English
This page is available in English
The whole page — instructions, buttons and results.
Switch to English按万相官方指南写文生图提示词:公式与用语取自指南,反向提示词单独给出,按模型检查字数并生成 API 请求体。
中文文生图提示词生成器
按万相官方文生图指南的公式写提示词,反向提示词单独给出;用 API 的话,再按所选模型检查字数,给出请求体。
模型与 API 请求
画面比例(size 参数)
这几个模型官方没有给比例表,尺寸是本工具按「每边 512 到 1440」的规则推算的。
n 写成 1:接口默认生成 4 张,按张计费。
怎样写一条文生图提示词
写主体和场景
主体是画面里的主要对象,写清它的特征和动作,比如「一个可爱的 10 岁中国小女孩,穿着红色衣服」;场景写它所处的环境。
点选风格和镜头语言
风格、景别、视角和镜头类型都可以点选,这些写法都出自万相的文生图提示词指南。也可以在旁边的框里写指南没有列出的风格。
加上氛围和细节
氛围词描述画面的感觉,比如梦幻、孤独、宏伟;细节修饰写光源的位置、道具、环境细节或「高分辨率」。
分开复制两个框
提示词放进工具的提示词栏,反向提示词只放进反向(负面)提示词栏。
用 API 时,选模型和比例
选好模型,计数器会按它的字数上限检查;超出的部分阿里云会直接截掉,不报错。再点一个比例,下方就是可以直接发送的请求体。
为什么按指南的公式写
这一页照着阿里云百炼的「文生图Prompt指南」(万相)来写提示词。指南给了两条公式:基础公式是「主体 + 场景 + 风格」;进阶公式是「主体(主体描述)+ 场景(场景描述)+ 风格(定义风格)+ 镜头语言 + 氛围词 + 细节修饰」。这里按进阶公式的顺序把你写的和点选的内容接成一段描述。
这一页以前把各栏用「## 标题」拼起来,标题文字会跟着粘进模型;还把反向提示词接在同一段文字后面,而指南说 prompt 和 negative_prompt 是两个参数。现在提示词是一段连贯的描述,反向提示词单独放在一个框里。
用 API 出图时还有三件事指南没讲,写在阿里云的文生图 API 参考里:每个模型的字数上限不同,超出部分直接截断;画面尺寸由 size 参数设定,不设就是默认的 1280*1280;prompt_extend 默认会改写你的提示词。模型、比例和改写开关就是为这三件事放的。
其他图像模型的官方指南没有在本页引用,所以这里不声称同一套写法对即梦、可灵或混元同样有效。
常见问题
出自阿里云百炼的「文生图Prompt指南」(万相)。进阶公式是「主体(主体描述)+ 场景(场景描述)+ 风格(定义风格)+ 镜头语言 + 氛围词 + 细节修饰」,按钮上的景别、视角、镜头类型、风格和光线用语都取自指南的提示词词典。
指南说明文生图模型有两个参数:prompt 是正向提示词,描述想要的画面;negative_prompt 是反向提示词,描述不希望出现的内容。两者是分开的参数,所以这里也分开给你。
每个模型有字数上限:wan2.6-t2i 为 2100,wan2.5-t2i-preview 为 2000,wanx2.0-t2i-turbo 为 800,wan2.2 和 wan2.1 只有 500;反向提示词都是 500。超出的部分会被自动截断,不会报错,所以要先选对模型,留意计数器。
默认会。prompt_extend 默认开启,会先用大模型优化正向提示词;想按原文出图,就取消勾选。阿里云还提醒,改写后的提示词可能触发版权内容审核,遇到这类报错时可以关闭改写再试。
能复制过去用,但这里的公式和用语是万相官方写的,别家的指南没有在本页引用。用在别的工具上,请以那个工具自己的说明为准。
不会。提示词完全在你的浏览器里拼好,不调用任何模型,也不发送到任何地方。
方法与来源
计算方式
In the browser, by a tested kernel: subject, scene, style, camera language, mood and detail joined into one description in the order of the advanced formula in Alibaba Cloud's 万相 (Wan) text-to-image prompt guide, with the negative prompt separate. Each is counted against the chosen model's limit, and the request body is shaped as that model's API reference shapes it (input.messages for wan2.6, input.prompt below it). No model is called.
本工具依据的标准
- Corrected 2026-10-21: the page sent readers to ERNIE ViLG (文心一格), but yige.baidu.com now redirects to Baidu's 文心助手, whose own description lists image creation (图片创作). Every mention, in all three languages and the tool config, now names Wenxin (文心). The other four products were checked on their own sites: Jimeng (ByteDance), Wan (万相, Alibaba Tongyi; the page's older name 通义万相 is kept as a recognisable alias), Hunyuan (Tencent) and Kolors (可图, Kuaishou).
- Rebuilt 2026-11-03 in all three languages. The page joined its fields under Markdown "##" headings, so each heading was pasted into the model with the picture, and it appended the negative terms to the same text, while the 万相 guide gives prompt and negative_prompt as separate parameters. The prompt is now one description in the guide's advanced-formula order, and the negative prompt has its own box and copy button.
- Every shot-size, angle, lens and lighting term is written as the guide's prompt column writes it (特写镜头, 俯视视角), each style as its example prompt uses it (国风水墨风格, 浅水彩), and 梦幻, 孤独 and 宏伟 are its own examples of mood words. 陶瓷 is left out because its example names a subject (瓷器小狗), not a style phrase.
- Extended 2026-11-04, when /tongyi-wanxiang-prompt-builder/ was merged into this page: its model picker and length counter, size values, prompt_extend switch and API request body now sit under the prompt, written by the same tested kernel. Every figure was re-read in Alibaba's API reference on 2026-11-04 and matched what that page carried.
- The sizes for wan2.2 and below are COMPUTED: Alibaba gives those models a rule (each side 512 to 1440 pixels, default 1024*1024) and no ratio table, and the page says so beside them. The body asks for n = 1 because the API default is 4 images, each billed.
- Merged 2026-11-04: /tongyi-wanxiang-prompt-builder/ now redirects here. Its model picker and length limits, size values, prompt_extend switch and API request body were moved onto this page first, each re-read in Alibaba's API reference the same day.
资料来源
- Jimeng (即梦AI), ByteDance: official site. https://jimeng.jianying.com/
- Wan (万相), Alibaba Tongyi: 「万相是阿里巴巴通义旗下的AI创意创作平台…提供文生图、图生图…」. https://tongyi.aliyun.com/wan/
- Tencent Hunyuan (腾讯混元): official site. https://hunyuan.tencent.com/
- Kolors (可图大模型), Kuaishou: official site. https://kolors.kuaishou.com/
- Baidu 文心助手: description lists 图片创作; yige.baidu.com (文心一格) redirects here. https://wenxin.baidu.com/
- Alibaba Cloud Model Studio. 文生图Prompt指南 (万相): the basic and advanced formulas, the prompt and negative_prompt parameters, and the term tables every button is taken from. https://help.aliyun.com/zh/model-studio/text-to-image-prompt — read 2026-11-03.
- Alibaba Cloud Model Studio. 文生图 V2 API reference (万相): prompt limits (wan2.6-t2i 2100, wan2.5-t2i-preview 2000, wanx2.0-t2i-turbo 800, wan2.2 and wan2.1 500), negative_prompt 500, 「超过部分会自动截断」; size rules and 「常见比例推荐的分辨率」; the wan2.6 input.messages shape; n 「默认为 4」 and billed per image; prompt_extend 「true:默认值」. https://help.aliyun.com/zh/model-studio/text-to-image-v2-api-refe… — read 2026-11-04.
Pick up where you left off
Stored only in this browser — never sent to our servers.