{ "battle": "gemini-vs-gpt-image-images", "category": "Images", "suite_version": "images-2026-07", "contestants": { "A": { "name": "Nano Banana 2", "model": "google/gemini-3.1-flash-image" }, "B": { "name": "GPT Image", "model": "openai/gpt-5.4-image-2" } }, "judge": "anthropic/claude-sonnet-5", "protocol": "blind position-swapped double VISION judging; cross-position disagreement scored as tie; images saved as receipts", "proxy_note": "Midjourney has no public API, so it cannot be tested under this protocol. We substituted the two leading API-testable image models at test time: Nano Banana 2 (Gemini 3.1 Flash Image — the model behind Gemini's image generation) vs GPT-5.4 Image 2 (the model behind ChatGPT's image generation). This is NOT a Midjourney result; Midjourney is discussed on the page from verified public claims only.", "started_at": "2026-07-17T08:36:12.529Z", "finished_at": "2026-07-17T09:06:39.818Z", "scoreline": { "Nano Banana 2": 2, "GPT Image": 4, "ties": 4 }, "total_cost_usd": 3.0853, "tasks": [ { "suite": "product", "id": "p1", "name": "Product shot", "prompt": "A studio product photograph of a matte-black insulated coffee flask on a light grey seamless background, soft top-left key light, subtle reflection under the flask, no text, no watermark, e-commerce quality.", "image_a": "/data/battles/images/gemini-vs-gpt-image-images/p1-a.jpg", "image_b": "/data/battles/images/gemini-vs-gpt-image-images/p1-b.png", "latency_ms": { "a": 7402, "b": 133907 }, "cost_usd": { "a": 0.068461, "b": 0.293967 }, "winner": "A", "judge_scores": { "a": 8, "b": 7 }, "judge_reasoning": "Pass 1 (A first): Both meet the prompt well; Image 1 has slightly better lighting gradient and composition with more natural top-left key light, though flip-cap design resembles bottle more than flask. Image 2 is classic flask shape but flatter lighting. | Pass 2 (B first): Both fit the brief; image 2 looks more like a classic insulated flask with better lighting gradient and subtle reflection, though it resembles a water bottle with handle/straw. Image 1 is clean but shadow is harsher and lighting flatter." }, { "suite": "product", "id": "p2", "name": "Lifestyle product", "prompt": "A pair of white running shoes on a wet city pavement at dawn, shallow depth of field, natural morning light, photorealistic, no people, no text.", "image_a": "/data/battles/images/gemini-vs-gpt-image-images/p2-a.jpg", "image_b": "/data/battles/images/gemini-vs-gpt-image-images/p2-b.png", "latency_ms": { "a": 8537, "b": 192951 }, "cost_usd": { "a": 0.06846150000000001, "b": 0.29334450000000006 }, "winner": "tie", "judge_scores": { "a": 7, "b": 7 }, "judge_reasoning": "Pass 1 (A first): Image 1 shoes are dirty, not clean white as requested. Image 2 shows cleaner white shoes, nicer dawn light, wet reflective pavement, better composition and aesthetic appeal, though pavement is more concrete than asphalt. | Pass 2 (B first): Image 2 shows more convincing wet pavement texture, natural dawn lighting, and realistic photographic quality. Image 1 has smoother, slightly artificial-looking shoes and less convincing wetness. Both lack text/people as required. Image 2's Nike logo isn't legible text, acceptable." }, { "suite": "text", "id": "t1", "name": "Text rendering — poster", "prompt": "A minimalist concert poster with the exact text \"MIDNIGHT ORCHARD — 14 AUGUST — ROUNDHOUSE\" in a clean sans-serif, dark navy background, a single stylised pear illustration, Swiss design style.", "image_a": "/data/battles/images/gemini-vs-gpt-image-images/t1-a.png", "image_b": "/data/battles/images/gemini-vs-gpt-image-images/t1-b.png", "latency_ms": { "a": 13770, "b": 154873 }, "cost_usd": { "a": 0.06722300000000003, "b": 0.293022 }, "winner": "A", "judge_scores": { "a": 8, "b": 6 }, "judge_reasoning": "Pass 1 (A first): Image 1 has cleaner Swiss minimalism with flat pear icon and correct text hierarchy including em-dashes; Image 2's textured pear is less minimalist/Swiss and lacks em-dash format, though still visually appealing. | Pass 2 (B first): Image 2 has correct minimalist Swiss style, cleaner text layout matching exact requested text with em-dashes, and simpler stylised pear. Image 1's pear is textured/realistic, less Swiss-minimalist, and text is split awkwardly without dash format." }, { "suite": "text", "id": "t2", "name": "Text rendering — sign", "prompt": "A photorealistic wooden café sandwich board on a cobbled street reading exactly \"TODAY'S SOUP: TOMATO & BASIL £4.50\" in white chalk handwriting.", "image_a": "/data/battles/images/gemini-vs-gpt-image-images/t2-a.png", "image_b": "/data/battles/images/gemini-vs-gpt-image-images/t2-b.png", "latency_ms": { "a": 12135, "b": 183146 }, "cost_usd": { "a": 0.06721849999999996, "b": 0.2928465 }, "winner": "tie", "judge_scores": { "a": 7, "b": 7.5 }, "judge_reasoning": "Pass 1 (A first): Both accurately show text and cobblestone street. Image 1 has richer context, café setting, better lighting and composition. Image 2 is clean but simpler background, slightly less photorealistic charm. | Pass 2 (B first): Image 1 has exact matching text including all elements. Image 2 omits some text formatting and appears more cropped, though scene context is richer. Image 1 is cleaner and more accurate to prompt." }, { "suite": "illustration", "id": "i1", "name": "Flat illustration", "prompt": "A flat vector-style illustration of a woman watering a giant potted monstera plant on an apartment balcony, warm sunset palette (coral, amber, deep teal), clean shapes, no gradients, no text.", "image_a": "/data/battles/images/gemini-vs-gpt-image-images/i1-a.png", "image_b": "/data/battles/images/gemini-vs-gpt-image-images/i1-b.png", "latency_ms": { "a": 14441, "b": 167800 }, "cost_usd": { "a": 0.0672204999999999, "b": 0.29246649999999996 }, "winner": "B", "judge_scores": { "a": 6.5, "b": 8 }, "judge_reasoning": "Pass 1 (A first): Image 2 has cleaner flat shapes, no gradients, better cohesive palette matching prompt. Image 1 has gradient shading in sky/plants violating 'no gradients' constraint, though both similarly adhere to composition. | Pass 2 (B first): Image 1 has cleaner flat shapes, no gradients, better color palette matching prompt (coral, amber, teal). Image 2 has visible gradients in sky and lighting effects, violating 'no gradients' constraint." }, { "suite": "illustration", "id": "i2", "name": "Children's book", "prompt": "A children's picture-book illustration of a small red fox in a yellow raincoat sharing an umbrella with a hedgehog at a rainy bus stop, soft watercolour texture, gentle and warm, no text.", "image_a": "/data/battles/images/gemini-vs-gpt-image-images/i2-a.jpg", "image_b": "/data/battles/images/gemini-vs-gpt-image-images/i2-b.png", "latency_ms": { "a": 8742, "b": 219133 }, "cost_usd": { "a": 0.06849550000000004, "b": 0.2935955000000001 }, "winner": "B", "judge_scores": { "a": 5.5, "b": 8 }, "judge_reasoning": "Pass 1 (A first): Image1 includes text ('BUS STOP', signage) violating no-text constraint, capping its score. Image2 has no text, warmer composition, gentle watercolor style, and better matches prompt overall. | Pass 2 (B first): Image 1 has yellow raincoat as requested, gentle warm watercolor, no text. Image 2 has text ('BUS STOP') violating no-text constraint and umbrella is yellow not distinct from coat, less adherence overall despite charm." }, { "suite": "photoreal", "id": "r1", "name": "Photoreal portrait", "prompt": "A photorealistic portrait of an elderly fisherman with weathered skin mending a green fishing net at golden hour, 85mm lens look, sharp eyes, softly blurred harbour background.", "image_a": "/data/battles/images/gemini-vs-gpt-image-images/r1-a.jpg", "image_b": "/data/battles/images/gemini-vs-gpt-image-images/r1-b.png", "latency_ms": { "a": 9254, "b": 197345 }, "cost_usd": { "a": 0.06864650000000005, "b": 0.29402149999999994 }, "winner": "tie", "judge_scores": { "a": 8, "b": 7.5 }, "judge_reasoning": "Pass 1 (A first): Image 1 has sharper eyes, better golden hour lighting, richer harbour bokeh, more convincing net-mending action. Image 2 is decent but eyes less sharp, lighting flatter, hands less engaged with net detail. | Pass 2 (B first): Image 1 better matches the 85mm close-up look with softly blurred harbour background and sharp eyes, per prompt. Image 2 has sharp background boats, contradicting the requested blur, though technically strong and detailed." }, { "suite": "photoreal", "id": "r2", "name": "Photoreal scene", "prompt": "A photorealistic interior of a small Tokyo ramen bar at night, steam rising from bowls, warm lantern light, two empty stools in the foreground, rain visible through the window.", "image_a": "/data/battles/images/gemini-vs-gpt-image-images/r2-a.png", "image_b": "/data/battles/images/gemini-vs-gpt-image-images/r2-b.png", "latency_ms": { "a": 12003, "b": 202092 }, "cost_usd": { "a": 0.06721749999999993, "b": 0.2935274999999997 }, "winner": "B", "judge_scores": { "a": 6.5, "b": 8 }, "judge_reasoning": "Pass 1 (A first): Image 2 matches prompt better with empty foreground stools, steam, rain visible, warm lantern light. Image 1 has stools occupied by people, contradicting 'empty stools' requirement, though technically detailed. Image 2 has cleaner composition and mood. | Pass 2 (B first): Image 1 shows empty stools as requested with steam, rain, lanterns; Image 2 has occupied stools, contradicting 'empty stools' requirement, though it's atmospheric and detailed. Image 1 better adheres to prompt with more coherent composition." }, { "suite": "diagram", "id": "d1", "name": "Infographic", "prompt": "A clean business infographic showing a 4-step process labelled exactly \"1. Capture 2. Sort 3. Review 4. Ship\" as connected rounded boxes, left to right, corporate blue and white palette, flat design.", "image_a": "/data/battles/images/gemini-vs-gpt-image-images/d1-a.jpg", "image_b": "/data/battles/images/gemini-vs-gpt-image-images/d1-b.png", "latency_ms": { "a": 8063, "b": 147654 }, "cost_usd": { "a": 0.0689194999999998, "b": 0.2953554999999999 }, "winner": "tie", "judge_scores": { "a": 7, "b": 7.5 }, "judge_reasoning": "Pass 1 (A first): Image 1 has richer, more polished corporate design with icons, arrows, and clear labeling matching the exact text; Image 2 is cleaner but simpler, less detailed, with plain title-less layout. Both meet text requirement, but Image 1 offers stronger visual appeal and completeness. | Pass 2 (B first): Image 1 matches the clean, minimal flat-design request precisely with correct text, simple icons, and clear arrows. Image 2 is more cluttered, adds extra unrequested text/title, deviating from 'clean' minimal brief despite polished visuals." }, { "suite": "diagram", "id": "d2", "name": "Instructional", "prompt": "A simple instructional diagram showing how to fold a paper crane in 6 numbered steps, line-art style, black lines on white, each step clearly separated in a 3x2 grid.", "image_a": "/data/battles/images/gemini-vs-gpt-image-images/d2-a.jpg", "image_b": "/data/battles/images/gemini-vs-gpt-image-images/d2-b.png", "latency_ms": { "a": 7169, "b": 149209 }, "cost_usd": { "a": 0.0684499999999999, "b": 0.29407799999999984 }, "winner": "B", "judge_scores": { "a": 6.5, "b": 7.5 }, "judge_reasoning": "Pass 1 (A first): Both show 3x2 grids with numbered steps in clean line-art. Image 2 has clearer, cleaner progression resembling actual crane folding sequence and ends with recognizable crane. Image 1's step 6 shows two separate cranes, less consistent with single sequence. | Pass 2 (B first): Both show clean 3x2 grids with line-art steps and final crane. Image 1 has clearer, more logical progression and cleaner final crane shape. Image 2's final step shows two cranes which slightly deviates from single instructional flow." } ] }