40 条 MOSS-Image 提示词:海报、信息图、标签与社交媒体图
按用途分组、可直接复制的提示词,能把精确文字放到图上,外加让 MOSS-Image 正确渲染文字的提示词写法。
MOSS-Image 就是为带文字的图片而生的。在作者的长文本基准 LLTB 上,它得分 96.3,而 Qwen Image 2512 为 64.9。下面这 40 条提示词正是围绕这一强项设计的。每条都是一个代码块,要渲染的文字放在引号里。刚接触这个模型?先看 ComfyUI 指南或模型概览。
MOSS-Image 接受哪些提示词格式?
MOSS-Image 接受三种格式,而且可以混用:纯自然语言、booru 风格的标签序列,以及 JSON 版式对象。USERGUIDE 推荐英文,同时支持中文和其他语言,并为三种格式统一给出了一条负面提示词。
自然语言。 描述场景,把文字放在引号里,说明它的位置。本文大多数提示词用的是这种格式。
Booru 标签。 逗号分隔的标签,顺序固定:质量 → 评分 → 元信息 → 年份 → 角色 → 系列 → 画师 → 通用。
JSON 版式。 一个 summary 加一组 items,每个 item 带一个边界框。在论文的人类偏好测试中,JSON 版式提示词同时胜过了简短和详细的自然语言提示词。结构说明见版式 JSON 页;下面第 8 组有可直接运行的示例。
四条规则适用于每一条提示词:
- 把精确文字放在引号里。模型渲染的就是引号内的内容。
- 在需要换行的地方用
\n。不写的话,由模型自行决定。 - 除非文字本身是中文,否则用英文写提示词。中文文字渲染效果不错,但训练描述以英文为主。
- 使用推荐的负面提示词:
nsfw, worst quality, bad anatomy, watermark, jpeg artifact。
所有提示词都按推荐设置运行:50 步、CFG 3、shift 5、Euler 搭配 sgm_uniform。如果短提示词丢字,把 CFG 往 6 调。你也可以在首页的 MossImage 生成器里试任何一条提示词。
活动海报
活动海报需要一个大标题、一行日期和一个场地。给每项单独一行,并用文字描述层级(大、中、小)。文字背后的场景保持简洁。
Concert poster, dark navy background with a grainy spotlight. Large white serif title "MIDNIGHT ORCHESTRA" at the top. Below it in smaller text: "Live at Harbor Hall\nSaturday 14 November, 8 PM\nTickets at the door". Silhouette of a cellist at the bottom.
Minimalist tech conference poster, cream background, one orange circle in the centre. Bold black sans-serif title "BUILD 2026" above the circle. Three lines in small grey text at the bottom: "Developer Conference\n3-4 December\nLisbon Congress Centre".
Vintage farmers market poster in the style of a 1950s letterpress print, kraft paper texture. Red block letters "SUNDAY MARKET" across the top. Illustrated basket of vegetables in the middle. Footer text in green: "Every Sunday, 8 AM to 1 PM\nRiverside Square".
Film festival poster, black and white photograph of an empty cinema seat row. Thin white uppercase title "NORTHERN LIGHTS FILM FESTIVAL" running vertically down the left edge. Bottom right: "Oslo\n20-27 February\nnorthernlightsfest.org".
Charity run poster, bright yellow background, flat illustration of three runners. Big black title "RUN FOR RIVERS 10K". Under it: "Sunday 5 April\nStart 7:30 AM, City Park Gate\nRegister: runforrivers.org". Small sponsor line at the very bottom: "Supported by Clearwater Trust".
可以换日期、场地和那一个背景物件;标题控制在四个词以内。
信息图
信息图是最考验换行控制的地方。给条目编号,每项配一个简短标签,并说明排布方式(列、行、网格),这样模型才能把文字块放得整齐一致。
Clean infographic on a white background titled "HOW COFFEE IS MADE" in bold dark-brown text. Five numbered steps in a vertical column, each with a small flat icon on the left and text on the right: "1. Harvest\n2. Dry\n3. Roast\n4. Grind\n5. Brew".
Infographic poster with a light-blue background and a large water-drop illustration. Title "WATER FACTS" at the top. Three stat boxes in a row, each with a big number and a caption: "71%\nof Earth's surface", "3%\nis fresh water", "1%\nis easy to reach".
Comparison infographic, two columns separated by a thin vertical line. Left column heading "ELECTRIC" in green, right column heading "PETROL" in grey. Under each heading, three short lines: left "Zero tailpipe emissions\nLower running cost\nQuiet", right "Longer range\nFaster refuel\nMore models".
Timeline infographic on a dark background, horizontal line with four glowing dots. Above each dot a year in bold, below each a label: "1990\nFirst website", "2004\nSocial networks", "2007\nSmartphones", "2023\nGenerative AI". Title at the top: "THE WEB IN FOUR STEPS".
Flat-style recipe infographic for pancakes, pastel yellow background. Title "PERFECT PANCAKES" in rounded letters. Ingredients list on the left: "2 cups flour\n2 eggs\n1.5 cups milk\n1 tbsp sugar\npinch of salt". Steps on the right: "Mix\nRest 10 min\nCook 2 min per side".
可以调整条目数量(三到五个最稳妥),每个标签控制在四个词以内。
产品标签与包装
标签由品牌名、产品名、一句卖点和重量组成。先描述容器,再描述标签,最后从上到下描述文字。指定材质和表面处理,让风格有依据。
Front label of a matte black coffee bag, studio product shot. Centered gold text: brand "NORTH & CO" in small letters, product "SINGLE ORIGIN ETHIOPIA" in large letters, then "Medium roast\nNotes of blueberry and cocoa\n250 g" in thin white text at the bottom.
Amber glass bottle of hot sauce with a cream paper label, illustrated red chili in the centre. Label text: "FIRE ROAD" in distressed red letters at the top, "Habanero Hot Sauce" in black script below, "Heat level: 8/10\n150 ml" at the bottom.
Kraft paper soap box, flat lay on a linen cloth. Printed text in dark green: "WILD LAVENDER" large, then "Handmade cold-process soap\nLavender, shea butter, oat\n100 g" in small type. Sprig of lavender drawn in one line beside the text.
White aluminium cosmetic tube, clinical minimalist style, soft shadow. Text in black Helvetica: "DAILY SPF 50" as the main line, "Mineral sunscreen\nFragrance free\n50 ml" below it, and a tiny line at the base: "Made in Portugal".
Craft beer can, teal background with a geometric white wave pattern. Big condensed uppercase "SALT LINE" on the can, under it "Coastal Pale Ale" in script, then "5.2% ABV\n440 ml" in small white caps near the bottom.
可以换容器和材质;品牌名控制在两个词以内,并且始终保留重量或容量那一行。
社交媒体语录卡
语录卡就是简单背景上的一块文字。指定对齐方式、字体气质(衬线、手写、无衬线)和署名行。用 \n 决定在哪里换行。
Square quote card, off-white background, centered black serif text: "The best time to plant a tree\nwas twenty years ago.\nThe second best time is now." Small grey attribution underneath: "Proverb".
Instagram quote post, deep green background with a thin gold border. Centered hand-lettered white text: "Slow is smooth,\nsmooth is fast." No other elements.
Motivational card, photo of a mountain ridge at dawn, dark gradient at the bottom. White bold text at the bottom left: "Start before you're ready." Below it in small text: "@trailnotes".
Minimal typographic card, pale pink background, large black grotesque text filling the frame: "DONE\nIS BETTER\nTHAN\nPERFECT". Tiny text in the bottom right corner: "mossimage.pro".
Quote card with a torn-paper collage style, mustard yellow and navy. Text in white on a navy strip: "Read what you love\nuntil you love to read." Attribution in black on the yellow area: "Naval Ravikant".
可以换背景和换行位置;文字本身控制在二十个词以内。
菜单与招牌
菜单有很多带价格的短行,考验的是对齐。说明价格格式和条目数量。招牌则需要一行大字和一个清晰的载体表面。
Café chalkboard menu, black board with white chalk lettering, wooden frame. Heading "TODAY'S COFFEE". Items with prices right-aligned: "Espresso ...... 3.00\nFlat White ...... 4.50\nCold Brew ...... 4.00\nMatcha Latte ...... 5.00".
Printed restaurant menu card on textured ivory paper, thin gold frame. Heading "DINNER" in small caps. Three dishes in serif text, each with a one-line description: "Roast Chicken\nlemon, thyme, root vegetables\n\nSea Bass\nfennel, white wine butter\n\nMushroom Risotto\nparmesan, truffle oil".
Wooden hanging shop sign on a brick wall, carved and painted letters: "OPEN\n9 AM - 6 PM\nMon - Sat". Warm afternoon light.
Neon sign on a dark tiled wall, pink glowing cursive text "good vibes only", with a small blue neon arrow underneath. Slight glow reflection on the tiles.
Airport-style directional sign, dark grey panel with white Helvetica text and arrows: "Gates A1-A12 →\nBaggage Claim ↓\nExit ←". Clean fluorescent lighting.
可以调整条目数量和价格格式;每道菜或每行招牌文字控制在五个词以内。
图书、播客与 YouTube 封面
封面需要标题、副标题和作者或频道名,各用不同字号。在提示词里和画布尺寸里(16 的倍数)都说明宽高比。
Book cover, 2:3 portrait, dark teal background with a single white line drawing of a lighthouse. Title "THE QUIET SHORE" in large white serif, subtitle "A Novel" in small italic, author "Elena Marsh" at the bottom.
Non-fiction book cover, bold yellow background, large black sans-serif title "DEEP FOCUS" filling the top half, subtitle "How to work without distraction" in the middle, author "Dr. Sam Okafor" at the bottom in small caps.
Square podcast cover art, warm gradient from orange to magenta, flat illustration of a microphone. Title "LATE SHIFT" in chunky white letters, subtitle "Stories from the night" in thin text underneath.
YouTube thumbnail, 16:9, close-up of a mechanical keyboard on a desk, blurred background. Huge yellow text with a black outline on the left: "I BUILT IT\nIN 7 DAYS". Small red badge in the corner reading "NEW".
Children's picture book cover, soft watercolor forest with a small fox. Playful rounded title "FOX FINDS A FRIEND" across the top in orange, author "by Mia Chen" in brown at the bottom.
可以换宽高比和插画;一张封面上的文字字号不要超过三种。
中文海报
MOSS-Image 可以渲染中文,不过它的中文长文本召回率(92.4)略低于英文(97.1)。保持每行简短,把文字放在中文引号里,并且整条提示词都用中文写,让风格描述和文字内容保持一致。
春节海报,红色背景,金色烫金纹样,顶部大字标题「新春快乐」,中间一盏灯笼插画,底部两行小字:「2027 丁未年」\n「阖家团圆 万事如意」。
茶叶包装海报,米白色宣纸质感,左侧一枝墨色竹叶,右侧竖排书法文字「武夷岩茶」,下方小字「大红袍 · 特级」\n「净含量 125 克」。
书店活动海报,浅灰背景,极简排版。标题「城市夜读会」黑色粗体居中,副标题「每周五晚八点」,底部一行「地点:三联书店二楼」。
咖啡店菜单板,深绿色背景,白色手写粉笔字。标题「今日推荐」,下面四行带价格:「美式 18」\n「拿铁 24」\n「燕麦拿铁 28」\n「手冲 32」。
励志语录卡片,正方形,深蓝色背景,居中白色宋体文字:“路虽远,行则将至。”\n下方小字:“——《荀子》”。
可以换字体描述词(宋体、黑体、书法),每行控制在八个字以内。
版式 JSON 提示词
JSON 提示词让你能以像素级精度控制每段文字或每个物体的位置。结构有四个键:summary、items、bbox_coord_space 和 bbox_format。框使用 yxyx 顺序([y_min, x_min, y_max, x_max]),单位是以画布中心为原点的绝对像素,所以 1024 x 1024 的画布在两个轴上都是 -512 到 +512。
{
"summary": {
"meta": ["poster"],
"canvas": "single",
"caption": "A concert poster on a dark navy background with a glowing spotlight, the title MIDNIGHT ORCHESTRA at the top, a cellist silhouette in the middle and the date at the bottom."
},
"items": [
{"label": "text", "caption": "Large white serif uppercase title, centered.", "text": "MIDNIGHT ORCHESTRA", "bbox": [-448, -416, -320, 416]},
{"label": "object", "caption": "Silhouette of a seated cellist lit from above.", "bbox": [-256, -256, 256, 256]},
{"label": "text", "caption": "Small white sans-serif date line, centered.", "text": "Saturday 14 November, 8 PM, Harbor Hall", "bbox": [320, -416, 384, 416]}
],
"bbox_coord_space": "abs_centralized",
"bbox_format": "yxyx"
}
{
"summary": {
"meta": ["label", "packaging"],
"canvas": "single",
"caption": "A front label for a matte black coffee bag with gold text, a coffee bean illustration and the weight at the bottom."
},
"items": [
{"label": "text", "caption": "Small gold brand name in letterspaced caps.", "text": "NORTH & CO", "bbox": [-416, -256, -368, 256]},
{"label": "text", "caption": "Large gold serif product name, two lines.", "text": "SINGLE ORIGIN\nETHIOPIA", "bbox": [-336, -352, -176, 352]},
{"label": "object", "caption": "A single gold line-drawn coffee bean.", "bbox": [-128, -128, 128, 128]},
{"label": "text", "caption": "Thin white tasting note line.", "text": "Medium roast, blueberry and cocoa", "bbox": [192, -352, 240, 352]},
{"label": "text", "caption": "Small white weight line.", "text": "250 g", "bbox": [320, -96, 368, 96]}
],
"bbox_coord_space": "abs_centralized",
"bbox_format": "yxyx"
}
{
"summary": {
"meta": ["infographic"],
"canvas": "single",
"caption": "A clean white infographic titled WATER FACTS with three stacked stat rows, each a big blue number and a short caption."
},
"items": [
{"label": "text", "caption": "Bold dark-blue title, centered at the top.", "text": "WATER FACTS", "bbox": [-448, -448, -368, 448]},
{"label": "text", "caption": "Big blue number with a grey caption to its right.", "text": "71% of Earth's surface is water", "bbox": [-256, -448, -128, 448]},
{"label": "text", "caption": "Big blue number with a grey caption to its right.", "text": "3% of that water is fresh", "bbox": [-64, -448, 64, 448]},
{"label": "text", "caption": "Big blue number with a grey caption to its right.", "text": "1% is easy to reach", "bbox": [128, -448, 256, 448]},
{"label": "text", "caption": "Tiny grey source line at the bottom.", "text": "Source: USGS", "bbox": [400, -448, 448, 448]}
],
"bbox_coord_space": "abs_centralized",
"bbox_format": "yxyx"
}
{
"summary": {
"meta": ["social", "quote"],
"canvas": "single",
"caption": "A square quote card on an off-white background with a centered serif quote and a small attribution."
},
"items": [
{"label": "text", "caption": "Centered black serif quote, three lines.", "text": "The best time to plant a tree\nwas twenty years ago.\nThe second best time is now.", "bbox": [-160, -384, 128, 384]},
{"label": "text", "caption": "Small grey attribution, centered.", "text": "Proverb", "bbox": [192, -384, 240, 384]}
],
"bbox_coord_space": "abs_centralized",
"bbox_format": "yxyx"
}
{
"summary": {
"meta": ["cover", "illustration"],
"canvas": "single",
"caption": "A children's book cover: a watercolor fox on the left half, the title and author on the right half over a soft forest background."
},
"items": [
{"label": "character", "caption": "A small orange watercolor fox sitting and looking right.", "bbox": [-320, -448, 384, -32]},
{"label": "text", "caption": "Playful rounded orange title, three lines, right-aligned.", "text": "FOX\nFINDS\nA FRIEND", "bbox": [-192, 64, 128, 448]},
{"label": "text", "caption": "Small brown author line.", "text": "by Mia Chen", "bbox": [256, 64, 304, 448]}
],
"bbox_coord_space": "abs_centralized",
"bbox_format": "yxyx"
}
只改框的坐标和 text 的值;bbox_coord_space 和 bbox_format 保持不变,并确保每个框都在 -512 到 +512 之内。
FAQ
为什么 MOSS-Image 会丢掉或拼错我提示词里的词?
通常是因为文字没放在引号里、提示词太短,或者步数太少。把精确文字放进引号,短提示词把 CFG 往 6 调,并且至少用 50 步。
我可以把 JSON 版式和自然语言风格描述混用吗?
可以。USERGUIDE 说三种格式可以混用。summary.caption 字段是放风格描述的自然位置,每个 item 的 caption 则描述该元素。
MOSS-Image 处理中文和英文一样好吗?
接近。论文报告中文长文本 92.4,英文长文本 97.1,原因是训练描述以英文为主。像上面海报里那样的短中文行,渲染是可靠的。
做 16:9 的 YouTube 缩略图应该用多大的画布?
任何宽高都是 16 的倍数、总像素在 1 到 2 百万左右的尺寸都可以,比如 1536 x 864。模型可以从 0.25M 外推到 4M 像素,但推荐范围是 1M 到 2M。