🎨 图像创作快速上手
目标只有一个：让你十分钟内从「不知道写什么」变成「知道每一句提示词在干什么」。下面是梦想家创作台的真实操作顺序。
1. 60 秒生成第一张图
- 打开创作台，切到「图片」。
- 选模型（新手直接用默认的旗舰 Seedream 5.0 Pro Spicy），再选画质（1K 先练手、2K 出成品）和画面比例（竖图 9:16、横图 16:9、方图 1:1，还有 3:4 / 4:3 / 2:3 / 3:2）。
- 在提示词框里写清你想要的画面（下一节有公式）。
- 按钮上方会先把价格亮出来——多少点、什么画质，看明白了再点生成，不玩盲盒扣费。
- 成品出现在我的作品，可以下载，也可以点「加入资产」存进我的资产，换设备登录还在。
第一次别贪多：默认模型 + 1K + 一句写清楚的提示词，先把流程跑通，再逐轮加细节。
2. 提示词公式：一句话的七个零件
主体 + 动作 / 状态 + 场景 + 构图 + 光线 + 风格 + 关键细节
例：一位留着黑色长卷发的年轻女子站在海边悬崖上，红裙被海风吹起，回头望向镜头，全身远景，人物位于画面右侧，落日金光从侧面打亮发梢，电影胶片质感，裙摆与发丝细节清晰。
- 主体先行：先说「谁 / 什么」，再说环境和风格。只写「好看、高清」这类空词，等于把方向盘交给运气。
- 构图要明说：特写、半身、全身、俯拍、仰拍、三分法、人物在左还是在右——你不说，模型就随便摆。
- 光线是气氛开关：柔和窗光、霓虹夜光、黄金时刻、高反差侧光，换个光线就是换个片子。
- 画面里要出现文字（海报、招牌、标题）时，把文字内容原样写出来，并说明位置和颜色；文案越短越稳。文字活儿可以交给 GPT Image 2，它是站内写字最准的。
- 一次只改一两处：每轮只动一两个变量，才知道是哪句话起的作用。
3. 文生图 vs 参考图编辑
文生图：从零造画面
适合海报、人物、场景、概念图这类从无到有的创作，把公式七个零件写全就行。
参考图编辑：保住主体，只改局部
上传一张参考图，再在提示词里交代三件事——保留什么、改变什么、目标风格与构图。这是锁住同一个人物的正道：
例：保留参考图中人物的面部特征与发型，把服装换成黑色晚礼服，背景换成暖色调的剧院大厅，全身构图，写实电影质感。
- 锁脸：先用一张满意的定妆图当参考，后续所有图都从它出发改，脸就不会每张都换一个人。
- 换装工坊：只想换衣服时不用自己写长提示词——进「换装工坊」，选晚礼服 / 泳装 / 制服 / 睡衣 / 运动装等档位，或者从站内衣服库里直接挑一件，系统自动按旗舰模型出图。
- 姿势库：动作想不出来时，去创作台的姿势库（288 组、每组 5 个角度）挑一个现成姿势当参考，人物立刻摆得有模有样。
4. 六个图像模型怎么选
| 模型 | 拿手活 | 什么时候用 |
|---|
| Seedream 5.0 Pro Spicy（默认旗舰） | 高细节人物、海报、复杂描述，参考图编辑稳 | 人物定妆、重要成品、图生图改造——先想它 |
| Seedream 5.0 | 标准档全能选手，出图快、性价比高 | 日常出图、批量试方向 |
| GPT Image 2 | 画面内文字最准，海报排版强 | 带标题的海报、招牌、封面字 |
| Nano Banana 2 | 编辑与融合自然，改图手感好 | 局部改造、多元素合成 |
| Qwen-Image 2.1 · 千问 | 中文语义理解细，东方题材对味 | 中文提示词、国风与亚洲面孔题材 |
| FLUX 3 Klein | 快、便宜，1K 只要 5 点 | 草稿、构图试验、批量抽卡 |
模型阵容和价格以创作台当前显示为准——站里只上各家最新最强的，列表会持续更新。
5. 常见问题
每次生成的人物都长得不一样？
固定用同一张清晰的定妆参考图，并在提示词里把年龄、发型、脸部特征写成固定句式；不要在同一轮里同时改发型、服装和画风。
画面元素乱、人物站位不对？
减少主体数量，并把位置关系写死，例如「人物 A 在左侧、人物 B 在右侧，两人面对面」。咱们的画布与导演台还会自动给人物加数量锚点，防双胞胎、防多出一个人。
结果总差一口气？
先缩短提示词，只留关键要求把构图定住，再一轮轮补材质、光线、氛围。失败不扣点，放心多试。
🎨 Image Creation — Quick Start
One goal: in ten minutes you go from “no idea what to type” to knowing exactly what every line of your prompt does. This is the real workflow inside the Dreamer Studio.
1. Your first image in 60 seconds
- Open the Studio and switch to Image.
- Pick a model (start with the default flagship, Seedream 5.0 Pro Spicy), then a quality tier (1K to practise, 2K for finals) and an aspect ratio (9:16 vertical, 16:9 wide, 1:1 square, plus 3:4 / 4:3 / 2:3 / 3:2).
- Describe the picture in the prompt box (formula in the next section).
- The price is shown before you generate — points and tier up front, no blind-box charges.
- Results land in My Works. Download them, or tap “Add to Assets” to keep them in My Assets, synced to any device you sign in on.
Keep the first run simple: default model, 1K, one clear sentence. Get the loop working, then add detail round by round.
2. The prompt formula: seven parts of one sentence
Subject + Action / State + Setting + Composition + Lighting + Style + Key details
Example: A young woman with long black wavy hair stands on a seaside cliff, her red dress lifted by the wind, glancing back at the camera, full-body wide shot, subject on the right of the frame, golden sunset light rimming her hair, cinematic film look, crisp detail in the fabric and strands of hair.
- Subject first. Say who or what it is before the environment and style. “Beautiful, HD” alone hands the steering wheel to luck.
- Name the composition. Close-up, half-body, full-body, high angle, low angle, rule of thirds, subject left or right — if you don't say it, the model guesses.
- Lighting is the mood switch. Soft window light, neon night glow, golden hour, hard side light — change the light and you change the film.
- Text inside the image (posters, signs, titles): write the exact words, their position and colour, and keep them short. For lettering, GPT Image 2 is the most accurate writer on the site.
- Change one or two things per round. That's how you learn which line did the work.
3. Text-to-image vs reference editing
Text-to-image: creating from zero
For posters, characters, scenes and concept art from nothing — just fill in all seven parts of the formula.
Reference editing: keep the subject, change the rest
Upload a reference image and state three things in the prompt — what to keep, what to change, and the target style and composition. This is the reliable way to keep the same person:
Example: Keep the face and hairstyle of the person in the reference image. Change the outfit to a black evening gown and the background to a warm theatre hall, full-body composition, realistic cinematic look.
- Face lock: settle on one good “official” portrait, then branch every later image from it — the face stops changing between pictures.
- Outfit Studio: for outfit swaps you don't need long prompts — pick a tier (evening gown / swimwear / uniform / sleepwear / sportswear) or choose a piece from the built-in clothing library, and the flagship model does the rest.
- Pose library: stuck on a pose? Pick one from the built-in library (288 sets, 5 angles each) in the Studio and use it as a reference.
4. Picking between the six image models
| Model | Best at | Reach for it when |
|---|
| Seedream 5.0 Pro Spicy (default flagship) | High-detail people, posters, complex descriptions; steady reference editing | Character portraits, hero images, image-to-image work — try it first |
| Seedream 5.0 | Fast, well-rounded standard tier | Everyday images, testing directions in batches |
| GPT Image 2 | Most accurate in-image text and poster layout | Posters with headlines, signs, cover lettering |
| Nano Banana 2 | Natural edits and blending | Local changes, multi-element composites |
| Qwen-Image 2.1 | Fine-grained semantics, a natural feel for East-Asian subjects | Chinese prompts, East-Asian faces and themes |
| FLUX 3 Klein | Fast and cheap — 1K from 5 points | Drafts, composition tests, bulk rolls |
The live lineup and prices are always the ones shown in the Studio — we only carry each maker's newest and strongest, and the list keeps growing.
5. FAQ
The person looks different in every image?
Always branch from the same clear reference portrait, keep a fixed description of age, hairstyle and facial features, and never change hairstyle, outfit and art style in the same round.
Elements are messy, people stand in the wrong places?
Use fewer subjects and pin positions down in words — “person A on the left, person B on the right, facing each other”. Our Canvas and Director Studio also add count anchors automatically, so you don't get twins or an extra person.
Close, but not quite right?
Shorten the prompt to the essentials and lock the composition first; then rebuild material, light and mood round by round. Failed runs don't charge points — experiment freely.