they are cases where the output followed the assignment more closely. GPT-5.4 had real wins, not consolation prizes. It placed all four elements correctly in the isometric room, produced the cleaner full-bleed navy text poster, and made the glass-kiosk scene richer in reflections and detail. When spatial placement, typography or material rendering took priority, it could beat Nano Banana on the brief’s central requirement. The numbers keep this verdict narrow: GPT-5.4 has the slightly higher aggregate score, 64.8 to 64.3, even as Nano Banana wins more tasks and receives the 75% statistical lean. That is a close matchup, not a runaway—and the small score gap makes the task-level pattern more persuasive than the aggregate edge. **Final call: Nano Banana 2 wins on more reliable instruction-following, especially around framing. GPT-5.4 Image 2 is competitive and sharper on a few specific tasks, but its repeated aspect-ratio misses cost it the head-to-head.**
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 clearly separated cups in a straight row on white marble and matches the requested wide 16:9 framing. Model B also gets the count and arrangement right, but its square framing departs from the prompt and leaves more unused space. (Second judge pass, order swapped — scores are the average of both: Model B shows exactly seven cups in a clean row, but its square framing misses the requested 16:9 format. Model A also shows exactly seven clearly separated cups and better m | Anthropic: Claude Fable 5.1: Both Model A and Model B render exactly seven i…
Courier leap with drones
A cinematic cyberpunk action scene in 16:9: a bike courier in a teal rain jacket is captured mid-leap from a narrow rooftop to a hovering delivery drone platform, one sneaker just leaving the ledge and a hard case of prototype foldable phones swinging from a shoulder strap; sheets of rain streak diagonally, neon magenta and amber billboards blur in the background, loose packaging slips whirl through the air, and the camera uses a low angle with slight motion blur on the city while the courier’s face and gadget case stay sharply frozen, creating convincing speed, weight, and dynamic energy.
OpenAI: GPT-6 Luna: Model A follows the requested 16:9 framing and clearly shows the courier leaping from a rooftop toward a hovering delivery platform amid rain, neon, and flying packaging. Model B has a striking low-angle action pose and a sharply rendered courier, but its square format and less explicit rooftop-to-platform leap make it a weaker match overall. (Second judge pass, order swapped — scores are the average of both: Model A better fulfills the requested widescreen framing and clearly stages the courier | Anthropic: Claude Fable 5.1: Model A delivers the requested 16:9 frame with…
Perspective & scale
A photorealistic one-point-perspective shot down a long empty library aisle, shelves receding to a single vanishing point, books getting consistently smaller with distance and nothing warped or wrongly sized, warm overhead light, 16:9.
OpenAI: GPT-6 Luna: Model A more fully matches the requested wide 16:9 view, with a long, empty aisle, strong single-point perspective, and warm overhead lighting. Model B has a convincing centered aisle, but its square framing and less expansive depth fit the prompt less closely. (Second judge pass, order swapped — scores are the average of both: Model A better matches the requested 16:9 framing and presents a long, polished aisle with strong, consistent convergence toward a single vanishing point. Model B also ha | Anthropic: Claude Fable 5.1: Model A delivers a clean 16:9 one-point perspec…
Watchmaker hands close-up
A detailed naturalistic studio portrait in 16:9 of a young wearable-tech designer seated at a workbench, shown from waist up and clearly facing camera while assembling a translucent smart ring; both hands are fully visible in the foreground, one hand delicately holding the ring with thumb and index finger and the other using a precision screwdriver, with all ten fingers anatomically correct, natural knuckles, believable nail shapes, and realistic proportions; scattered tiny components, a magnifying lamp, and matte-black gadgets sit on the desk, and soft side lighting from the left creates gentle shadows that emphasize accurate human anatomy and hand structure.
OpenAI: GPT-6 Luna: Model A better matches the requested 16:9 naturalistic workbench portrait, with the magnifying lamp, scattered components, and both hands clearly engaged in assembling the ring. Model B has a strong centered portrait and convincing hand detail, but its square framing misses the specified aspect ratio. (Second judge pass, order swapped — scores are the average of both: Model A better matches the requested 16:9 workshop portrait, with a clear view of the designer, both hands assembling the ring, a | Anthropic: Claude Fable 5.1: Model A delivers the requested 16:9 frame, a na…
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 B places the floor lamp in the front-right area and otherwise follows the requested layout, though its rendered look is less flat-vector than requested. Model A better matches the flat isometric style, but its lamp sits near the front center rather than in the front-right corner. (Second judge pass, order swapped — scores are the average of both: Model B places all four requested elements correctly, including the floor lamp in the front-right corner, and has a polished, coherent isometric | Anthropic: Claude Fable 5.1: Both place the bed on the left, the rug centered…
Legible multi-line text
A minimalist event poster with three lines of crisp, correctly-spelled text stacked and centered: 'NIGHT MARKET' large on top, 'Fridays · 6–11pm' in the middle, 'Riverside Pier 4' at the bottom, on a deep navy background, clean sans-serif, subtle grain.
OpenAI: GPT-6 Luna: Model B presents all three correctly rendered lines directly on a deep navy, subtly grained background, with a clean centered layout. Model A also renders the text accurately, but the gray mockup surround and shadow are less faithful to the requested minimalist poster. (Second judge pass, order swapped — scores are the average of both: Model B presents all three correctly spelled lines in a centered, full-bleed navy layout with subtle grain, closely matching the prompt. Model A also renders the | Anthropic: Claude Fable 5.1: Both render all three lines correctly spelled an…
Quantum Lane poster
A clean editorial poster mockup pinned inside a startup hardware lab, rendered as crisp vibrant flat-vector design in 16:9, lit by soft morning window light: the poster must prominently and legibly read exactly "QUANTUM LANE" on the top line, "BUILD SMALL. THINK STRANGE." below it, and a small footer line "LAB 7B"; include illustrated microchips, soldering irons, and a tiny robot arm around the typography, with generous spacing, high contrast, and a straight-on composition so every letter is easy to read.
OpenAI: GPT-6 Luna: Model A follows the requested 16:9 editorial-vector look and keeps “QUANTUM LANE” together on the top line, with the requested subtitle, footer, and lab illustrations clearly visible. Model B is striking and polished, but its title is split across two lines and its overall rendering is less like the requested flat-vector poster. (Second judge pass, order swapped — scores are the average of both: Model A follows the requested flat-vector style and keeps both headline lines and the footer legible | Anthropic: Claude Fable 5.1: Model A nails the brief: true 16:9, consistent f…
Glass kiosk reflections
A hyperreal product-shot scene in 16:9 inside a midnight metro station: a transparent glass repair kiosk displays disassembled smartwatches and tiny screwdrivers on clear shelves, surrounded by chrome edges, mirrored floor tiles, and a puddle from tracked-in rain; through the kiosk walls you can see a woman in a saffron coat studying a holographic route map, while the glass panels correctly reflect the tiled station, refract the tools and watch parts behind them, and show subtle double reflections from layered panes; cool cyan overhead lighting mixes with a warm orange ad screen for a convincing test of glass, transparency, reflections, and refraction.
OpenAI: GPT-6 Luna: Model A better honors the requested 16:9 station scene and places the woman and route map visibly behind the kiosk glass. Model B has especially detailed watches and tools with strong reflections, but its square framing departs from the prompt and the woman is less clearly seen through the kiosk walls. (Second judge pass, order swapped — scores are the average of both: Model B more fully delivers the glass-testing brief, with abundant watch parts and tools, layered reflections, a rain puddle, an | Anthropic: Claude Fable 5.1: Model A delivers the required 16:9 frame, a con…
Matchup powered by OpenRouter.