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
Two image models face prompts that test more than visual polish: they must handle exact framing, constrained styles, readable text and precise object counts. The comparison turns on which kinds of prompt fidelity matter most.
The saved aggregate scores are **63.2** for Google’s Nano Banana 2 and **61.9** for OpenAI’s GPT-5.4 Image 2, but the recorded verdict is a tie: each model won four tasks, with no task ties. The narrow score gap does not override that split, so neither takes the matchup overall. Google’s strongest results came when composition and format were central. It better matched the Bauhaus transit-room brief, the restricted four-color campsite, the perspective scene and the object-free reading nook; recurring advantages included the requested wide framing and closer adherence to visual constraints. Its weaknesses were concrete, too: signage was garbled in the transit scene, and it appeared to include ten gears where exactly nine were required. OpenAI took the anxious curator, salt-flat convoy, legible-text poster and tungsten-gears tasks. It made the curator’s panic more immediate, organized the convoy with stronger depth, presented the requested text cleanly, and got the gear count right. But square framing cost it fidelity on several prompts that specified 16:9, even when the underlying scene was otherwise convincing. **Final call: Too close to call (not statistical).** The scores lean slightly toward Google, while the task wins are evenly split and each model’s misses matter. This assessment applies only to the saved task sample; it is not a claim of general capability superiority.
Bauhaus transit control room
A night-shift rail operations manager at a regional train control desk, rendered in strict Bauhaus style: bold geometric forms, flat planes of red, yellow, blue, black, and cream, asymmetrical but balanced layout, sans-serif signage, and reductive industrial design. The scene should still clearly depict a real person at work monitoring route diagrams and timetables, with crisp circles, rectangles, and diagonal track graphics integrated into the composition; frontal wide view, poster-like clarity, clean matte lighting, 16:9.
OpenAI: GPT-6 Luna: Model A more fully captures the strict Bauhaus poster style, wide frontal control-room view, and a clearly working rail manager, though some signage is garbled. Model B has a convincing worker and readable information, but its realistic rendering and square format depart more from the prompt. (Second judge pass, order swapped — scores are the average of both: Model A more fully captures the strict Bauhaus brief with flat geometric color blocks, crisp route graphics, and a clear wide control-room | Anthropic: Claude Fable 5.1: Model A delivers the requested 16:9 poster in a…
Anxious restoration curator
A museum paper conservator in a midnight-blue lab coat stands alone at a long workbench, realizing a tiny illuminated manuscript page has just begun to tear under her tweezers; her face must unmistakably convey sharp, immediate panic mixed with professional restraint—eyes widened, jaw tense, breath caught, brows drawn upward—while her hands hover frozen above the damage. Show the conservation studio around her with Japanese tissue, wheat-starch paste jars, magnifying lamp, and cotton gloves, lit by cool overhead fluorescents plus one harsh task lamp creating dramatic highlights on her expression; tight medium shot, cinematic realism, shallow depth of field, 16:9.
OpenAI: GPT-6 Luna: Model B makes the curator’s immediate panic unmistakable and uses a tighter, more cinematic composition focused on her face and hovering hands. Model A better conveys the full studio setting and matches the requested widescreen format, but its expression and moment of damage feel less urgent. (Second judge pass, order swapped — scores are the average of both: Model B better captures the requested tight, frozen moment, with especially clear panic and hovering hands above the manuscript. Model A s | Anthropic: Claude Fable 5.1: Model A honors the 16:9 frame, includes the mag…
Salt-flat survey convoy
A wide 16:9 scene of a geospatial survey team working on a blinding white salt flat at sunrise, using a laser scanner beside a utility truck in a perfectly consistent one-point perspective: in the foreground, a bright orange survey prism pole and a kneeling field engineer; in the midground, two standing technicians and stacked equipment cases; far in the distance, the truck and a line of marker flags shrinking naturally toward the horizon. Relative scale must feel physically correct—near objects substantially larger than distant ones—with straight, coherent perspective lines and no warped proportions; photorealistic, low golden light, razor-sharp atmospheric clarity.
OpenAI: GPT-6 Luna: Model B better conveys the intended depth, with a prominent foreground surveyor and pole, a working team and cases in the midground, and a naturally diminished truck and flags near the horizon. Model A matches the requested wide format, but its truck is too close and visually dominant; Model B’s main miss is that it is square rather than 16:9. (Second judge pass, order swapped — scores are the average of both: Model B better matches the requested depth and scale, with the kneeling engineer and o | Anthropic: Claude Fable 5.1: Model A delivers the requested wide 16:9 photor…
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 more fully meets the prompt with a balanced 16:9 campsite scene and a cohesive four-color flat-vector palette. Model B has a clean graphic style, but its square format and visible tonal shading make it less faithful to the restrictions. (Second judge pass, order swapped — scores are the average of both: Model A better matches the requested 16:9 flat-vector campsite scene and has a more balanced, richly composed dusk landscape. Model B has a clear campsite motif, but it is square and uses | Anthropic: Claude Fable 5.1: Model A delivers the required 16:9 format with…
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 the exact three lines as a direct, centered poster with a deep navy, subtly grainy background and strong typographic hierarchy. Model A renders the text accurately too, but the photographed mockup and surrounding pale background make it feel less like the clean minimalist poster requested. (Second judge pass, order swapped — scores are the average of both: Model B presents all three correctly spelled lines crisply, centered on a deep navy, subtly grained background, with the req | Anthropic: Claude Fable 5.1: Both render all three lines correctly and legib…
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 closely matches the requested 16:9 framing while delivering a clear, balanced one-point perspective, warm light, and convincing scale recession. Model B also has strong perspective and detail, but its square framing departs from the prompt’s aspect ratio. (Second judge pass, order swapped — scores are the average of both: Model A better matches the requested 16:9 format while maintaining a convincing centered vanishing point, warm overhead lighting, and consistent depth and book sca | Anthropic: Claude Fable 5.1: Model A delivers the requested 16:9 frame, warm…
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 follows the requested 16:9 framing and includes the cozy chair, books, tea, and warm window light without any forbidden objects. Model B is also polished and respects the exclusions, but its square framing misses the requested aspect ratio. (Second judge pass, order swapped — scores are the average of both: Model B and Model A both include the cozy reading essentials and avoid plants, lamps, and artwork, but Model A follows the requested 16:9 format and has a well-balanced wide compositi | Anthropic: Claude Fable 5.1: Both Model A and Model B successfully omit plan…
Nine tungsten gears
An isometric 3D render of a precision watchmaking bench where a technician in finger cots is arranging EXACTLY 9 separate tungsten gears on a black anti-static mat for inspection; every gear must be fully visible, distinct, and individually countable, with no overlaps, no extra parts that could be mistaken for gears, and varied diameters to test counting accuracy. Include a loupe, brass caliper, notebook with sketched measurements, and a small desk lamp casting crisp shadows; ultra-clean workshop detail, cool gray palette with metallic highlights.
OpenAI: GPT-6 Luna: Model B shows exactly nine separate, countable gears and includes the finger-cotted technician, loupe, caliper, measurement notebook, and desk lamp. Model A is richly detailed, but it appears to show ten gears on the mat, violating the exact-count requirement. (Second judge pass, order swapped — scores are the average of both: Model B presents exactly nine distinct, fully visible gears on the mat and includes the requested bench tools, notebook, finger cots, and lamp. Model A is richly detailed, | Anthropic: Claude Fable 5.1: Model B places exactly nine clearly separated t…
Matchup powered by OpenRouter.