Seedream 5.0 Pro

Introduction

In-image text is one of the weakest surfaces of most image models. Latin scripts drift on tricky accents, tracking wobbles, and anything beyond a short brand name tends to garble. Non-Latin scripts break outright: hangul syllables lose their composition, Arabic letter joins collapse, Japanese kanji strokes get invented, Thai tone marks land in the wrong place. For any deliverable where the copy has to be correct, that's a launch-blocker.

Seedream 5.0 Pro is trained natively on 15 languages: Arabic, English, Russian, Indonesian, Spanish, German, Turkish, Portuguese, Malay, Vietnamese, French, Japanese, Korean, Tagalog, and Thai. The model renders their scripts, diacritics, and layout conventions well enough for commercial poster, cover, and packaging work. Other languages still generate, but in-image text and local cultural rendering are weaker than for the 15 natively supported ones.

A French jazz festival poster in portrait 3:4 orientation, warm mid-century modern style. Stylised silhouette of a saxophonist in midnight-blue on cream. Title FESTIVAL DE JAZZ DE MONTPELLIER in tall condensed serif capitals with accented é clearly formed. Subtitle Trente-huitième édition in italic serif with both è accents accurate. Dense credits column reading dates and venues in fine capitals at the bottom.

The poster above is one prompt with four separate blocks of French copy: title, subtitle, mid-block credits column, and a bottom date-and-venue line. Every accented character (é, è, à) renders correctly, ligatures hold, and the type hierarchy reads as a designer laid it out. The rest of this guide covers how to write prompts that hit that quality across the 15 native languages.

Request shape

Multilingual text rendering doesn't need a special parameter. The text you want in the image goes into positivePrompt inline, quoted so the model treats it as literal copy.

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'bytedance:seedream@5.0-pro',
  positivePrompt: 'A French jazz festival poster in portrait 3:4 orientation. Title FESTIVAL DE JAZZ DE MONTPELLIER in tall condensed serif capitals. Subtitle Trente-huitième édition in italic serif. Warm mid-century poster aesthetic on textured cream paper.',
  width: 1328,
  height: 1776
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "bytedance:seedream@5.0-pro",
            "positivePrompt": "A French jazz festival poster in portrait 3:4 orientation. Title FESTIVAL DE JAZZ DE MONTPELLIER in tall condensed serif capitals. Subtitle Trente-huitième édition in italic serif. Warm mid-century poster aesthetic on textured cream paper.",
            "width": 1328,
            "height": 1776
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "imageInference",
      "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
      "model": "bytedance:seedream@5.0-pro",
      "positivePrompt": "A French jazz festival poster in portrait 3:4 orientation. Title FESTIVAL DE JAZZ DE MONTPELLIER in tall condensed serif capitals. Subtitle Trente-huitième édition in italic serif. Warm mid-century poster aesthetic on textured cream paper.",
      "width": 1328,
      "height": 1776
    }
  ]'

runware run bytedance:seedream@5.0-pro \
  positivePrompt="A French jazz festival poster in portrait 3:4 orientation. Title FESTIVAL DE JAZZ DE MONTPELLIER in tall condensed serif capitals. Subtitle Trente-huitième édition in italic serif. Warm mid-century poster aesthetic on textured cream paper." \
  width=1328 \
  height=1776

{
  "taskType": "imageInference",
  "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "model": "bytedance:seedream@5.0-pro",
  "positivePrompt": "A French jazz festival poster in portrait 3:4 orientation. Title FESTIVAL DE JAZZ DE MONTPELLIER in tall condensed serif capitals. Subtitle Trente-huitième édition in italic serif. Warm mid-century poster aesthetic on textured cream paper.",
  "width": 1328,
  "height": 1776
}

[
  {
    "taskType": "imageInference",
    "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
    "imageUUID": "9c1b2d3a-4e5f-6789-abcd-ef0123456789",
    "imageURL": "https://im.runware.ai/image/os/a14d18/ws/2/ii/9c1b2d3a-4e5f-6789-abcd-ef0123456789.jpg"
  }
]

The essentials are the same as any Seedream 5.0 Pro request:

• positivePrompt is required, up to 3000 characters. The provider recommends under 600 English words, though the character budget rises for scripts with different word conventions.
• width and height are required for text-to-image. Their product must fall between 1,048,576 and 4,194,304 pixels.
• Text renders wherever the prompt names it. There is no dedicated text parameter and no separate typography field.

Rendering across scripts

The four posters below are commercial use cases across four of the harder scripts to render: Arabic (right-to-left, complex letter joins), Thai (tone marks and vowel positions), Korean (syllable-block hangul), and Japanese (mixed kanji, hiragana, katakana with vertical layouts). Each is one prompt with the copy quoted inline.

An elegant Arabic-language luxury fragrance advertisement. A tall faceted amber glass perfume bottle with a polished gold cap on warm cream marble. The Arabic brand name قصر العنبر in elegant modern thuluth-style calligraphy in warm rose gold across the upper third, with correct letter joins and dot placement. Below in smaller naskh, the tagline عطر خالد يحمل روح الشرق in cream tone. Discreet English signature MAISON AMBER at the bottom.

A vibrant Thai street food promotional poster. A steaming bowl of pad kra pao with basil, a fried egg on top, and jasmine rice on a distressed wooden table. Golden hour side light. Title ผัดกะเพราต้นตำรับ in warm hand-painted brush script in deep terracotta on the wood, correct tone marks and vowel positions. Subtitle รสจัดจ้าน เผ็ดร้อน หอมกลิ่นใบกะเพรา in warm cream. Small price panel ราคาพิเศษ ฿89 at lower right.

A Korean streaming drama key art poster. Close cinematic three-quarter portrait of a young female Korean actor under warm neon-lit rain on a Seoul back-alley at night, wearing a black waterproof trench coat. Title 밤의 상속자 in bold modern condensed hangul in warm cream. Subtitle 복수는 밤에 시작된다 in elegant serif hangul. Streaming credit 넷플릭스 오리지널 · 2026년 8월 공개 in tracked cream capitals at the bottom.

A Japanese fashion magazine cover. Confident female Japanese model in her early thirties with a sleek dark bob against a soft muted sage-green studio backdrop, wearing a structured cream wool coat over a deep navy roll-neck. Masthead 静けさの季節 in tall vertical mincho serif down the left side in warm charcoal. Horizontal cover line 北の光と、冬の色。 across the top in fine sans-serif. Small tag ATRIUM JAPAN in the lower right.

Look closely at each: the Arabic letter joins hold across the calligraphic curves and the diacritic dots land on the correct sides. The Thai tone marks ( ่ ้ ) sit above the correct consonants and the sara-a vowels lean the right way. The Korean hangul syllable blocks compose cleanly with proportional spacing. The Japanese vertical mincho masthead reads down the left side with correct kanji stroke count and hiragana proportions, and the horizontal cover line sits at the top without wrestling with the vertical block.

None of these outputs use a special typography flag. The prompt names the script, the type style ("thuluth calligraphy", "hangul", "mincho serif"), and the exact copy in quotes, and the model treats that as the deliverable.

Prompting for in-image text

Three rules do most of the work for accurate text rendering.

Quote the exact copy inline in the prompt. Not "a Korean drama title", not "an Arabic fragrance name", but the literal string with the actual characters. Seedream reads the quoted string as the copy to render. Paraphrased text ("a fragrance name in Arabic") produces an invented string that reads plausibly to a Latin-script eye but is often nonsense in the target script.

Name the script and the type style. "In hangul" or "in mincho serif" or "in Arabic thuluth calligraphy" tells the model both the character system and the visual treatment. Without the type style the model defaults to a plain sans-serif of the script, which is fine for signage but wrong for a poster.

Name where the copy lives in the composition. "Across the upper third", "down the left side vertically", "in the lower right corner", "under the masthead". The model treats those clauses as layout instructions and reserves the region for the text, which stops the copy from getting hidden behind the subject or awkwardly cropped.

Multi-script layouts

Bilingual deliverables (a Russian novel with an English publisher line, a Japanese product with an English brand mark, a Korean streaming series with an English platform tag) don't need two calls. Seedream handles multiple scripts in one composition when the prompt names each script's copy and placement separately.

A minimalist Russian literary fiction book cover. A stylised watercolour of a lone silver birch tree in early morning mist against a soft grey-blue countryside. Russian title Тени берёз across the upper third in elegant modern Cyrillic serif in warm cream, with the ё correctly rendered with its two dots. Author name Александра Соколова below in smaller cream serif tracked capitals. English publisher line NORTHWICK PRESS · CONTEMPORARY LITERATURE at the bottom in small tracked capitals. Editorial matte paperback aesthetic.

The Russian title sits at the top in Cyrillic serif, with the diacritic-marked ё composing correctly among the other letterforms. The author name sits directly below in the same script at a smaller weight. The bottom publisher line switches to Latin script in a much smaller tracked treatment, visually anchoring the cover without competing with the title above. All three text blocks were named separately in the prompt with their own script, weight, and position, and the model composed them into a cover that reads as one design.

Match visual weight across scripts. When two scripts share a layout, matching their weight (title cream and bold, subtitle cream and light) makes the design read as intentional rather than as a translation retrofit. The prompt does the balancing work by describing each block's weight and colour in the same terms.

Tips

1. Quote the exact copy inline in the prompt. The literal string in the target script is what the model renders. Paraphrased or English-described text produces invented copy that reads as noise in the target script.

2. Name the script and the type style together. "Hangul", "mincho serif", "Arabic thuluth calligraphy", "Devanagari" combined with a treatment ("bold condensed", "elegant italic", "warm hand-painted brush") gets the character system and the visual finish right in the same clause.

3. Name where the copy lives in the composition. "Across the upper third", "down the left side vertically", "in the lower right corner" reserves layout space for the text and stops the model from hiding it behind the subject.

4. Stick to the 15 natively supported languages when text quality matters. Arabic, English, Russian, Indonesian, Spanish, German, Turkish, Portuguese, Malay, Vietnamese, French, Japanese, Korean, Tagalog, Thai. Other languages generate, but in-image text and local cultural rendering are weaker.

5. Keep in-image text under a paragraph. Title, subtitle, tag line, small credit block all render cleanly. A full paragraph of body copy is where any current model, including Seedream, starts producing typos. If the deliverable needs body text, generate the layout and composite the paragraph externally.

6. For multi-script layouts, describe each block separately. Its script, its style, its colour, its position. The model composes multiple scripts into one design cleanly when each is given its own directive rather than left to the model's inference.

7. Small text is a soft ceiling. Small copy (fine tracked capitals, footer text) renders significantly better than in previous generations, but occasional typos still appear. For legally-required small text (regulatory notices, ingredient lists), composite it externally rather than trusting the render.

8. Use the 2K tier for text-dense deliverables. Above 2.36 megapixels the pixel budget covers finer type more cleanly. Titles hold at 1.5K, but dense credit blocks and small subtitles read better at 2K.