Nano Banana 是 Google 的图像模型,可以根据文字和参考图生成、编辑图片(Google 的 Nano Banana 文档)。让图片出彩的并不是更长的提示词,而是先想清楚这张图用来做什么,明确告诉模型每张参考图各自承担什么任务,并在结果成为下一张图的输入之前先检查一遍。Bald Rhino 的智能体就是这样使用它的,下面每一张图都来自真实的制作项目。提示词按发送时的原文(英文)展示。
-
霓虹字 · Nano Banana Pro Cinematic 16:9 night exterior shot of a cozy coffee shop, viewed from the rainy street looking in through a large rain-streaked window… a neon sign glows in cursive-script lettering reading exactly "MOONBEAN"… -
美食广告 Premium cinematic food commercial still, 16:9. A perfect golden, flaky croissant on a rustic plate on a wooden bakery counter, fresh steam rising, flakes of pastry catching warm backlight… -
产品图 Create a premium product photo for a reusable water bottle ad… clean negative space above the bottle for headline text. The bottle is plain: no lettering, numbers, logos or watermarks. -
动画电影剧照 Still from a high-end stylized 3D animated feature film. A 10-year-old girl named Lia with big expressive brown eyes, warm brown skin, a voluminous cloud of curly black hair… -
同一角色,新场景 Using Image 1 only for Lia's character design… and the stylized 3D animated feature film rendering: create a new scene. Lia kneels on a mossy path in an enchanted forest at golden hour… -
带精确文字的封面图 YouTube thumbnail, 16:9… huge bold yellow text with a thick black outline reads exactly: "EAT FOR $5"… text readable at small size. -
绘本关键帧 Create a charming 2D cartoon storybook-style keyframe, wide shot: two cartoon owls side by side on a thick gnarled tree branch at night… -
大理石胸像 A classical white marble portrait bust of Socrates: snub nose, broad bald dome, bushy curly beard, wide-set amused eyes, a faint knowing half-smile… -
角色设计肖像 Character design portrait, head and shoulders, of CHIEF KUI of the Tufang grassland peoples, northern frontier of Shang China c. 1250 BCE… -
竖屏特写 Vertical 9:16 close-up of a craftsman's weathered hands holding a small bronze graver, its tip pressed against the edge of a turtle shell… -
封面背景 Create a bright, appetizing food photograph for a YouTube thumbnail background: three loaded tacos with seasoned beef, shredded lettuce, tomato, cheese… -
短提示词 A paper boat floating on a rain puddle at dusk, warm street lights reflected in the water, cinematic still -
扁平矢量漫画 Create a bold flat vector cartoon illustration for a vertical social video: two mathematicians in a dramatic face-off on either side of a big blackboard… -
历史动画片 Create an illustration for an animated history film: a lone scholar in a cream robe seen from behind on a flat rooftop in Alexandria at night… -
岩画风格 Cave-painting-style illustration: a crowd of black stick figures raising arms toward the sky, rain streaks, rice fields, a small war party, on warm ochre… -
宣纸水墨 Full-bleed warm off-white xuan rice paper with the same paper colour and texture as Image 1. In the centre, a single large drop of black ink has landed… -
编辑成空场景 Edit Image 1 into an empty set plate. Keep the exact camera angle, the square wooden table, the steaming red-and-white split hotpot… Remove all people completely…
1. 围绕图片的用途来构图
动笔之前,先决定观众第一眼应该注意到什么,以及这张图必须传达什么:一张认得出的脸、产品的某个卖点、两个角色之间的关系、一个地点,或者一个讲解。先描述这个看得见的情境,再写让它一目了然的构图、位置、光线、材质和处理方式。
几个习惯最能拉开差距:
- 一张图只表现一个可见的瞬间。“她躺下、站起来、跑开”是三个状态。选出这张图需要交代的那一个。
- 先定裁切,再构图。为视线、动作或之后要压在上面的标题留出空间,并确保重要细节在图片实际展示的尺寸下依然清楚。
- 描述你想要什么,而不是要避免什么。“下班后空无一人的办公室”比一串“不要出现什么”的清单更有用。
- 化解风格冲突。绘画质感的表面和完整渲染的 CG 场景会互相拉扯,除非你说明两者如何结合。
Create a premium product photo for a reusable water bottle ad: a single insulated stainless steel bottle, matte brushed silver steel body with a deep teal silicone grip band near the base and a teal screw cap with a small carry loop, standing upright on a wet slate surface, covered in fine condensation droplets to suggest ice cold. Centered, slightly low angle, 85mm lens, shallow depth of field, soft cool morning light with a rim light, pale aqua gradient background, clean negative space above the bottle for headline text. The bottle is plain: no lettering, numbers, logos or watermarks.
-
for a reusable water bottle ad
用途放在最前面:这是广告,产品必须一眼就能看懂。
-
covered in fine condensation droplets to suggest ice cold
要卖的卖点(保冷)是画出来的,而不是写出来的。
-
clean negative space above the bottle for headline text
先定裁切再出图:广告标题会放在那里。
-
no lettering, numbers, logos or watermarks
文字之后用准确的字体再加,所以要求模型让瓶身保持干净。
2. 每张参考图只承担一项任务
只附上这次输出需要的图片,并说明每张图贡献什么:
| 用作……的参考图 | 告诉模型…… |
|---|---|
| 身份或产品 | 要保留哪张脸、轮廓、标记、比例、服装或产品细节。人物所在的房间或光线不必一起带过来。 |
| 风格 | 要借用的渲染方式、材质、配色或光线,与图中的主体分开说明。 |
| 地点和构图 | 版式、机位和重要物体;是保留这个视角,还是在同一地点拍一个新镜头。 |
| 姿势 | 要迁移的姿态或手势,同时保留已确认的人物身份。 |
| 要编辑的图片 | 哪些元素要改;其他参考图只为这次修改服务,而不是替换整张图。 |
按附上的顺序给参考图编号(“图 1”“图 2”),并在每个名字旁边加一句看得见的简短描述:左边沙黄色的恐龙,右边灰色毛茸茸的哺乳动物。画面中有多个角色时,只写名字很容易混淆。如果两张参考图互相矛盾,就决定由哪一张来管脸、服装、风格或背景。
Using Image 1 only for Lia's character design (the girl with the voluminous curly black hair, warm brown skin, big brown eyes, glossy yellow raincoat and long red knitted scarf) and the stylized 3D animated feature film rendering: create a new scene. Lia kneels on a mossy path in an enchanted forest at golden hour, curiously reaching toward a tiny glowing blue mushroom, fireflies around her, god rays through tall trees. Medium-wide shot, Lia slightly right of centre. Same character proportions and outfit. 16:9.
-
Using Image 1 only for Lia's character design … and the … rendering
写明参考图的任务:她的角色设计和影片的画风。屋顶和夜空则不要。
-
the girl with the voluminous curly black hair, warm brown skin …
名字旁边配上看得见的描述,模型才会把正确的特征绑定到 Lia 身上。
-
create a new scene
新的地点、新的时刻,所以不要求沿用旧构图。
3. 从你认可的那张图开始编辑
修改时,附上已认可的图片,同时说明要改什么、什么必须保留。改姿势可以保留机位和场景;改机位就意味着换一个视角,这时再要求完全保留旧构图就和修改本身矛盾了。把“保留”和“修改”分开写,并检查两者是否真的都做到了。
Edit Image 1 into an empty set plate. Keep the exact camera angle, the square wooden table, the steaming red-and-white split hotpot, all plates of beef, tofu, meatballs, greens, sauce bowls, chopsticks, beer bottles and glasses, the red paper lanterns, wooden lattice screens and warm tungsten light with red glow. Remove all people completely: the four friends at the table are gone, leaving four empty wooden stools around the table, and the restaurant background is empty of diners, just empty tables, lanterns and soft warm bokeh. Photorealistic, vertical. No people, no faces, no text.
除非是在与模型的同一段对话中继续,每次请求都是从头开始:要重新提交图片和需要的参考图。如果某次修改破坏了脸、产品或场景,就回到上一张好图,修正原因,而不是把问题一路带下去。另外,“只改这一块”只是引导,而不是像素级蒙版:其余部分也要看一遍。
4. 只在文字能看清时才把它放进图里
生成之前就定好确切的文字、数字和语言。说明每段文字放在哪里、以什么形式呈现:标题、标签、小字。给重要的文字留出空间和对比度。对于以文字和版式为核心的图(信息图、海报、图解),Bald Rhino 使用 Nano Banana Pro;只需要一个标题或几个标签时,Nano Banana 2.1 就够了。
YouTube thumbnail, 16:9, punchy and appetising. Close-up photo-style shot of a hand holding a steaming skewer of takoyaki and a convenience-store onigiri in a glowing neon-lit Osaka alley at night, red paper lanterns and Japanese signage blurred behind. A stack of shiny 100 yen coins sits in the lower right. On the left, huge bold yellow text with a thick black outline reads exactly: "EAT FOR $5". A small white label at the top reads exactly: "JAPAN ON A BUDGET". Warm saturated colours, shallow depth of field, the food is the clear focal point, text readable at small size.
-
reads exactly: "EAT FOR $5"
文案已定稿,并逐字放在引号里。
-
On the left … A small white label at the top
每段文字都有位置和大小:一个标题,一个标签。
-
text readable at small size
封面图会以很小的尺寸被看到:检查标准直接写进了请求。
即便如此,也要按实际尺寸检查结果。模型仍可能拼错字,或让小字和细节变形(Google 的 Nano Banana Pro 提示词技巧,英文)。如果文案必须准确或可编辑,就先生成画面,再单独排版文字。
5. 在图片成为下一步输入之前先检查
以合适的尺寸把结果和参考图对比:脸、手、标记、产品细节、人物数量、阴影以及与地面的接触。分清有意的表情或光线变化和设计被改掉之间的区别。修复之后,再把整个构图检查一遍。要作为视频场景起点的图片最值得严格把关:它里面的任何错误都会进入每一帧。