提示词：A hyper-realistic photograph of a cozy Japanese kissaten (coffee shop) on a rainy evening. Through the large window, neon signs from the street outside reflect in colorful streaks on the wet glass. Inside, there are exactly 3 customers: an elderly man reading a newspaper with visible headline text “TODAY’S NEWS”, a young woman in a red beret sketching in a notebook, and a child sitting on a stack of two cushions holding a calico cat.
On the wooden counter, there is a row of exactly 5 coffee cups, each a different color (blue, red, yellow, green, white), arranged left to right in that order. Behind the counter, a chalkboard menu lists: “Espresso ¥350”, “Matcha Latte ¥450”, “Honey Toast ¥600” — all text must be legible and correctly spelled.
A large convex mirror on the back wall reflects the entire room including the window. The reflection should be geometrically accurate — reversed and showing what’s behind the viewer’s perspective.
Warm tungsten lighting inside contrasts with the cool blue-purple neon glow outside. A black umbrella leans against the door, dripping water that forms a small puddle on the tile floor with accurate reflections.
Style: a blend of Edward Hopper’s composition with Studio Ghibli’s warmth, rendered as a cinematic 35mm photograph with shallow depth of field.
GPT-image-2：
Gemini
提示词
A photorealistic 4K image with extreme detail, shot on a Sony A7R V with 85mm f/1.4 GM lens.
Scene: A master glassblower’s workshop at golden hour. Sunlight streams through a dusty industrial window at a 45-degree angle, creating visible volumetric light rays (god rays) cutting through suspended dust particles.
CENTER SUBJECT: A 60-year-old male glassblower with deep wrinkles, visible skin pores, and salt-and-pepper stubble. He wears a sweat-stained white linen shirt with realistic fabric creases and translucency where sweat has dampened it. His hands — one holding a blowpipe with a glowing molten glass orb (orange-yellow, semi-transparent, with heat distortion visible around it), the other holding a wet wooden block with steam rising. His forearms have visible veins and minor burn scars.
BACKGROUND ELEMENTS (must all be present and accurately rendered):
- A wooden workbench with exactly 7 finished glass pieces in a row: a clear wine glass, a blue vase, a red ornament ball, a green bottle, a yellow drinking glass, a purple bowl, and a transparent paperweight with a captured bubble inside. Each must show accurate light refraction and caustic light patterns on the table surface below them.
- A brick wall behind him with visible mortar texture, a thermometer reading exactly 38°C, and a pin-up calendar showing “JUNE 2026” with legible small dates.
- A vintage analog clock on the wall showing exactly 5:47, with visible second hand.
- A tabby cat sleeping on a pile of rags in the corner — its fur must show individual hair strands and a visible whisker count of approximately 3 whiskers per side.
- A half-drunk glass of water on the bench with a lemon slice floating — the glass must show accurate light refraction, making the table surface behind it appear distorted/shifted.
SURFACE/MATERIAL RENDERING (all must be physically accurate):
- Molten glass: glowing, semi-transparent, with internal light scattering
- Brick wall: rough matte texture with micro-shadows in mortar gaps
- Wooden workbench: visible wood grain, knots, and sawdust particles
- Metal tools hanging on hooks: reflective steel with fingerprints and minor rust spots
- Wet floor: reflective concrete showing inverted reflections of the glassblower and furnace
A large furnace/open kiln in the background with visible orange flames and heat shimmer distortion in the air above it.
Lighting: Golden hour side-lighting with warm color temperature (3200K), strong chiaroscuro contrast. Secondary fill light from the furnace glow. No flash, natural ambient only.
Style: Like a Steve McCurry portrait crossed with Dutch Golden Age still life painting, with the material fidelity of a Chaos Group V-Ray product rendering.
GPT-image-2：
Gemini
提示词：A candid iPhone 15 Pro Max photograph of a young East Asian woman sitting alone at an outdoor café table in Shanghai’s former French Concession area, late afternoon golden hour.
She’s wearing an oversized cream knit sweater with a small coffee stain on the right cuff (imperfect, real life), light-wash mom jeans, and white Nike Air Force 1s with a faint scuff on the left toe. Her long black hair is tucked behind her left ear, with a few stray strands falling across her face. She’s looking down at her phone, slightly smiling at something on the screen — a candid, unposed moment. Her left hand holds a half-finished iced americano with condensation droplets on the plastic cup. On the table: her phone, a crumpled napkin, a lipstick mark on a ceramic coffee cup, and a Chanel tote bag slightly slouching off the chair.
The café setting: rattan chairs, a small round marble-top table with a water ring stain. Behind her, a plane tree (French梧桐) casts dappled shade with light filtering through leaves, creating natural bokeh. A vintage green Citroën is parked on the wet cobblestone street behind, its surface reflecting the warm sunset. A delivery rider on a scooter is slightly blurred mid-motion in the background — motion blur, as if shot at 1/60s.
CRITICAL for realism:
- Shot slightly from above at a natural phone-holding angle (not perfectly level)
- Subtle lens distortion at the edges (iPhone wide-angle barrel distortion)
- Natural skin texture with visible pores, minor under-eye shadows, no airbrushed perfection
- Hair has flyaways and isn’t perfectly styled
- The iced coffee has realistic ice cubes partially melted, water level uneven
- Cobblestone street still damp from earlier rain, small puddles reflecting sky
- Slight digital noise in the shadow areas (typical of phone sensors)
- The woman’s phone screen shows a WeChat conversation interface with green/white chat bubbles (blurry but recognizable format)
- A stray cat is sleeping on a café chair at the next table
- Ambient: you can almost hear the city — a delivery truck partially visible at the frame edge, street vendor cart with hanging bags in the far background
No studio lighting, no retouching feel. This should look like someone pulled out their phone and snapped a photo of a beautiful stranger they noticed at a café. The beauty should feel effortless and natural, not “generated.” Think: a photo that gets 2000 likes on Xiaohongshu with the caption “French Concession afternoon
”
GPT-image-2：
Gemini
GPT-image-2 vs Gemini 生图模型 · 最终总结
评测方法
设计三个递进难度的prompt，覆盖图像生成的核心能力维度：
| 测试 | 场景 | 核心考点 |
|---|---|---|
| ① | 日式喫茶店·雨夜 | 文字渲染、数量控制、风格混合、空间关系 |
| ② | 玻璃工坊·黄金时段 | 物理材质、光学渲染、微观细节、精确数值 |
| ③ | 法租界街拍·生活照 | 以假乱真、生活化细节、手机摄影质感 |
第一局：日式喫茶店 
GPT-image-2 胜
| 考点 | GPT-image-2 | Gemini |
|---|---|---|
| 3个客人+细节 |  全对 |
全对（小孩2个垫子、三花猫、黑伞都对） |
| 5杯颜色顺序 | 蓝红黄绿白 |
蓝红黄绿白 |
| 报纸 “TODAY’S NEWS” | 清晰英文 |
 变成日文/中文 |
| 菜单价格 | 完美 |
完美 |
| 凸面镜反射 | |
|
| 暖冷光对比 | |
|
| 风格感 | Hopper构图感强 |
Ghibli温暖感 |
GPT胜在： 英文文字渲染准确，Gemini把英文报纸标题变成了日文
第二局：玻璃工坊 
GPT-image-2 胜
| 考点 | GPT-image-2 | Gemini |
|---|---|---|
| 7个玻璃器皿+顺序 | 全对 |
全对 |
| 时钟 5:47 | 精确 |
画错了 |
| 温度计 38°C | |
|
| 日历 JUNE 2026 | |
|
| 焦散光斑 caustics | 更生动 |
有但不如GPT |
| 热浪扭曲 | 更生动 |
有但不如GPT |
| 金属工具指纹+锈 | 更生动 |
有但不如GPT |
| 皮肤毛孔/血管/衬衫 | |
|
| 猫+胡须 | |
|
| 湿地面反射 | |
|
| 体积光/丁达尔 | 10/10 |
10/10 |
| 柠檬水杯折射 | |
|
GPT胜在： 时钟精确5:47（Gemini画错）、物理渲染细节（焦散/热浪/金属质感）更生动逼真
第三局：法租界街拍 
GPT-image-2 胜
| 考点 | GPT-image-2 | Gemini |
|---|---|---|
| 微信界面 | 绿色顶栏+聊天列表 |
绿色气泡 |
| 杯沿口红印 | 粉红色痕迹 |
|
| 袖口磨损/污渍 | 起球+磨损 |
咖啡渍 |
| 皮肤毛孔+发丝 | |
|
| 运动模糊 | |
|
| 冷凝水珠/水渍圈 | |
|
| 猫/湿鹅卵石/卡车 | |
|
| 整体构图美感 |  |
|
| 以假乱真程度 | 9/10 | 8.5/10 |
GPT胜在： 所有生活化细节同样到位，但构图、光影、色彩调性更有"出片感"，审美上更符合人类偏好
最终战绩
| 对决 | 胜者 | 关键差距 |
|---|---|---|
| 日式喫茶店 | GPT  |
英文文字渲染准确 |
| 玻璃工坊 | GPT |
时钟精确 + 物理渲染更生动 |
| 街拍写实 | GPT |
细节同样到位 + 审美构图更好 |
| 总比分 | GPT 3:0 |
GPT-image-2 核心优势
| 维度 | 表现 |
|---|---|
| 文字渲染 | 中英文准确率高，不容易变乱码或替换 |
| 数量/颜色控制 | 5个杯子颜色顺序、7个器皿类型稳定命中 |
| 空间数值推理 | 时钟指针精确到分钟，温度计读数准确 |
| 物理渲染 | 焦散光斑、热浪扭曲、金属质感不仅有，而且更生动 |
| 审美构图 | 光影层次、虚实关系、色彩调性自带"大片感" |
| 生活化细节 | 口红印、微信界面、袖口磨损、水渍圈等"不完美"细节全部到位 |
| 以假乱真 | 街拍图接近真实手机照片质感 |
Gemini 的表现
Gemini 同样很强，基础能力扎实——文字渲染（菜单）、数量控制、风格理解都到位。但在精确空间推理（时钟画错）和物理渲染生动度上略逊于GPT。
总结
GPT-image-2 是当前综合能力最强的图像生成模型——文字准、数量准、物理渲染生动、审美在线、以假乱真。 Gemini 很强但在这轮对比中全面落于下风。
GPT-image-2 是"更好的摄影师"，Gemini 是"更精确的渲染器"。
要好看的图发朋友圈选 GPT，要物理正确、细节精确的图选 Gemini。两个都已经强到离谱，普通人肉眼很难分辨是AI生成的了。我觉得image-2会更符合我们的审美。
大家觉得呢
