Head to head: Wan vs Google: Nano Banana 2 (Gemini 3.1 Flash Image)
Wan vs Google: Nano Banana 2 (Gemini 3.1 Flash Image)
By Ryan Merket · Published
Wan and Google’s Nano Banana 2 face off across eight image-generation prompts, from a painterly dawn hallway to a chaotic pantry accident. The comparison tests atmosphere, composition, prompt detail, and expression across very different scenes.
Google’s Nano Banana 2 (Gemini 3.1 Flash Image) leads the saved task sample, with aggregate scores of **64.9 to Wan’s 60.5**. It won five tasks to Wan’s two, with one tie; the editorial assessment is a **clear result (not statistical)**. Google’s edge came from scenes where atmosphere and multiple prompt details had to work together. It was stronger on the painterly apricot-hallway dawn and the ukiyo-e wave, and more fully realized both the pantry avalanche and rainy repair bench. It also won the reading-nook task, though both models missed the strict no-plants instruction by showing greenery outside the window. Wan had the sharper moments in controlled composition and expression: its library aisle delivered cleaner one-point perspective, and its locksmith portrait better captured a restrained, almost-breaking laugh. The tabletop attribute-binding task was a tie; the judges found strengths in both images’ object placement, while noting shortcomings in their respective settings. **Call: Google’s Nano Banana 2 wins this saved task sample, 64.9 to 60.5.** The verdict applies to these eight prompts, not to image generation in general.
Apricot hallway dawn
Painterly realism interior scene, 16:9 — at exactly 6:12 a.m., first golden-hour light pours through a narrow apartment hallway and strikes a crowded shoe rack: dusty patent-leather loafers, a pale linen grocery tote, a ribbed glass umbrella stand with two umbrellas still dripping, and a small aluminum parcel locker key tagged 5B; the low sun cuts long warm bands across cool blue walls and a terrazzo floor, catching floating dust and turning the scene quietly hopeful, composed from knee height looking down the hall with the far doorway softly blown out.
OpenAI: GPT-6 Luna: Model B better captures the painterly, quietly hopeful mood, with strong warm light bands, floating dust, cool blue walls, and a softly blown-out doorway. Model A includes more exact shoe-rack and parcel-locker details, but its lighting and overall treatment feel less atmospheric. (Second judge pass, order swapped — scores are the average of both: Model B more fully captures the painterly, quietly hopeful dawn, with long golden bands, visible dust, cool blue walls, and a softly blown-out far doo | Anthropic: Claude Fable 5.1: Model B delivers a more convincing knee-height…
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 directly fulfills the prompt with a long, empty, symmetrical aisle, a clear single vanishing point, and consistently diminishing shelves and books. Model B is also attractive and convincingly library-like, but its prominent repeated shelf bays make the central perspective feel less uninterrupted. (Second judge pass, order swapped — scores are the average of both: Model A has the cleaner, more consistent one-point perspective, with the shelves and overhead lights converging precisely | Anthropic: Claude Fable 5.1: Model A delivers exactly the requested long, em…
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 captures the bold indigo wave, cream foam, and dawn, but its boats feel awkwardly fragmented and the scene is less fully realized. Model B more convincingly evokes ukiyo-e with a coherent fishing crew, layered seascape, and strong print-like composition, though its palette is warmer than requested. (Second judge pass, order swapped — scores are the average of both: Model B more clearly depicts a fishing boat with people riding the cresting wave, and its dawn colors, bold outlines, and el | Anthropic: Claude Fable 5.1: Model B delivers a convincing ukiyo-e print wit…
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: Both images include visible greenery outside the window, violating the strict no-plants requirement, while otherwise showing the requested reading nook without lamps or artwork. Model B has a more complete, inviting reading setup with the side table, stacked books, tea, and open book, and its composition is especially polished. (Second judge pass, order swapped — scores are the average of both: Model A presents a clear, cozy reading nook with the required chair, book stack, tea, and warm window | Anthropic: Claude Fable 5.1: Both models correctly include the armchair, book…
Almost-laughing portrait
Photorealistic close portrait — a middle-aged locksmith in a cramped key-cutting shop is caught in the instant of trying not to laugh, the emotion unmistakably amused restraint: one eyebrow lifted, eyes creased and glinting, lips pressed together around a smile that is about to break, cheeks beginning to rise; shallow depth of field with pegboards of odd duplicate keys and a tiny handwritten sign reading RIVA 24 MIN in soft blur behind, gentle fluorescent shop light mixed with cool window fill, natural skin texture, and a tight off-center composition that prioritizes the face and expression.
OpenAI: GPT-6 Luna: Model A more precisely captures the almost-laughing restraint with pressed lips, creased eyes, and a tight off-center portrait, while clearly establishing the locksmith shop and the requested sign. Model B is compelling and photorealistic, but its smile is more openly formed and its wider composition gives less priority to the face. (Second judge pass, order swapped — scores are the average of both: Model A more clearly captures the restrained, almost-breaking laugh in a tight portrait, and its | Anthropic: Claude Fable 5.1: Model A nails the core brief — lips pressed toge…
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 more closely matches the clean, softly lit studio tabletop scene while clearly placing the duck on the green cube and the blue cylinder behind the other objects. Model B includes the requested objects and positions, but its busier indoor background and cloth-covered surface are less faithful to the prompt. (Second judge pass, order swapped — scores are the average of both: Model B clearly presents the green cube to the left of the glossy red sphere, with the blue cylinder behind them and | Anthropic: Claude Fable 5.1: Both bind colors correctly with the duck on the…
Dropped pantry avalanche
Hyperreal action still, 16:9 — a split second after a pantry shelf fails, an avalanche of household objects bursts outward toward the camera: a 3-kilogram sack of red lentils tearing open midair, cinnamon dust blooming, a green enamel colander spinning, clothespins scattering, a glass jar of pickled shallots tipping with brine droplets suspended, and a tabby cat springing sideways off a washing machine; dramatic frozen motion with sharp subject detail, motion streaks on the fastest objects, hard side light from a bare utility bulb, and a slightly low angle that amplifies the chaos.
OpenAI: GPT-6 Luna: Model B more completely captures the pantry avalanche, with a visibly failed shelf, a sideways-leaping cat, a labeled lentil sack, suspended brine, and scattered clothespins. Model A is clean and polished, but its cat is walking and several requested action details are less evident. (Second judge pass, order swapped — scores are the average of both: Model B captures more of the specified split-second chaos, including the leaping cat, bursting lentils, cinnamon cloud, spinning colander, scattered | Anthropic: Claude Fable 5.1: Model B delivers the explosive shelf-failure wi…
Rainy repair bench
Photorealistic editorial still life, 16:9 — a rainy-day kitchen repair bench improvised on a maple dining table: a brushed stainless moka pot opened beside a tiny pile of damp coffee grounds, a scratched brass watch gear, a cobalt-blue ceramic mug with a chipped rim, folded olive dish towel, translucent skin-toned fingertips holding a jeweler’s screwdriver just out of frame, and a clear puddle reflecting a window grid; cool overcast daylight from the left with a weak warm under-cabinet bulb from the back, creating believable mixed-color shadows, soft specular highlights on metal, saturated glaze on ceramic, visible wood grain, wet surfaces, and realistic fabric texture.
OpenAI: GPT-6 Luna: Model B more clearly conveys an active repair, with the moka pot opened and the screwdriver engaging it, while also including the chipped cobalt mug, damp grounds, towel, rainy window, and reflective puddle. Model A is polished and well-lit, but its moka pot is assembled and the repair elements feel less integrated. (Second judge pass, order swapped — scores are the average of both: Model B better captures the opened moka pot, damp grounds, rainy window, and puddle reflecting a window grid, with | Anthropic: Claude Fable 5.1: Model B nails more of the specific prompt detai…
Matchup powered by OpenRouter.