Seedream 5.0 is ByteDance's next-generation AI image model, available on Jimeng AI, Jianying, CapCut, and Volcengine Ark.
Seedream 5.0 is a new-generation AI image generation model launched by ByteDance, and is now available on Jimeng AI, Jianying, CapCut, and Volcano Ark.
Core Capabilities | Core Capabilities
| Capability | Description |
|---|---|
| Real-time Web Search | Auto-fetches trending info when prompt contains timely keywords |
| Multi-step Reasoning | Interprets abstract concepts (e.g. "serene tech feel" → desaturated + clean lines + cold light) |
| Multi-round Editing | Iterative refinement: local edits, style transfer, element add/remove, text rendering |
| High Resolution | Native 2K, AI-enhanced 4K, 2-5 second generation |
| Character Consistency | Maintains face, clothing, pose across multiple images (storyboard-ready) |
| Text Rendering | 99%+ accuracy for Chinese/English text, use quotes for best results |
Prompt Structure
Basic Structure (Text-to-Image)
[主体描述] + [行为/动作] + [环境/背景] + [材质/质感] + [光影效果] + [构图要求] + [风格关键词]
- Subject + action + environment usenatural languagedescription
- Style/color/lighting/composition useshort phrasesas accents
- Text content should be marked with quotation marks, such as:
"Hello World"
Four-Part Structure (Advanced)
主体 → 环境 → 材质/质感 → 光影
Editing Prompt Formula
变化动作 + 变化对象 + 变化特征
示例："将骑士的头盔变为金色"
Style Vocabulary Library
Realistic Photography
写实电影剧照商业摄影纪实摄影超写实RAW 原片质感- Lens:
85mm定焦35mm广角长焦压缩感鱼眼镜头 - Lighting:
伦勃朗光环形光分割光黄金时刻暖光蓝调时刻冷光霓虹光
Anime/Illustration
- Japanese Anime:
吉卜力动画风格新海诚风格日系少女漫画赛璐璐质感 - Western:
美漫风格DC漫画风格欧美写实人物Pop Art波普艺术 - Chinese:
国潮插画水墨画风格中式工笔画赛博国风 - Other:
像素风格低多边形扁平插画厚涂油画水彩手绘
Design/Commercial
极简主义包豪斯风格磨砂玻璃质感高质感金属赛博朋克电影海报级别品牌VI视觉信息图Infographic知识卡片
Lighting and Shading Modifiers
戏剧性侧光柔和漫射光高对比度低饱和度莫兰迪色调赛博霓虹暖橙调冷蓝调胶片颗粒感
Common Prompt Templates
Realistic Portraits
[性别年龄外貌]，[服装描述]，[表情神态]，[环境背景]，85mm定焦，自然光，写实电影剧照风格，超高清，细节丰富
Landscapes/Scenes
[场景描述]，[时间/天气]，[光线描述]，[构图]，[风格词]，电影感构图，8K超清
Knowledge Cards (Complete Template)
生成一张[格式/载体]风格的图像，向[目标受众]解释/展示"[核心概念]"。
图像需具备[风格特征A]、[风格特征B]和[排版要求C]，整体感觉类似于[熟悉参照物]。
Brand/Poster (Blank Space Template)
[视觉主体描述]，[材质描述]，[光影效果]，
所有视觉主体集中在画面[左/右]侧，为[右/左]侧留出大面积干净的背景区域，方便后期排版添加文字。
背景：[背景描述]
Continuous Storyboard (Character Consistency)
参考[图1]的面部和发型，将其更改为[场景风格]装束，
生成N张连续的[场景描述]分镜图，[风格]，需要在一个场景中，连续动作。
E-commerce Products
为这件[产品]创建[平台]风格的展示图，风格类似于[品牌参照]，
背景简洁，突出产品质感，专业商业摄影
Scene Quick Reference
| Scene | Prompt Keywords | Notes |
|---|---|---|
| Avatar | 头像图标 正方形构图 纯色背景 |
Specifying a style reference image produces better results |
| Knowledge Cards | 信息图 知识图谱 排版清晰 |
Explain the target audience and core concept |
| PPT Background | 留白构图 偏向[左/右]侧 哑光背景 |
Emphasize blank space on one side for layout |
| Character Cosplay | 保持人脸不变 写实质感服饰 相同姿势 |
Upload the original image + target character image |
| Journal Diary | 手写字体 纸张纹理 拼贴风格 米黄底色 |
Specify the date and weather to add atmosphere |
| Glass Icons | 磨砂玻璃质感 渐变色 C4D OC渲染 |
Pure white background + simple composition |
| Poster Design | 电影海报级别 戏剧光 大面积留白 |
Clarify the text content and position |
| Talisman/China-Chic | 山海经 国潮票据 水墨 篆刻印章 |
Can add "wish" text to increase emotion |
Advanced Techniques
1. Web Search Trigger
Automatically searches the web when the prompt contains time-sensitive words:2026年流行色 最新款XX 今年XX趋势 米兰冬奥会
2. Image Editing
- Specify Region: "Replace [region] in the image with..."
- Style Transfer: "Keep the content unchanged, change it to [style]"
- Element Control: "Add/remove [element] to/from the image"
- Lighting Adjustment: "Change the image lighting to [lighting name]"
- Filter Addition: "Add a [filter name] filter to the image"
- Makeup Modification: "Add [makeup description] to the character"
3. Text Rendering
Put the text to be generated in quotation marks:图片中央写着"创意无界"
4. Composition Control
- Golden Ratio:
三分法构图黄金螺旋 - Viewpoint:
俯视鸟瞰仰视正面平视45度斜角 - Blank Space:
大量留白简洁背景主体偏[方向]
5. Multi-Image Fusion
Supports up to 14 reference images; when fusing, specify which element from which reference image to use:
参考图1的风格，图2的色调，图3的人物姿势
6. Image Set Generation
Trigger Words:一系列 组图 生成N张连续的 分镜图
Negative Prompt Writing
Clearly state the unwanted elements, placed at the end of the prompt:
背景简洁，不要杂乱元素保持人脸，不要改变面部特征不要文字水印不要过度曝光
平台入口 | Platforms
| Platform | URL | Description |
|---|---|---|
| 即梦AI Jimeng AI | https://jimeng.jianying.com/ | Main site, about 20 free 2K generations per day |
| 火山方舟 Volcengine Ark | https://console.volcengine.com/ark | Enterprise API, supports 4K |
| 剪映 Jianying | App Store | AI 绘画 → Seedream 5.0 |
| CapCut (Overseas) | App Store | AI Image |
API 生图脚本 | Image Generation Script
generate.py Calls the 即梦 4.0 API, images are automatically downloaded to --output-dir(default output/）。
Environment Preparation
In generate.py create in the same directory as .env write in VOLC_ACCESSKEY、VOLC_SECRETKEY, or export in the terminal. The script automatically reads the same directory .env。pip install -r requirements.txt。
Usage
# 文生图
python generate.py --prompt "一只猫在花园里玩耍，水彩风格"
# 图像编辑（输入参考图）
python generate.py --prompt "将背景换成海滩" --image-urls "https://example.com/photo.jpg"
# 指定分辨率 + 强制单图
python generate.py --prompt "电商主图，产品特写" --width 2560 --height 1440 --force-single
# 组图生成
python generate.py --prompt "生成4张分别关于春夏秋冬的盲盒组图"
Use in the Skill workflow
- Generate a prompt according to this Skill's rules, and have the user confirm it.
- Soft prompt before initiating: default is 1 image; if multiple images (image set) are needed, add
--no-force-singleor keep words such as "image set" or "a series". - Execute
python generate.py --prompt "<confirmed_prompt>"(for image sets, add--no-force-single）。 - After the script polling completes, the images are in
output/, display the paths and URLs.
Parameter Description
| Parameter | Description |
|---|---|
--prompt |
Required, prompt |
--image-urls |
Input reference image URLs (up to 10) |
--width / --height |
Specify output width and height (must be passed together); if not passed, intelligent adaptation is used |
--size |
Output area (pixels), default 2K (2048×2048) |
--scale |
Text influence degree 0~1 (default 0.5); the larger, the stronger the text |
--force-single |
Output only 1 image (default） |
--no-force-single |
Allow multiple images (image set); the model decides the number based on the prompt |
--watermark |
Add AI watermark |
--output-dir |
Directory to save generated images (default output/); both URLs and base64 will be written here |
References | 参考文件
- Detailed examples & use cases → examples.md
- Official docs, API params, size chart, full style dictionary → reference.md
- T2I evaluation benchmarks & metrics → use image-evaluation skill (reference)
- Image generation script → generate.py
- Dependencies → requirements.txt