Head to head: Google: Nano Banana 2 (Gemini 3.1 Flash Image) vs Krea 3 Edit
Google: Nano Banana 2 (Gemini 3.1 Flash Image) vs Krea 3 Edit
By Ryan Merket · Published
This image-model matchup tests more than visual polish: the prompts demand precise object placement, spatial layouts, lighting, and detailed scene direction. The verdict turns on how consistently each model handles those constraints across varied tasks.
Google’s Nano Banana 2 takes the saved task sample, scoring **65.0** to Krea 3 Edit’s **56.0** and winning all **8** task verdicts, with no ties. That’s a **decisive result for this sample**, not a statistical claim or a general ranking of the models. The clearest separation came on composition-heavy prompts. Nano Banana 2 better placed objects in the attribute-binding and bedroom-layout tasks, and more faithfully followed the technical stage blueprint and high-corner rehearsal-hall view. Krea often included the requested elements, but its perspective, layout choices, or prominence of key objects weakened the match. It also had an edge in scene direction: its rainy loading-dock quartet felt more candid and convincingly wet, while its backstage portrait better captured the requested relieved expression and split lighting. On the teapot-and-reflections task, neither model nailed every reflection requirement, but Nano Banana 2 produced the more plausible teapot and made the apple reflection clearer. Both also showed foliage despite the strict no-plants instruction in the reading-nook task. Krea’s outputs were not without strengths, and some individual details favored it; the overall verdict reflects the saved task outcomes, not a claim that it is broadly inferior. **Call: Nano Banana 2 wins this head-to-head decisively on the saved task sample, especially when the prompt demands precise composition and faithful scene direction.**
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 clearly places the yellow duck on the matte green cube, with the cube left of the glossy red sphere and the blue cylinder behind them, in a convincing softly lit tabletop scene. Model B follows the arrangement too, but the sphere has a distracting duck-like reflection or artifact and the scene feels less photographic. (Second judge pass, order swapped — scores are the average of both: Model A more clearly places the blue cylinder behind the cube and sphere, and its tabletop textures and | Anthropic: Claude Fable 5.1: Both place the correct colors on the correct obj…
Rainy loading-dock quartet
Photorealistic backstage scene after a club set: at 2:13 a.m. on a rain-slick loading dock behind the fictional venue Meridian Coil, a tired jazz quartet packs up under a buzzing sodium-vapor wall light and a cool spill from an open doorway; a brushed brass tenor sax on a dented road case, a chrome snare stand, wet black leather boots, dark walnut bass neck, red nylon instrument straps, damp denim jackets, and realistic skin with light perspiration must all show physically believable reflections, roughness, and shadow behavior; low eye-level documentary photography, shallow puddles mirroring the figures, natural color, 16:9.
OpenAI: GPT-6 Luna: Model A more convincingly captures a candid, rain-soaked loading-dock teardown, with distinct quartet members, the requested instruments and gear, and strong warm-versus-cool lighting reflected in the puddles. Model B includes the venue name and several relevant details, but feels more posed and less convincingly wet, with a weaker sense of documentary realism. (Second judge pass, order swapped — scores are the average of both: Model B includes more of the specified gear, especially the saxophon | Anthropic: Claude Fable 5.1: Model A delivers a convincing documentary frame…
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 A closely matches the requested isometric bedroom layout, with all four objects clearly placed as specified and a cohesive flat-vector finish. Model B includes the requested objects, but its more frontal, less consistently isometric perspective and sparse presentation weaken the match. (Second judge pass, order swapped — scores are the average of both: Model B places all four requested items clearly, including the lamp in the front-right corner, and best matches the clean flat-vector style | Anthropic: Claude Fable 5.1: Model B places every element where the prompt a…
Relief after the encore
Cinematic close portrait of a young flamenco singer in a narrow side-stage corridor the instant after an overwhelming encore, conveying unmistakable relieved joy mixed with disbelief: eyes glossy with tears, mouth open in a shaky half-laugh, shoulders dropped, one hand pressed to the sternum, the other still clutching a worn ivory fan; smudged stage makeup, embroidered charcoal traje corto jacket, warm tungsten spill from the stage on one side and cool fluorescent backstage light on the other, shallow depth of field, intimate 85mm photography, 16:9.
OpenAI: GPT-6 Luna: Model A more convincingly captures relieved, disbelieving joy and follows the requested costume, fan, hand placement, corridor, and split warm-cool lighting. Model B includes the key gestures and tears, but its heavy orange cast and more distressed expression make the mood and lighting less faithful. (Second judge pass, order swapped — scores are the average of both: Model A more clearly captures the charcoal embroidered jacket, backstage corridor, ivory fan, tearful disbelief, and split warm-co | Anthropic: Claude Fable 5.1: Model A nails nearly every prompt beat—tear-glo…
Blueprint stage layout puzzle
Clean blueprint-style technical drawing of a small experimental music stage, white linework on deep cyan, perfectly organized top-down plan view: a hexagonal drum riser sits exactly in the center; a vibraphone is directly left of the riser; a double bass is behind the riser; a vocal mic stand is on top of the front edge of the riser; a pedalboard is between the vibraphone and the riser; two wedge monitors are symmetrically in front of the riser; a folding stool is right of the riser; a cable reel is behind the stool but left of the double bass; and a narrow fog machine is in the back-right corner of the stage; precise spatial placement, labeled like a technical production diagram.
OpenAI: GPT-6 Luna: Model A follows the requested layout more closely, especially with a centered hexagonal riser, clearly labeled instruments, and paired front monitors; Model B uses a rectangular riser and mislabels the stool. Model A also reads more like a detailed technical production drawing, while Model B leaves much of the canvas unused. (Second judge pass, order swapped — scores are the average of both: Model A more closely follows the requested layout, especially with a centered hexagonal riser and clearly | Anthropic: Claude Fable 5.1: Model A delivers a dense, legible technical dia…
Rehearsal hall crane view
Ultra-detailed architectural interior rendering of a contemporary orchestra rehearsal hall seen from a high corner balcony, with strict one-point perspective toward the far conductor podium: nearest to camera a giant suspended line-array speaker and lighting truss, midground rows of empty black chairs with music stands, farther back a chamber organ console, and at the vanishing end a single conductor on the podium, all with correct relative scale so the grand piano, timpani, cello cases, exit doors, and ceiling coffers feel proportionally accurate and unwarped; crisp afternoon daylight from tall clerestory windows, 16:9.
OpenAI: GPT-6 Luna: Model A closely matches the requested high-corner view and prominently includes the giant foreground line-array speaker, lighting truss, empty orchestra setup, organ, and distant conductor. Model B includes many of the same elements, but the speaker is much less prominent and the hall feels less detailed and less convincingly composed toward a far podium. (Second judge pass, order swapped — scores are the average of both: Model A more fully realizes the prompt with a prominent foreground line-ar | Anthropic: Claude Fable 5.1: Model A nails the brief: a true one-point persp…
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 A has a more natural teapot, convincing kitchen reflections, and a balanced close-up composition; Model B has dramatic lighting, but its teapot shape is less plausible. Neither clearly shows both the window and apple correctly reflected on the water’s surface, though the apple reflection is more apparent in Model A. (Second judge pass, order swapped — scores are the average of both: Model A presents a more clearly recognizable chrome teapot and convincingly reflects the window and red appl | Anthropic: Claude Fable 5.1: Model B delivers the requested dramatic hard si…
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 a more inviting, richly detailed reading nook, with a clear stack of books and a mug beside the chair; Model B is attractive but sparser, with the mug perched on the books. Both avoid lamps and artwork, but both show foliage through the window, which conflicts with the strict no-plants instruction. (Second judge pass, order swapped — scores are the average of both: Model B and Model A both include the requested nook elements and avoid indoor plants, lamps, and artwork; foliage is vis | Anthropic: Claude Fable 5.1: Both outputs correctly include the armchair, bo…
Matchup powered by OpenRouter.