scenario-multi-reference

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Scenario Multi-Reference Composition

Scenario 多参考图合成

Overview

概述

A reference image lends everything it shows unless the prompt says otherwise: hand a model a raccoon chef as a style reference and the raccoon walks into the scene. The fix is to give each picture exactly one job, then say, for every picture, what not to take from it. Four or five pictures, one idea each (a drawing, a photo, a mannequin, a film still, a color swatch), compose into one image that keeps each contribution and nothing else. Connection and the generation loop: see the
scenario
skill. Per-family reference caps and syntax:
scenario-image
lists the image model-family skills. One subject held across many images:
scenario-consistency
. Animating the result:
scenario-video
. Refusals:
scenario-moderation
. If a sibling skill named here is missing from your available skills, load it with
skill_get
(Scenario MCP) or ask the user to install it (https://github.com/scenario-labs/skills#install); unattended, proceed from tool schemas and flag the gap.
除非提示词另有说明,否则参考图会把它所呈现的一切都带入结果:如果把一张浣熊厨师图片作为风格参考交给模型,浣熊也会走进场景。解决方法是为每张图片指定恰好一个职责,然后针对每张图片说明不要从中采用什么内容。四张或五张图片,每张只提供一个想法(绘画、照片、人体模型、电影剧照、色卡),可以合成一张既保留各项贡献又不包含其他元素的图像。连接方式和生成循环:参见
scenario
skill。各模型系列的参考图数量上限与语法:
scenario-image
列出了图像模型系列 skill。需要在多张图像间保持同一主体:参见
scenario-consistency
。为结果添加动画:参见
scenario-video
。处理拒绝情况:参见
scenario-moderation
。如果此处提到的同级 skill 不在可用 skill 列表中,请使用
skill_get
(Scenario MCP)加载,或请用户安装(https://github.com/scenario-labs/skills#install);在无人值守模式下,则根据工具 schema 继续执行并标记这一缺失。

Quick reference

快速参考

JobThe picture lendsTypical "not" line
Main subject (WHO)the character, product"Use only the idea of a gecko. Do not copy the photo's colors."
Stage (WHERE)the place"redrawn as woodblock architecture" when the medium differs
Posethe body pose"Do not show the mannequin or any wood."
Anglethe camera"Do not include the lighthouse."
Stylemedium, render, texture"Do not include the raccoon, the kitchen, the apron or the pot."
Lightthe light only"Do not include the flowers or vase."
Palettethe only allowed colors"No green anywhere."
Materiala surface"Do not show any fish, fins or water."
Expressionthe mood of a face"No human skin, no nose, no ears, no hair."
Wearclothing or armor"The cat's own face stays visible; no human face mask."
Inspirationa design idea, not the thing"a gown designed from this jellyfish. Do not show a jellyfish."
Layoutgrid and composition"Do not copy the word MONDAY."
The prompt, one line per picture (the kept prompts wrote Main subject and Stage as WHO and WHERE):
text
<One sentence naming the medium.> Each reference picture has exactly one job:
Image N (<what it shows>) = <JOB> only. <What it lends, concretely.> Do not <what to leave out>.
...
<Framing: shot size, placement, mood, text rules.>
职责图片提供的内容典型的“不要”指令
主体(WHO)角色、产品“仅采用壁虎的概念。不要复制照片中的颜色。”
场景(WHERE)地点当媒介不同时:“重绘为木版画建筑”
姿势身体姿势“不要展示人体模型或任何木材。”
角度镜头视角“不要包含灯塔。”
风格媒介、渲染方式、纹理“不要包含浣熊、厨房、围裙或锅。”
光线仅光线“不要包含花朵或花瓶。”
配色唯一允许使用的颜色“任何地方都不要出现绿色。”
材质表面材质“不要显示任何鱼、鱼鳍或水。”
表情面孔的情绪“不要出现人类皮肤、鼻子、耳朵或头发。”
穿着服装或盔甲“猫自己的脸仍然可见;不要使用人脸面具。”
灵感设计理念,而非实体本身“设计一件以这只水母为灵感的礼服。不要出现水母。”
布局网格与构图“不要复制 MONDAY 这个词。”
提示词模板,每张图片一行(经过验证的提示词将 Main subject 和 Stage 写作 WHO 和 WHERE):
text
<One sentence naming the medium.> Each reference picture has exactly one job:
Image N (<what it shows>) = <JOB> only. <What it lends, concretely.> Do not <what to leave out>.
...
<Framing: shot size, placement, mood, text rules.>

Rules that held

始终成立的规则

  • One job per picture. Two jobs on one card (subject plus palette) blur both. A card kept in the call but unused gets "Image 4 = no job: ignore it completely."
  • One main subject. Two identities in one run split attention; compose them per
    scenario-consistency
    .
  • The job most likely to be ignored goes first, usually Style, with "This is the most important job". A clay scene came back as chrome CG until the Style line moved to the top.
  • Lend concretely. "Copy its 3D render look, soft materials, warm bounce light" beats "use its style". A palette job names each element's color ("the gecko is Prussian blue with vermilion spots"); "palette only" barely recolors.
  • Settle conflicts between jobs. Every picture carries more than its job, so a Style card brings its own colors and a subject photo its own fur. With a Palette card, the Style line adds "do not copy its colors", and the subject's colors are named from the swatch. When that recolors something the user cares about (their dog, their product), ask before running; unattended, follow the task instructions, else keep the subject's own colors and report the conflict.
  • Idea, not copy. When a subject photo must not survive literally: "Use only the idea of a gecko ... Do not copy the photo's pose, lighting, colors or realism."
  • Text is quoted exactly with "No extra words, no duplicate text"; copy that must be letter-perfect goes to
    scenario-text-overlay
    .
  • Clean cards. Plain backgrounds, no readable signs or logos. The user's own pictures are best; generate a missing card with a text-to-image run, one idea per card.
  • 每张图片只有一个职责。 同一张图片承担两个职责(主体加配色)会模糊两者。调用中保留但未使用的图片应写为“Image 4 = 无职责:完全忽略它。”
  • 只有一个主体。 一次生成中出现两个身份会分散注意力;请按照
    scenario-consistency
    合成它们。
  • 最容易被忽略的职责放在最前面,通常是风格,并加上“这是最重要的职责”。一个黏土场景曾生成成镀铬 CG,直到把风格行移到最前面才解决。
  • 具体说明可提供的内容。 “复制其 3D 渲染外观、柔和材质、温暖反弹光”优于“使用其风格”。配色职责要说出每个元素的颜色(“壁虎为普鲁士蓝色,带朱红色斑点”);仅写“只采用配色”几乎无法完成重新着色。
  • 解决职责之间的冲突。 每张图片都包含超出其职责的内容,因此风格图会带入自身颜色,主体照片会带入自身毛色。使用配色图时,风格行要加上“不要复制它的颜色”,并根据色卡指定主体颜色。如果这会改变用户在意的内容(他们的狗或产品),运行前应先询问;无人值守时遵循任务说明,否则保留主体自身颜色并报告冲突。
  • 采用理念,而非复制。 当主体照片不能以写实形态保留时:“仅采用壁虎的概念……不要复制照片的姿势、光线、颜色或写实风格。”
  • 文字要准确加引号,并写明“不添加额外文字,不重复文字”;必须逐字准确呈现的复制内容交给
    scenario-text-overlay
    。
  • 使用干净的参考卡。 采用纯色背景,不带可辨识的标牌或 Logo。优先使用用户自己的图片;缺少某张图片时,通过文生图运行生成,每张图只提供一个想法。

Worked example: a crayon drawing becomes a film still

实例:蜡笔画变成电影剧照

Cards: a child's crayon monster (Main subject), a backyard photo (Stage), a wooden mannequin leaping (Pose), a 3D animated film still of a raccoon chef (Style).
  1. upload_asset
    each card, then
    upload_asset_complete
    (see
    scenario
    ). Keep the asset ids in the order you will number them.
  2. recommend
    with
    capability="img2img"
    ,
    features=["referenceImages"]
    (keeps only members exposing that exact field; drop it to see members that name the slot differently), and the user's own words as
    prompt
    . Handle
    next_step
    as
    scenario
    directs.
  3. model_schema_get
    on the pick. Read: the reference field's cap (four or five cards fit the leading members at authoring time, up to ten on some),
    array: true
    , how the description says the prompt addresses references (by order as "Image 1", or by tags such as
    @Image1
    on some families), the sizing family, and the quality tier. When the schema offers a dedicated slot for a job (a style-reference field, a pose control), put that card there instead of in the reference array (
    scenario-consistency
    ), and read the field description for how the prompt refers to it, since prompt numbering follows the array.
  4. model_run
    with
    dry_run=true
    (top-level, beside
    model_id
    , never inside
    parameters
    ) and the schema's own names, for example
    parameters={"prompt": "<below>", "referenceImages": ["asset_monster", "asset_yard", "asset_mannequin", "asset_raccoon"], "quality": "high"}
    plus its sizing fields. Quote the user the total, any cards still to generate and the film included when one is wanted (a dry run rejects a payload missing its required image, so price against any uploaded picture, a small stand-in when no card exists yet), and run nothing paid until they agree (unattended, launch only inside a budget the task states; otherwise stop and report the quote). Then the same call with
    wait=false
    ,
    jobs_wait
    , and
    asset_display
    .
text
A still frame from a 3D animated family feature film. Each reference picture has exactly one job:
Image 1 (child's crayon drawing) = WHO. Turn this crayon monster into a soft, plush 3D animated character: round purple body, three uneven eyes, spiky green hair, tiny orange legs, big toothy smile. Keep its lopsided, hand-drawn charm.
Image 2 (backyard photo) = WHERE. Use this exact backyard: lawn, wooden fence, red hose, blue kiddie pool, swing set, lemon tree.
Image 3 (wooden mannequin) = POSE only. The monster leaps mid-air with both arms thrown up and one knee raised, exactly like the mannequin. Do not show the mannequin or any wood.
Image 4 (raccoon chef frame) = STYLE only. Copy its 3D render look, soft materials, warm bounce light and rich color. Do not include the raccoon, the kitchen, the apron or the pot.
Wide shot, the monster in the center above the lawn, afternoon sun, joyful.
  1. Check every picture for a leak (below). Fix by tightening that picture's line and running again, never by retouching.
  2. To animate it, hand the result to
    scenario-video
    as the first frame. The film prompt carries no jobs, only motion, camera, and sound: the frame already holds every job.
Seven more kept prompts across claymation, product ad, woodblock print, fashion editorial, pixel art, tilt-shift miniature, and risograph poster, each with its film prompt: references/prompt-patterns.md.
参考卡:儿童蜡笔怪兽(主体)、后院照片(场景)、跳跃的木制人体模型(姿势)、浣熊厨师的 3D 动画电影剧照(风格)。
  1. 对每张参考卡调用
    upload_asset
    ,然后调用
    upload_asset_complete
    (参见
    scenario
    )。保留 asset id,并确保其顺序与之后的编号顺序一致。
  2. 调用
    recommend
    ,设置
    capability="img2img"
    、
    features=["referenceImages"]
    (只保留公开该确切字段的成员;移除该字段可查看以其他名称称呼该槽位的成员),并将用户自己的原话作为
    prompt
    。按照
    scenario
    的指示处理
    next_step
    。
  3. 对选中的模型调用
    model_schema_get
    。读取:参考字段的上限(编写时主流成员可容纳四张或五张卡,部分成员可容纳至十张)、
    array: true
    、字段描述如何通过顺序(如“Image 1”)或标签(如某些模型系列中的
    @Image1
    )在提示词中引用参考图、尺寸系列以及质量档位。当 schema 为某项职责提供专用槽位(如风格参考字段、姿势控制)时,将对应卡片放入该槽位,而不是参考图数组(
    scenario-consistency
    );同时阅读字段描述,了解提示词如何引用它,因为提示词编号遵循数组顺序。
  4. 调用
    model_run
    ,设置
    dry_run=true
    (位于顶层,与
    model_id
    并列,绝不能放在
    parameters
    内),并使用 schema 自身的字段名,例如
    parameters={"prompt": "<below>", "referenceImages": ["asset_monster", "asset_yard", "asset_mannequin", "asset_raccoon"], "quality": "high"}
    ,再附上其尺寸字段。向用户报出总费用、仍需生成的卡片数量,以及需要时是否包含影片费用(dry run 会拒绝缺少必需图片的 payload,因此应使用任意已上传图片进行估价;若尚无卡片,则使用小型占位图),在用户同意前不要运行任何付费操作(无人值守时,仅在任务规定的预算内启动;否则停止并报告报价)。然后使用相同的调用,设置
    wait=false
    、
    jobs_wait
    和
    asset_display
    。
text
A still frame from a 3D animated family feature film. Each reference picture has exactly one job:
Image 1 (child's crayon drawing) = WHO. Turn this crayon monster into a soft, plush 3D animated character: round purple body, three uneven eyes, spiky green hair, tiny orange legs, big toothy smile. Keep its lopsided, hand-drawn charm.
Image 2 (backyard photo) = WHERE. Use this exact backyard: lawn, wooden fence, red hose, blue kiddie pool, swing set, lemon tree.
Image 3 (wooden mannequin) = POSE only. The monster leaps mid-air with both arms thrown up and one knee raised, exactly like the mannequin. Do not show the mannequin or any wood.
Image 4 (raccoon chef frame) = STYLE only. Copy its 3D render look, soft materials, warm bounce light and rich color. Do not include the raccoon, the kitchen, the apron or the pot.
Wide shot, the monster in the center above the lawn, afternoon sun, joyful.
  1. 检查每张图片是否发生泄漏(见下文)。通过收紧对应图片的指令并重新运行来修复,绝不要使用后期修图。
  2. 如需为结果添加动画,请将结果作为第一帧交给
    scenario-video
    。影片提示词不承载任何职责,只描述动作、镜头和声音;该帧已经承载了所有职责。
另外还有七组经过验证的提示词,涵盖黏土动画、产品广告、木版画、时尚大片、像素艺术、移轴微缩模型和 Risograph 海报,每组都附有相应的影片提示词: references/prompt-patterns.md。

Leak check before keeping

保留结果前的泄漏检查

For each picture, ask two questions: is its job visible, and is anything else from it visible?
  • Pose and Angle: no mannequin, no wood, no prop from the sketch.
  • Style: none of the style reference's own subject, set, or objects.
  • Expression: the mood moved, the face did not (no human nose or skin on a toaster).
  • Layout: the grid moved, the reference's own words did not.
  • Palette: no color outside the swatch.
Before rewriting a line for an ignored job,
asset_get
the output:
metadata.referenceImages
lists what the run consumed, and a reference sent as a bare string where the schema says
array: true
was dropped silently, so no wording brings its job back.
对于每张图片,提出两个问题:它的职责是否可见?其中是否还出现了不应可见的其他内容?
  • 姿势和角度:不能出现人体模型、木材或草图中的道具。
  • 风格:不能出现风格参考图自身的主体、场景或物体。
  • 表情:应迁移情绪,而不是迁移面孔(例如,烤面包机上不能出现人类鼻子或皮肤)。
  • 布局:应迁移网格,而不能迁移参考图中的原有文字。
  • 配色:不能出现色卡之外的颜色。
在改写指令以修复被忽略的职责之前,对输出调用
asset_get
:
metadata.referenceImages
会列出本次运行实际使用的参考图;如果 schema 指定
array: true
,却以裸字符串形式发送参考图,该参考图会被静默丢弃,因此仅改写提示词无法恢复其职责。

Common mistakes

常见错误

  • The main subject copied the photo. A green gecko photo came back as the same green gecko in a woodblock print. Put Style first, use "only the idea of", and make the palette strict.
  • A logo by accident. Red, white and blue swooshes on generated billboards read as a soda brand: the image-to-video model refused the frame, and the image editor then refused it as input. Say what billboards and signs show (waves, clouds, cranes).
  • Brands in a reference. A street photo with real shop signs was refused as a reference. Regenerate the card with "invented, blank signs, no logos".
  • A protected design named in the prompt. A famous chair named outright was refused. Describe an original design instead.
  • A garment blocked by the safety filter. A translucent gown was refused; "opaque, fully covered, high-necked" passed.
  • Drift in the film. A guitar changed color mid-clip. Ask for "almost static", list what must not change, and move only what should.
  • Hardcoding a model id or price. Discover with
    recommend
    , read caps off
    model_schema_get
    , price with
    dry_run=true
    .
  • 主体复制了照片。 一张绿色壁虎照片生成后,仍是木版画中的同一只绿色壁虎。将风格放在最前,使用“仅采用……的概念”,并严格限定配色。
  • 意外出现 Logo。 生成广告牌上的红、白、蓝色弧形标志看起来像汽水品牌,导致图生视频模型拒绝该帧,图像编辑器随后也拒绝将其作为输入。明确说明广告牌和标牌上应显示什么(波浪、云、鹤)。
  • 参考图中包含品牌。 带有真实商店标牌的街景照片被拒绝作为参考。重新生成参考卡,并写明“虚构的空白标牌,无 Logo”。
  • 在提示词中直接点名受保护的设计。 著名椅子被拒绝。改为描述原创设计。
  • 服装被安全过滤器拦截。 半透明礼服被拒绝;改为“不透明、完全覆盖、高领”即可通过。
  • 影片中出现漂移。 吉他在片段中途改变了颜色。要求“几乎静止”,列出不得变化的内容,并且只移动应该移动的部分。
  • 硬编码 model id 或价格。 使用
    recommend
    发现模型,通过
    model_schema_get
    读取数量上限,并使用
    dry_run=true
    估价。