全部内容基于 BFL 官方仓库原文:
black-forest-labs/flux-3-image-klein-pup→system_prompts.jsonblack-forest-labs/flux-3-image-klein(GitHub)→ README 与 examples 本文的 system prompt 与官方逐字一致,未做改写。
一、为什么短提示词不生效
BFL 官方 README 原话:
Do not use either model without the prompt upsampler. The model is trained on the captions it writes; a prompt sent as typed gives noticeably weaker results. It is on by default in the CLI, the demo and the Python API.
关键在于 "trained on the captions it writes" —— 这个模型不是被 SD 那种关键词喂大的, 它是被自己生成的长篇场景描述训练出来的。所以它的"母语"是 200 词左右的完整画面描述, 你塞给它 "a girl on the beach" 这种词串,等于用外语跟它说话。
官方实测对比(同一份模型、同一提示词意图):
| 长度 | 内容 | |
|---|---|---|
| 用户输入 | 178 字符 | double exposure portrait: a young East Asian man facing the camera, a larger translucent profile of him facing left layered over him, muted blues, film grain, silver hoop earring |
| 展开后(模型实际吃的) | 1325 字符 | (见下方"展开样例") |
差了 7.4 倍。这就是"简单提示词不生效"的真正原因。
二、三层方案,按省力程度排序
方案 1:直接用工作流里的展开器(推荐,零成本)
你下载的 flux3_klein_int8_t2i.json 工作流里已经内置了 Prompt Upsampler 子图,
里面嵌的就是官方那份 3404 字符的 system prompt(我已逐字比对,与 BFL 原文一致)。
用法: 在 Prompt 节点里写你的短句(中文也行),点运行 →
Upsampled prompt 预览节点会显示展开后的完整描述 → 再喂给扩散模型。
也就是说:你不用学怎么写长提示词,写短句就行,展开器负责翻译。
- 短句上限:不要超过 2-3 句话,太长的输入反而会让展开器"保留原文措辞"而不是重新规划
- 展开器输出的永远是英文(官方规定 All prose is English),中文短句 → 英文完整描述
- 想微调:直接改
Upsampled prompt预览出来的文本,再跑一次
方案 2:在 ComfyUI 外自己展开(想反复打磨时用)
把下面这段原文,作为系统提示词粘到任何大模型(DeepSeek / 豆包 / ChatGPT / 本地 Qwen), 然后把你的想法当作用户消息发过去。DeepSeek 免费而且中文理解好,性价比最高。
You are the prompt upsampler of a text-to-image model. Expand the user's short request into ONE detailed, self-contained caption: the exact text the image model receives. Output only the caption, with no greetings, explanations, refusals, questions, Markdown, headings, JSON, coordinates, ids or angle-bracket tokens, and no quotes around it. If the input is not clearly an image request, caption a visualization of its concept. The request may begin with a "TARGET CANVAS: W:H" line: compose for that canvas (wide spreads content sideways, tall stacks it) and never mention the ratio or the canvas.
Fidelity and planning: keep every explicit subject, attribute, count, spatial constraint, style cue and piece of wording; never contradict, substitute, drop or narrow the request, and keep an abstract concept abstract. Decide everything left open as ONE concrete, consistent plan; never offer alternatives or hedge. Never re-express a stated constraint through another mechanism; add camera optics (depth of field, bokeh, blur) only when asked for or implied. Without a stated style, plan an actual photograph and open with the medium convention ("Full-color outdoor photograph, ...", "Full-color indoor photograph, ...", "Studio photograph, ...", "Close-up photograph, ..."); keep "Photorealistic" for renders, and turn "photorealistic" or "realistic" requests for photos into the photograph convention. Give a photograph a few ordinary, non-ideal specifics drawn from the scene's own objects instead of the stock-photo ideal. Do not invent names, brands, logos or celebrities; keep user-named ones. A long request keeps its content but not its wording: drop appraisal and sales vocabulary, turn moods into observable physical facts, and add only new concrete content with positions. Time of day is lighting, not rendered text.
Caption format: one continuous prose block in the generative voice, never "this image shows", "we can see" or "is visible". Lead with medium and style, then scene, lighting and camera, then the subjects with concrete colors, materials and state. Every person in focus gets an observable profile: ethnicity, gender, age range, skin tone ("fair skin", "brown skin"), hair and clothing. Make the layout reconstructable from words: frame-relative positions (frame-left, frame-right, upper, lower, foreground, midground, background), sentence-initial anchors ("On the left, ...", "In the background, ..."), relations (holds, faces, behind) and a consistent order; describe up to about ten instances one by one, beyond that give the exact count and arrangement. Out-of-focus or distant planes are shapes, colors and silhouettes, never identities or readable text, and the background takes about a sixth of the caption. Be definite: no hedging adverbs, no quality or emotion judgments, no closing sentence about mood; state negations as what is present. Rendered text: transcribe every wording verbatim inside double quotes in its original language and script, keep the line breaks of multi-line text, and give its typography, color and placement; author concrete, domain-correct wording when the request needs text it does not supply, and name canonical structures (calendar, keyboard, periodic table) exactly with their full layout. All prose is English; only quoted text keeps its language. Typical length is 100 to 350 words, more for complex scenes.
Output only the caption.
省 token 的用法:这个 system prompt 一次贴进去,之后在同一会话里连续发短句即可,不用重复贴。
方案 3:直接找我
把想法告我,我按上面这份官方规则展开,你复制回去用。适合你懒得切窗口的时候。 我同时会告诉你「哪几个词是模型最可能偏离的」,方便你二次修正。
三、手工写的时候,硬性规则清单
如果你要自己写,下面这些是官方 prompt 里明文规定的,违反任意一条画质都会掉:
结构(按顺序)
- 先写媒介和风格 —— 照片用
Full-color outdoor photograph,/Studio photograph,/Close-up photograph,;注意官方明确要求:照片类不要写 "photorealistic",那个词是给渲染图用的 - 然后场景 → 光照 → 相机
- 然后主体,带具体颜色、材质、状态
人物(每个主体都必须有)
- 每个人物给一份可观察的档案:族裔、性别、年龄段、肤色(写
fair skin/brown skin这种)、发型、服装
布局(必须能用文字重建画面)
- 用相对画面的位置词:
frame-left、frame-right、upper、lower、foreground、midground、background - 句首锚点:
On the left, .../In the background, ... - 关系词:
holds、faces、behind,保持顺序一致 - 十个以内逐个描述;超过十个,改成"精确数量 + 排列方式"
- 背景约占全文 1/6,别写成主体
- 虚化/远处的平面只写形状、颜色、剪影 —— 不要写身份,也不要写可读的文字
语气
- 必须明确:不用 "maybe"、"somewhat" 这类虚词;不写画质评价、不写情绪总结; 不要结尾补一句"画面氛围很温馨"。否定要写成"画面里有什么",不是"没什么"
- 不要出现
this image shows/we can see/is visible—— 直接用生成式语态
画面里的文字
- 要出现的文字,用双引号原样括起来,保留原语言和原有换行
- 同时给出:字体风格、颜色、位置
- 该有文字但你没提供的,模型会自己编 —— 所以你最好自己写死
时长
- 典型 100–350 词,复杂场景更多。太短(<80 词)就会退化成"关键词模式"
四、可套用的模板
[媒介风格] photograph of [场景],
[光照与相机] lit by [光源], shot on [镜头/视角],
[主体A] [位置] [族裔/性别/年龄段/肤色/发型/服装] [动作/姿态] [手上拿着什么],
[主体B] [位置 + 与主体的关系] ...,
[环境细节] [背景占 1/6],
[画面中的文字] "原文" in [字体] [颜色] placed [位置]
中文想法 → 按这五段填 → 自然就落在 150–250 词区间。
五、进阶:用边界框精确控制布局
FLUX 3 Klein 支持在提示词里用边界框摆放元素(这是它区别于 FLUX.1 的核心能力)。
- 坐标系:0–1000 的网格,拉伸到画布上
- 工作流里对应节点:
CreateBoundingBoxes - system prompt 会自动切换成
t2i_bbox变体(3302 字符,也在工作流里) - 支持的语法:在描述里引用元素 id,然后把该元素"换颜色、移动、替换、删除"
物理宽高比为 a 的元素,其 x 跨度 ≈ y跨度 × a × H/W(官方原文给出的换算)。
适合场景:海报排版、产品摆位、"这个放左上角,那个放右下角"这类位置敏感的需求。
六、常见错误对照表
| 你可能会写 | 问题 | 改成 |
|---|---|---|
a beautiful girl, masterpiece, 8k | 画质词对 FLUX 3 无效,是 SD 时代的习惯 | 删掉,换成具体描述 |
photorealistic photo of ... | 官方要求照片用 Full-color outdoor photograph, | 改开头媒介词 |
no people in the background | 否定表述效果差 | 写成 The background is empty, showing only ... |
a warm, cozy mood | 情绪词会被丢弃 | 换成可观察事实:warm amber lamplight on the wooden table |
a girl | 缺人物档案,模型随机发挥 | a young woman in her mid-twenties with fair skin and shoulder-length black hair |
at night 当成文字渲染 | 官方:时间是光照,不是文字 | lit by a single streetlamp at night |
| 逐步堆关键词到 50 个 | 超 10 个同类元素会乱 | 写 exactly twelve identical chairs arranged in two rows of six |
| 中文长描述直接喂扩散模型 | 展开器输出的规范是英文 | 中文喂展开器,让它出英文 |
七、展开样例(官方 examples 原文)
输入(178 字符):
double exposure portrait: a young East Asian man facing the camera, a larger translucent profile of him facing left layered over him, muted blues, film grain, silver hoop earring
模型实际吃到的(1325 字符):
Digital double-exposure portrait in a muted blue monochrome palette, framed as a square
head-and-shoulders composition with visible 35mm film grain. Centered in the foreground, a young
East Asian man in his twenties faces the camera directly, shown from the shoulders up with fair
skin, short black hair, dark eyes, and a neutral closed mouth. He wears a plain dark crew-neck
shirt that merges into the shadowed lower portion of the image, and a polished silver hoop earring
hangs from his left ear. His face remains sharply defined beneath a larger translucent profile of
the same man layered across it, facing left. The oversized profile rises above and extends beyond
the frontal head, its forehead, nose, lips, chin, and swept-back hair rendered as pale blue
silhouettes with feathered edges. The translucent profile overlaps the frontal face so that both
sets of eyes, cheekbones, and strands of hair intersect through the center. Behind them, the double
exposure dissolves into a hazy blue field containing indistinct cloud forms and fine tonal texture,
with no hard horizon. Cool diffused light catches the bridge of the frontal nose and the rim of the
silver hoop, while deep navy shadows gather around the outer edges. Fine scratches, dust specks,
uneven exposure, and dense analog grain run across the entire frame.
注意它做了什么:没有加任何新主体,但把"young East Asian man"补全成了完整人物档案、 把"layered over him"展开了两段空间关系、把"muted blues"落实成具体的光影描写。 这就是你要的效果。
八、配套参数(distilled 版)
| 参数 | 值 |
|---|---|
| 步数 | 16 |
| Guidance | 无(distilled 版不用 CFG) |
| 采样器 | euler / simple |
| 分辨率 | 原生 1024×1024,最高 4K |
| 提示词长度 | 100–350 词 |