Seedream 5 Pro Prompt Guide: Text and Reference Images

Build a Seedream 5 Pro prompt in five parts: subject, layout, exact text, what changes, and what stays. For a readable poster, put every string in its own zone; for an edit, keep the change list separate from the keep list. Seedream 5 Pro here means Seedream 5.0 Pro.

The six examples below ran through OmniaKey on November 12, 2026 (local), all at 1440x1920 with watermark false. All six kept the requested Chinese strings. Poster B missed its sans-serif request, and winter edit A added props; both are visible in the pairs. Each prompt ran once, so these are single samples.

The five parts of a Seedream 5 Pro prompt

Volcengine's reference-edit guide separates what to change from where. In a text prompt, name the region, then name the change. The five parts below follow that order.

Adaptable template

This is a fill-in template, not a tested prompt. Replace each bracket with your own content. The template is in English so you can reuse the structure with your own Chinese or English strings; the six prompts below stay exactly as tested.

[shot type] of [subject] in [setting], [lighting or style]. Place [main element] in [zone] and [secondary element] in [zone]. Show only these exact strings: "[text 1]" at [zone] in [size], "[text 2]" at [zone] in [size], and "[text 3]" at [zone] in [size]. Keep [details that must stay]. Change only [edit 1], [edit 2], and [edit 3]. Do not add [unwanted elements].

The short prompts (A) name the job and the strings. The zoned prompts (B) add positions, type instructions, exclusions and keep lists. Both styles produced readable Chinese in these six outputs. The differences landed in placement and in what else entered the frame. To send one of these prompts through the API, the Seedream 5.0 Pro API tutorial has the create, poll and download code.

Poster text: four exact Chinese lines

Poster A sets the event, palette and four strings in one sentence. Poster B assigns each string a zone and asks for a sans-serif Chinese face.

Prompt A

设计一张竖版咖啡店周末活动海报，暖米色背景，深绿色文字，橙色点缀，一杯拿铁和一枝桂花。海报上只出现这些中文文字：“桂花咖啡日”“周六 10:00–18:00”“手作拿铁 · 秋日限定”“松间咖啡”。排版清楚，所有文字易读。

Prompt B

设计一张竖版咖啡店周末活动海报。暖米色纯色背景，深绿色无衬线中文字体，少量橙色点缀。画面分三个区：顶部约三分之一留给标题，中间是一杯拿铁和一枝桂花的静物，底部留给活动信息。只排以下四行文字，逐字保留，不加引号、不加英文、不新增标语：
顶部最大字号：桂花咖啡日
标题下方中等字号：手作拿铁 · 秋日限定
底部第一行：周六 10:00–18:00
底部第二行：松间咖啡
文字不压在杯子或花枝上，四边留出宽裕空白，所有文字在画面内完整可见。

Settings: generate, November 12, 2026 (local), 1440x1920, watermark false, one call each.

What we saw: poster A shows all four strings, spelled as asked, with the time near the top, the subtitle below it and the store name at the bottom. A large photoreal cup, leaves and small graphic flourishes sit against a cream and orange field. Poster B shows the same four strings, with the subtitle after the title, the time and store name at the bottom and the cup central. The zone instructions were followed. The requested sans-serif face was not: the headings still show serif, brush-like stroke variation. B is more minimal, with less orange accent than A, and its cup angle and latte art differ. These are two independent generations, so the cup and latte art are not a layout-only comparison.

What to change: if a specific typeface matters, set the type in a layout tool after generation; this output did not honor the sans-serif instruction. If placement matters, copy B's zone structure. Keep the four strings quoted and unchanged.

Product label: three lines that stay readable

Product A describes the bottle and three label lines. Product B adds the camera angle, the full bottle height, the label shape and a rule against extra text or icons.

Prompt A

拍摄一只虚构品牌的琥珀色玻璃冷萃咖啡瓶，黑色旋盖，正面贴奶油色长方形纸标签，标签上只有“松间”“桂花冷萃”“250 mL”三行文字。瓶子居中直立，放在浅色木桌上，背景为柔和米色，旁边放一小枝桂花。自然柔光，清晰的电商产品照片，标签文字易读。

Prompt B

拍摄一只虚构品牌的琥珀色玻璃冷萃咖啡瓶，黑色旋盖，正面贴奶油色长方形纸标签。镜头与标签正面平行，瓶子居中直立，完整展示瓶盖到瓶底，瓶身约占画面高度三分之二。标签由三行深绿色印刷文字组成，逐字使用，禁止添加其他文字、图标或徽章：
第一行大字：松间
第二行中等字：桂花冷萃
第三行小字：250 mL
标签边缘平直，文字水平，保持字间距。瓶子放在浅色木桌上，背景为柔和米色，旁边放一小枝桂花，花枝不挡标签。自然柔光，清晰的电商产品照片。

Settings: generate, November 12, 2026 (local), 1440x1920, watermark false, one call each.

What we saw: product A renders 松间, 桂花冷萃 and 250 mL correctly. The text is brown, the bottle amber, the cap black, the label cream and the flower sits on the right; the whole bottle and label are visible. Product B renders the same three lines correctly, in green type with a larger brand name and a larger cream label, with the flower on the left. The whole bottle is visible at about two thirds of the image height. Its cap shape and reflections differ from A. These are two separate bottles, not a before-and-after pair.

What to change: for a brand name that must read at thumbnail size, B's larger brand and label are the direct difference. If the cap or reflections must match, supply a reference and expect a redraw; these two were generated separately. We used B as the common source for both edits below.

Reference edits: change the label, keep the bottle

Both edits start from product B. Prompt A asks for a winter product photo in one sentence. Prompt B names three changes and lists what must stay.

Prompt A

把参考图改成冬季产品照片：背景换成浅灰蓝色，去掉桂花，把标签上的“桂花冷萃”改为“冬日冷萃”。保持同一只瓶子和其他标签文字。

Prompt B

以参考图为唯一产品依据，只做三项修改：一、把米色背景改成浅灰蓝色；二、移除瓶子旁的桂花花枝；三、仅把标签第二行“桂花冷萃”替换为“冬日冷萃”。保留原瓶子的轮廓、琥珀色玻璃、黑色旋盖、奶油色标签形状与位置、拍摄角度、瓶子大小和浅色木桌。标签第一行“松间”和第三行“250 mL”逐字保留。新增文字沿用原标签的深绿色印刷字体和字号。不要增加雪花、围巾、装饰品或其他文字。

Settings: edit, product B source, November 12, 2026 (local), 1440x1920, watermark false, one call each. Method: both edits used the same product-B image, uploaded as a 1024 px long-edge JPEG; no post-generation repair or text retouching.

What we saw: edit A replaces 桂花冷萃 with 冬日冷萃 correctly and keeps 松间 and 250 mL. The background turns pale blue-grey and the original flowers are gone. It also adds pine branches, pinecones, berries, snow and drifting white flecks beyond the three requested edits. The bottle silhouette and label placement remain recognizable; the bottle looks darker. Edit B makes the same text replacement and keeps the brand and volume. It keeps a recognizable bottle silhouette, cap and label placement, turns the background pale blue-grey, removes the flowers and adds no winter props or snow. Its amber tone is closer to the source. In this output B stayed closer to only the three changes.

What to change: write the three edits as numbered items, then add a short exclusion list for snow, props and new text. Repeat the keep list every round. For label type that must stay pixel-identical, set it again in a layout tool; neither edit preserved the label at pixel level. For another look at how reference details survive an edit, see our Seedream 5.0 Pro vs GPT Image 2 test.

What missed, and what to change next

• The sans-serif request in poster B did not hold. The headings still show serif and brush-like variation. Brand-critical type needs a layout pass after generation.
• Edit A added pine branches, pinecones, berries, snow and white flecks beyond the three edits. Name the three edits and exclude the rest, as B does.
• All six outputs kept the requested Chinese strings, including the four poster lines and the label replacement. Each prompt ran once, so check the text again at the final display size.
• Both edits used the same product-B source, and both redrew the bottle, cap and reflections. Treat reference edits as a redraw, then rebuild any area that must stay exact.
• Volcengine's image generation API reference suggests keeping Chinese prompts under about 300 characters and English under about 600 words, and says overlong prompts can drop elements. That is guidance, not a hard cap. All six prompts here are Chinese and fit inside it.

Send a prompt through the API

To send a prompt, create a task with POST https://api.omniakey.com/v1/media/tasks, then poll GET /v1/media/tasks/{id} and download the asset URL. Use model Dola-Seedream-5.0-pro, type image, and operation generate; put the prompt in input.prompt. OmniaKey's image API docs list the input and size rules: parameters.size accepts 1K, 1.5K, 2K, or WxH, and parameters.n must be 1. The request below is an example with the poster A prompt.

{
  "model": "Dola-Seedream-5.0-pro",
  "type": "image",
  "operation": "generate",
  "input": {
    "prompt": "设计一张竖版咖啡店周末活动海报，暖米色背景，深绿色文字，橙色点缀，一杯拿铁和一枝桂花。海报上只出现这些中文文字：“桂花咖啡日”“周六 10:00–18:00”“手作拿铁 · 秋日限定”“松间咖啡”。排版清楚，所有文字易读。"
  },
  "parameters": {
    "size": "1440x1920",
    "watermark": false,
    "n": 1
  }
}

The example uses the same 1440x1920 size and watermark false as the six images. Set watermark to false to remove the default “AI 生成” label in the bottom-right; JPEG is the default output. For an edit, switch operation to edit and put the source image in input.images as an array of URLs or data URLs. The request body is capped at 1 MiB, so a URL is easier than inline Base64. The full create, poll and download code is in the Seedream 5.0 Pro API tutorial. For the current rate, see the Seedream 5.0 Pro model page. To try a prompt without code, use the image generator.

Frequently asked questions

Is Seedream 5 Pro the same model as Seedream 5.0 Pro?

Yes. Here, Seedream 5 Pro is the shorter name for Seedream 5.0 Pro. The OmniaKey model ID is Dola-Seedream-5.0-pro, with generate and edit operations.

Why did the sans-serif instruction come back with serif strokes?

In poster B, the requested sans-serif Chinese face did not hold: the headings still show serif and brush-like variation. Set brand-critical type in a layout tool, or test the exact face before relying on it.

Can an edit keep a label pixel-identical?

Treat it as a redraw, not a pixel copy. Both winter edits used the same product-B source and kept the label text and placement recognizable, but the bottle, cap and reflections were redrawn. Rebuild that layer after generation when exact type or packaging matters.

How long should a Seedream 5 Pro prompt be?

Volcengine's image generation API reference suggests Chinese prompts under about 300 characters and English under about 600 words, and says overlong prompts can drop elements. That is guidance, not a hard cap. All six prompts here fit inside it.

Sources checked

• Volcengine Ark: Seedream 5.0 Pro reference-edit guide
• Volcengine Ark: image generation API
• OmniaKey: image API docs
• OmniaKey: Seedream 5.0 Pro model page

Checked November 12, 2026 (local).