Head to head: OpenAI: GPT-5.4 Image 2 vs Google: Nano Banana Pro (Gemini 3 Pro Image)
OpenAI: GPT-5.4 Image 2 vs Google: Nano Banana Pro (Gemini 3 Pro Image)
By Ryan Merket · Published
A task-by-task comparison of two image models across text rendering, stylistic control, composition, and prompt constraints. The results hinge less on broad image quality than on which details each model gets right.
The aggregate scores are **62.9 for GPT-5.4 Image 2 and 60.6 for Nano Banana Pro**—a narrow gap, not a convincing separation. At **limited confidence** that either model is genuinely better, this is a sample tie. Calling a winner would claim more than the results support. GPT-5.4 Image 2 has the edge when the brief depends on finish and specificity: it produced the more balanced multi-line text, the more convincing ukiyo-e print, and the stronger instrument-restorer portrait. It also earned the archive-corridor verdict, though that comparison exposed a weakness of its own: square framing and repeated floor labels undercut an otherwise strong vanishing point. Nano Banana Pro did better on several prompts where following the requested frame and scene mattered most. Its widescreen hotline image and Aegean pastry poster better matched the requested 16:9 format; it also handled the negation prompt more faithfully. The restricted-palette campsite was a tie, with the models trading polish against stricter adherence to format and palette. The pattern is useful, but not decisive: GPT-5.4 Image 2 more often excels at visual refinement, while Nano Banana Pro more often gets the requested composition and constraints into frame. **Final call: a tie. These models are effectively even here; choose by task, not by the 2.3-point score gap.**
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 A presents all three lines crisply and correctly, with clear hierarchy, centered spacing, and a restrained navy-and-grain treatment. Model B also renders the text correctly, but its oversized middle line and heavier, slightly blurred lettering make the layout feel less balanced and polished. (Second judge pass, order swapped — scores are the average of both: Both models render the requested wording correctly on a deep navy, subtly textured background. Model A has a more refined, balanced p | Anthropic: Claude Fable 5.1: Both render all three lines correctly, but Mode…
Relief on the call floor
A cinematic editorial photograph, 16:9, of a night-shift emergency translator named Imani Vale in a municipal language hotline office at 2:13 a.m., the instant she realizes a missing child has just been found safe; her expression must read unmistakably as overwhelming relief mixed with exhausted joy, with wet eyes, loosened jaw, and a shaky half-laugh, one hand still on a headset, the other hovering above a notepad covered in hurried Amharic and Spanish scribbles; cool monitor glow on her face, sodium-vapor city light through rain-streaked windows, shallow depth of field, desks and coworkers softly blurred behind her.
OpenAI: GPT-6 Luna: Model A captures the relief vividly and includes the exact time, identity, headset, and handwritten notes, but its square framing misses the requested 16:9 format. Model B delivers the widescreen rainy-night setting and a convincing tearful, relieved expression, though its notes and identity are less specific. (Second judge pass, order swapped — scores are the average of both: Model B more completely captures the rainy night-shift hotline setting, with wet eyes, a headset, hurried notes, coworke | Anthropic: Claude Fable 5.1: Model A nails the relieved tear-streaked smile…
Named art style
A ukiyo-e woodblock print of a fishing boat riding a large cresting wave at dawn, faithful to the flat color planes, bold outlines, and stylized foam of the tradition, muted indigo and cream palette.
OpenAI: GPT-6 Luna: Model A more closely matches the requested muted indigo-and-cream ukiyo-e print, with textured print-like surfaces, bold contours, stylized foam, and a dawn horizon. Model B is also compelling and crisp, but its vivid red sun and cleaner, more modern finish stray slightly from the specified restrained palette and traditional print feel. (Second judge pass, order swapped — scores are the average of both: Model A more convincingly evokes a traditional ukiyo-e print, with stylized foam, a fishing b | Anthropic: Claude Fable 5.1: Model A delivers a more authentic ukiyo-e print…
Instrument restorer hands check
A hyper-detailed studio photograph of a violin restorer at work, shown from mid-thigh up at a maple workbench, clearly visible both hands delicately resetting a tiny ivory saddle onto a 1912 Mirecourt violin while her left thumb stabilizes the instrument and her right hand holds fine tweezers; all ten fingers must be anatomically correct and naturally posed, with realistic wrists, forearms, shoulders, and body proportions; resin dust, magnifying loupe, spool clamps, and a labeled jar reading 'Berenz 4B hide glue' on the bench; soft north-window lighting, crisp focus, neutral background.
OpenAI: GPT-6 Luna: Model A better matches the requested studio portrait, with both hands clearly engaged, a detailed violin-restoration setting, and the labeled glue jar; Model B has a convincing delicate repair action but crops the restorer’s face and offers less of the requested portrait composition. Both include recognizable tools and readable glue labeling, though neither makes the tiny saddle especially unmistakable. (Second judge pass, order swapped — scores are the average of both: Model A more fully matche | Anthropic: Claude Fable 5.1: Model A satisfies more of the brief—figure show…
Aegean six-color pastry poster
A vibrant flat-vector promotional poster, 16:9, showing a pastry chef at work glazing eight pear-and-saffron tarts in a compact bakery kitchen, restricted exactly to the named 'Aegean Harbor' palette and no other colors: deep navy, sea-teal, sun-bleached cream, terracotta, olive green, and muted coral; use only those six palette colors for every element including skin, utensils, shadows, highlights, steam, and background tiles, with no stray black, white, gray, or off-palette hues; bold geometric shapes, clean outlines, side lighting from a single window, balanced composition.
OpenAI: GPT-6 Luna: Model A depicts the chef glazing eight tarts in a polished, palette-coherent poster, but its square format misses the requested 16:9 layout. Model B matches the widescreen format and clearly presents the bakery scene, eight tarts, readable promotional text, and a cohesive restricted palette. (Second judge pass, order swapped — scores are the average of both: Model B follows the requested 16:9 format and clearly depicts a chef glazing eight pear tarts in a geometric bakery kitchen using the named | Anthropic: Claude Fable 5.1: Model B nails the core constraints: a true 16:9…
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 is polished and warmly lit, but its square framing misses the requested 16:9 format. Model B matches the wide format and includes the requested nook elements while avoiding plants, lamps, and artwork, though its chair looks worn and its composition is less refined. (Second judge pass, order swapped — scores are the average of both: Model B better matches the requested 16:9 framing and includes the chair, books, tea, window, and warm light; Model A is more polished but is square rather th | Anthropic: Claude Fable 5.1: Both outputs honor the negations (no plants, la…
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 B follows the requested 16:9 format and presents a cohesive flat-vector campsite scene using the restricted palette. Model A has appealing detail, but its square format and subtly graded sky depart from the prompt. (Second judge pass, order swapped — scores are the average of both: Model B follows the four-color flat-vector brief and 16:9 format closely, with a balanced campsite scene. Model A is polished, but its square format and visible tonal variation conflict with the specified aspect | Anthropic: Claude Fable 5.1: Model A holds the four-color palette strictly w…
Archive corridor perspective
A precise architectural illustration, 16:9, of a subterranean film archive where a conservator pushes a small ladder cart down a long central corridor lined with identical steel canisters, emphasizing rigorous one-point perspective with a clearly consistent vanishing point at the far end; include oversized refrigeration ducts crossing overhead, floor markings numbered A17 through A29 shrinking correctly with distance, and a forklift parked in a side bay that reads believably smaller than the foreground cart; nothing warped, no impossible shelf angles, accurate relative scale between the worker, cart, doors, ducts, and canisters; clean contemporary technical-render style with cool white industrial lighting.
OpenAI: GPT-6 Luna: Model B better matches the requested 16:9 technical-illustration style and clearly shows the ladder cart, overhead ducts, and a forklift in a side bay. Model A has a particularly strong central vanishing point, but its square format and repeated, out-of-order floor labels weaken adherence. (Second judge pass, order swapped — scores are the average of both: Model B better matches the requested 16:9 technical illustration and clearly includes a conservator pushing a ladder cart, while Model A has | Anthropic: Claude Fable 5.1: Model A delivers a convincing one-point perspect…
Matchup powered by OpenRouter.