AI Image Art Direction Skill
Skill Name
ai-image-art-direction
Purpose
用于 AI image generation / editing 的 art direction,重点处理视觉世界观、镜头语言、材质、光线、风格约束、生成限制和 AI 常见失真。
When to Use
- 需要为品牌、campaign、海报、社媒、产品或 editorial 生成 AI 图像。
- 需要把创意方向转成可执行 image prompt。
- 需要审查 AI 图像是否有 artifact、品牌不符或生产不可用。
- 需要定义图像的视觉世界、镜头、材料、光线和限制条件。
When Not to Use
- 任务只是普通版式或图文关系时,使用
visual-design 或 graphic-design。
- 任务是品牌系统而非单张图像时,使用
brand-design。
- 任务是 campaign KV 系统时,先使用
campaign-kv,再调用本 Skill。
- 不要用本 Skill 直接模仿活跃艺术家、品牌视觉或受版权保护的图像。
Inputs
- Image purpose
- Brand / campaign context
- Subject and message
- Intended medium
- Aspect ratio and size
- Desired visual world
- References with safety notes
- Constraints: legal, brand, product accuracy, text, logos
- Model or tool limitations if known
Outputs
- Image art direction
- Visual world definition
- Composition and camera logic
- Material and lighting rules
- Prompt-ready instruction
- Negative constraints
- Artifact risk checklist
- Production notes
- Evaluation score
Core Judgment
先用 art direction、visual semiotics 和 studio image-system 经验定义视觉世界观,再用 camera / material / lighting 约束生成,最后用 brand safety、artifact evidence 和 layout usability 拒绝不可生产图像。
Decision Logic
- First decision: 定义 image intent:解释产品、建立情绪、承载 campaign、制造封面停留或提供背景场景。
- Second decision: 定义 visual world:真实摄影、editorial still life、documentary、3D、illustration、collage、diagrammatic。
- Third decision: 定义 camera and composition:shot type、angle、lens feel、focal point、depth、negative space。
- Evidence calibration: contemporary AI image sources 可提示风险,但不能替代 art direction 和 production proof。
- Experience rule: campaign image-system 案例显示,图像世界必须可重复生成或延展,否则只是一张 mood image。
- Experience rule: contemporary AI failure patterns 显示“更真实”不等于更可用,产品准确性、裁切和文字安全区更重要。
- Prioritize: concept clarity、material credibility、brand fit、artifact control、layout usability。
- Reduce: 多余主体、无意义细节、风格堆叠、不可控文字、复杂手部或产品结构。
- Suppress: generic cinematic look、plastic AI skin、fake luxury render、过度景深、无来源光。
- What breaks the output: 图像漂亮但不服务 intent;品牌不符;产品细节错误;无法裁切;承载不了文字;AI artifact 明显。
Workflow
- Diagnose: 明确图像用途、媒介、品牌、需要传达的信息。
- Define World: 写出 visual world、时代感、空间、材质和真实度。
- Compose: 定义主体、焦点、镜头、构图、留白和文字安全区。
- Light and Material: 指定光线来源、阴影、表面、质感、颜色温度。
- Validate: 对照 visual semiotics、brand systems、studio art direction 和 contemporary AI risk sources。
- Constraint: 写出 negative constraints、法律限制、品牌限制、产品准确性限制。
- Prompt: 生成 prompt-ready instruction,避免词堆。
- Critique: 检查 artifact、brand fit、layout fit、production risk。
- Refine: 用具体修正指令迭代,不随机加风格词。
Output Format
- Handoff Target: 交付给 image generation operator / art director / designer / retoucher / production team。
- Asset Type: prompt direction、reference boundary、image spec、retouch notes、layout-safe asset。
- Image Purpose: 图像要完成什么任务。
- Visual World: 画面世界、媒介逻辑、真实度。
- Subject / Scene: 主体、环境、动作、关系。
- Camera / Composition: 景别、角度、焦点、留白、裁切。
- Lighting / Materiality: 光线、阴影、材质、表面。
- Color Direction: 色温、主色、强调色、饱和度。
- Prompt-ready Direction: 可直接给图像模型的指令。
- Negative Constraints: 必须避免的内容和失真。
- Locked Image Controls: 必须保持不变的主体、镜头、材质、光线、品牌限制。
- Adaptation Requirements: poster、KV、social、web、retouch、crop 的要求。
- Validation Binding: canonical / studio / contemporary references。
- Production Notes: ratio、resolution、safe area、retouching。
- Evaluation Score: 按本 Skill 评分。
Quality Bar
AI 图像必须服务明确设计意图,视觉世界一致,材质和光线可信,artifact 可控,并能在目标版面中被裁切、排版和交付。
Anti-Patterns
- Prompt stuffing: 堆大量风格词但没有视觉世界和构图逻辑;不能作为 validated art direction。
- Generic cinematic: 任何图都写 cinematic lighting,导致无差异影像。
- Plastic AI skin: 人物皮肤过度光滑、毛孔消失、表情假。
- Fake material: 金属、玻璃、布料、纸张没有真实物理逻辑。
- Impossible product: 产品结构、logo、文字、包装细节错误。
- Layout unusable image: 没有留白或安全区,无法放标题或裁切。
- Unsafe imitation: 直接模仿活跃艺术家、品牌广告或可识别摄影风格。
- Surreal without concept: 超现实元素存在但不服务 message。
- Case-backed evidence: Gretel image systems 显示重复行为和系列一致性比单张惊艳更重要。
- Case-backed evidence: Collins expressive worlds 显示视觉世界必须有品牌语义边界。
- Case-backed evidence: Studio Dumbar controlled experimentation 显示实验图像必须有系统限制。
- Case-backed evidence: contemporary AI imagery patterns 显示 plastic skin 和 fake materials 会迅速降低可信度。
- Case-backed evidence: campaign layout use 显示没有 safe area 的图像会在后期排版中失败。
Evaluation Criteria
1-5 分:
- Intent fit: 图像是否服务任务。
- Visual world coherence: 世界观是否一致。
- Composition clarity: 主体和焦点是否明确。
- Camera language: 镜头、角度、景深是否有意图。
- Material credibility: 材质、光线、表面是否可信。
- Artifact control: 人体、文字、logo、产品细节是否安全。
- Reference safety: 是否避免直接模仿和版权 / 品牌风险。
- Production usability: 是否可裁切、排版、导出和修图。
- Real-world performance: 图像在 poster、KV、social crop 中是否保留主体和文字空间。
- Real-world performance: 产品、人物、logo、文字细节是否能通过修图成本检查。
- Generation stability: 多轮生成是否能保持主体、材质、镜头和品牌约束。
- Prompt loss tolerance: prompt 被模型弱化后,关键限制是否仍可通过 recovery 找回。
- Handoff integrity: 生成、修图、排版、授权边界是否足以让团队接手。
- Production survivability: 图像经过裁切、放大、压缩和 retouch 后是否仍可用。
Production Notes
必须作为 production handoff 记录:aspect ratio、resolution、safe area、是否需要透明背景、可放文字区域、产品准确性要求、参考限制、授权假设、negative constraints、seed / model notes if available、retouching needs、crop variants、final / concept 状态和人工复核点。
Canonical References
- Visual Semiotics: 支持图像符号、场景意义和品牌语义。
- Brand Systems: 支持图像与品牌识别的一致性。
- Grid & Layout: 支持图像裁切、留白和图文关系。
Studio References
- Gretel: 可验证 motion / image system and campaign visual worlds。
- Collins: 可验证 expressive brand worlds and storytelling。
- Studio Dumbar: 可验证 dynamic visual systems and controlled experimentation。
Validation Notes
- Production validation conclusion: AI 图像进入生产后,最先损耗的是 prompt constraint、产品准确性和 layout safe area。
- Production validation conclusion: model drift 会让同一 art direction 在多轮生成中变成不同视觉世界。
- Production validation conclusion: AI image handoff 必须包含 negative constraints、retouch notes 和 reference safety,否则无法稳定交付。
- Contemporary Validation Sources: D&AD、Creative Review、It’s Nice That 可作为 art direction quality signal;Behance Featured limited 只能作为 weak contemporary signal。
Evidence Gaps
- 缺 prompt-to-retouch deployment proof:从生成到修图到最终版式的损耗链路。
- 缺 model drift proof:不同模型、seed、迭代轮次下 art direction 的稳定性记录。
- 缺 product / logo / text accuracy proof:产品图像在真实交付中的人工修正成本。
Production Proof
First production proof focus: generation stability, model drift, prompt loss and handoff integrity.
本 Skill 的生产验证目标是判断 AI image art direction 从 prompt 进入多轮生成、修图、裁切、版式和交付后,visual world、subject accuracy、materiality、safe area 和 brand safety 是否稳定保留。
Real-world Stress Conditions
- Model drift: 同一 prompt 在不同模型或多轮迭代中偏离视觉世界。
- Prompt loss: 模型忽略 negative constraints、safe area、材质或产品细节。
- Product accuracy: 包装、logo、文字、结构生成错误。
- Human / skin artifacts: 人物表情、手、皮肤和比例失真。
- Crop adaptation: poster、KV、social 比例变化导致主体或留白丢失。
- Retouch handoff: 修图师不知道哪些内容必须保留或禁止改变。
- Reference risk: 参考图过强,生成结果接近受保护品牌或风格。
Production Constraints
- Prompt must separate subject, world, camera, material, lighting and constraints.
- Negative constraints must include product, text, logo, anatomy and reference safety.
- Safe area and crop variants must be defined before generation.
- Image cannot be considered final until artifact and retouch review passes.
- Product and brand assets may require manual compositing instead of generation.
- Handoff must state final vs concept status.
- Model / seed / iteration notes should be preserved when available.
Output Loss Layer
- What degrades first: fine product details、text accuracy、hands、material surface、brand-specific cues。
- What collapses under pressure: visual world consistency、safe area、lighting continuity、series cohesion。
- What survives: clear subject hierarchy、simple composition、strong material direction、locked crop rule。
- What becomes noise: extra surreal details、fake bokeh、gloss、over-texture、irrelevant props。
- What loses brand signal: off-style lighting、generic AI premium look、incorrect product / logo / color。
Failure Under Execution
- Prompt 写得完整,但模型持续忽略文字安全区,导致无法排版。
- 产品图生成漂亮,但 logo、包装结构或文字错误,无法交付。
- 多张系列图使用同一 direction,却出现不同光线、材质和世界观。
- 修图阶段删除了原本用于品牌识别的关键材质或构图。
- 社媒裁切后主体被截断,poster 版本又缺少留白。
- 参考图过强,结果接近可识别品牌广告或摄影风格。
Production Recovery Rules
- If prompt loss appears, split prompt into subject, composition, material, lighting and constraints blocks.
- If product accuracy fails, generate environment separately and composite approved product assets.
- If safe area fails, regenerate with explicit negative space and crop anchor before retouching.
- If model drift appears, reduce style variables and lock camera, material and lighting.
- If artifacts persist, simplify anatomy / object complexity or switch medium logic.
- If brand signal weakens, reintroduce approved color, material, product or layout cue.
- If reference risk is high, abstract structure and remove identifiable surface features.
Injected Cases
- Case Name: Visual semiotics for image meaning
Type: canonical
Why It Matters: 支持图像符号与品牌语义。
What It Validates: image must communicate meaning。
What It Corrects: 纠正无概念超现实。
What It Adds: meaning check。
- Case Name: Photography / art direction composition principles
Type: canonical
Why It Matters: 支持镜头、主体、焦点和光线。
What It Validates: camera language before prompt。
What It Corrects: 纠正 prompt stuffing。
What It Adds: shot logic。
- Case Name: Grid & layout principles for image use
Type: canonical
Why It Matters: 支持留白、裁切和文字安全区。
What It Validates: image must fit layout。
What It Corrects: 纠正 wallpaper image。
What It Adds: layout usability。
- Case Name: Gretel image / motion systems
Type: studio
Why It Matters: 展示图像作为系统的一部分。
What It Validates: repeatable visual behavior。
What It Corrects: 纠正单张惊艳。
What It Adds: image system logic。
- Case Name: Collins expressive brand worlds
Type: studio
Why It Matters: 展示视觉世界和品牌语义的结合。
What It Validates: strong world needs boundaries。
What It Corrects: 纠正风格泛化。
What It Adds: brand world logic。
- Case Name: Studio Dumbar controlled visual experimentation
Type: studio
Why It Matters: 展示实验图像需要系统规则。
What It Validates: controlled experimentation。
What It Corrects: 纠正随机超现实。
What It Adds: constraint-led art direction。
- Case Name: D&AD / Creative Review art direction patterns
Type: contemporary
Why It Matters: 提供 campaign image quality benchmark。
What It Validates: image must support idea。
What It Corrects: 纠正漂亮但无任务图像。
What It Adds: communication benchmark。
- Case Name: Contemporary AI image failure patterns
Type: contemporary
Why It Matters: 提醒 plastic skin、fake materials、text errors、layout failure。
What It Validates: negative constraints must be explicit。
What It Corrects: 纠正“生成后再修”依赖。
What It Adds: artifact anticipation。
Extracted Heuristics
- Learn: define visual world before prompt;Strengthens: art direction;Apply: all AI image tasks。
- Learn: shot logic beats style-word stacking;Strengthens: camera language;Apply: photography-like generation。
- Learn: material credibility must be specified;Strengthens: taste and realism;Apply: product and premium imagery。
- Learn: safe area is an image requirement, not layout afterthought;Strengthens: production fit;Apply: posters and KV。
- Learn: negative constraints should include brand and legal risk;Strengthens: safety;Apply: reference-based generation。
- Learn: repeatability matters for campaign images;Strengthens: system thinking;Apply: series generation。
- Learn: surrealism needs message function;Strengthens: semiotic fit;Apply: conceptual imagery。
- Learn: artifact risk is part of evaluation, not technical cleanup;Strengthens: critique;Apply: final image review。
Case-backed Refinements
- Reinforced: art direction must precede prompt.
- Corrected: realism alone is not enough; layout usability and product accuracy matter more.
- Narrowed: contemporary AI inspiration is weak evidence unless tied to production context.
Inherits
- Canonical System: composition、color、hierarchy、image-text layout。
- Trend Overlay: 用于识别 AI aesthetic fatigue 和饱和风格。
- Taste System: materiality、precision、premium vs generic、maturity。
- Anti-Pattern System: generic AI gloss、plastic skin、unsafe imitation。
- Critique System: artifact and production findings。
- Evaluation System: image-specific scoring。
Version
- Version: V1.0 (production-proven)
- Status: production-proven
- Last updated: 2026-05-02
1---2name: ai-image-art-direction3description: AI Image Art Direction Skill4---5# AI Image Art Direction Skill67## Skill Name89ai-image-art-direction1011## Purpose1213用于 AI image generation / editing 的 art direction,重点处理视觉世界观、镜头语言、材质、光线、风格约束、生成限制和 AI 常见失真。1415## When to Use1617- 需要为品牌、campaign、海报、社媒、产品或 editorial 生成 AI 图像。18- 需要把创意方向转成可执行 image prompt。19- 需要审查 AI 图像是否有 artifact、品牌不符或生产不可用。20- 需要定义图像的视觉世界、镜头、材料、光线和限制条件。2122## When Not to Use2324- 任务只是普通版式或图文关系时,使用 `visual-design` 或 `graphic-design`。25- 任务是品牌系统而非单张图像时,使用 `brand-design`。26- 任务是 campaign KV 系统时,先使用 `campaign-kv`,再调用本 Skill。27- 不要用本 Skill 直接模仿活跃艺术家、品牌视觉或受版权保护的图像。2829## Inputs3031- Image purpose32- Brand / campaign context33- Subject and message34- Intended medium35- Aspect ratio and size36- Desired visual world37- References with safety notes38- Constraints: legal, brand, product accuracy, text, logos39- Model or tool limitations if known4041## Outputs4243- Image art direction44- Visual world definition45- Composition and camera logic46- Material and lighting rules47- Prompt-ready instruction48- Negative constraints49- Artifact risk checklist50- Production notes51- Evaluation score5253## Core Judgment5455先用 art direction、visual semiotics 和 studio image-system 经验定义视觉世界观,再用 camera / material / lighting 约束生成,最后用 brand safety、artifact evidence 和 layout usability 拒绝不可生产图像。5657## Decision Logic5859- First decision: 定义 image intent:解释产品、建立情绪、承载 campaign、制造封面停留或提供背景场景。60- Second decision: 定义 visual world:真实摄影、editorial still life、documentary、3D、illustration、collage、diagrammatic。61- Third decision: 定义 camera and composition:shot type、angle、lens feel、focal point、depth、negative space。62- Evidence calibration: contemporary AI image sources 可提示风险,但不能替代 art direction 和 production proof。63- Experience rule: campaign image-system 案例显示,图像世界必须可重复生成或延展,否则只是一张 mood image。64- Experience rule: contemporary AI failure patterns 显示“更真实”不等于更可用,产品准确性、裁切和文字安全区更重要。65- Prioritize: concept clarity、material credibility、brand fit、artifact control、layout usability。66- Reduce: 多余主体、无意义细节、风格堆叠、不可控文字、复杂手部或产品结构。67- Suppress: generic cinematic look、plastic AI skin、fake luxury render、过度景深、无来源光。68- What breaks the output: 图像漂亮但不服务 intent;品牌不符;产品细节错误;无法裁切;承载不了文字;AI artifact 明显。6970## Workflow71721. Diagnose: 明确图像用途、媒介、品牌、需要传达的信息。732. Define World: 写出 visual world、时代感、空间、材质和真实度。743. Compose: 定义主体、焦点、镜头、构图、留白和文字安全区。754. Light and Material: 指定光线来源、阴影、表面、质感、颜色温度。765. Validate: 对照 visual semiotics、brand systems、studio art direction 和 contemporary AI risk sources。776. Constraint: 写出 negative constraints、法律限制、品牌限制、产品准确性限制。787. Prompt: 生成 prompt-ready instruction,避免词堆。798. Critique: 检查 artifact、brand fit、layout fit、production risk。809. Refine: 用具体修正指令迭代,不随机加风格词。8182## Output Format8384- Handoff Target: 交付给 image generation operator / art director / designer / retoucher / production team。85- Asset Type: prompt direction、reference boundary、image spec、retouch notes、layout-safe asset。86- Image Purpose: 图像要完成什么任务。87- Visual World: 画面世界、媒介逻辑、真实度。88- Subject / Scene: 主体、环境、动作、关系。89- Camera / Composition: 景别、角度、焦点、留白、裁切。90- Lighting / Materiality: 光线、阴影、材质、表面。91- Color Direction: 色温、主色、强调色、饱和度。92- Prompt-ready Direction: 可直接给图像模型的指令。93- Negative Constraints: 必须避免的内容和失真。94- Locked Image Controls: 必须保持不变的主体、镜头、材质、光线、品牌限制。95- Adaptation Requirements: poster、KV、social、web、retouch、crop 的要求。96- Validation Binding: canonical / studio / contemporary references。97- Production Notes: ratio、resolution、safe area、retouching。98- Evaluation Score: 按本 Skill 评分。99100## Quality Bar101102AI 图像必须服务明确设计意图,视觉世界一致,材质和光线可信,artifact 可控,并能在目标版面中被裁切、排版和交付。103104## Anti-Patterns105106- Prompt stuffing: 堆大量风格词但没有视觉世界和构图逻辑;不能作为 validated art direction。107- Generic cinematic: 任何图都写 cinematic lighting,导致无差异影像。108- Plastic AI skin: 人物皮肤过度光滑、毛孔消失、表情假。109- Fake material: 金属、玻璃、布料、纸张没有真实物理逻辑。110- Impossible product: 产品结构、logo、文字、包装细节错误。111- Layout unusable image: 没有留白或安全区,无法放标题或裁切。112- Unsafe imitation: 直接模仿活跃艺术家、品牌广告或可识别摄影风格。113- Surreal without concept: 超现实元素存在但不服务 message。114- Case-backed evidence: Gretel image systems 显示重复行为和系列一致性比单张惊艳更重要。115- Case-backed evidence: Collins expressive worlds 显示视觉世界必须有品牌语义边界。116- Case-backed evidence: Studio Dumbar controlled experimentation 显示实验图像必须有系统限制。117- Case-backed evidence: contemporary AI imagery patterns 显示 plastic skin 和 fake materials 会迅速降低可信度。118- Case-backed evidence: campaign layout use 显示没有 safe area 的图像会在后期排版中失败。119120## Evaluation Criteria1211221-5 分:123124- Intent fit: 图像是否服务任务。125- Visual world coherence: 世界观是否一致。126- Composition clarity: 主体和焦点是否明确。127- Camera language: 镜头、角度、景深是否有意图。128- Material credibility: 材质、光线、表面是否可信。129- Artifact control: 人体、文字、logo、产品细节是否安全。130- Reference safety: 是否避免直接模仿和版权 / 品牌风险。131- Production usability: 是否可裁切、排版、导出和修图。132- Real-world performance: 图像在 poster、KV、social crop 中是否保留主体和文字空间。133- Real-world performance: 产品、人物、logo、文字细节是否能通过修图成本检查。134- Generation stability: 多轮生成是否能保持主体、材质、镜头和品牌约束。135- Prompt loss tolerance: prompt 被模型弱化后,关键限制是否仍可通过 recovery 找回。136- Handoff integrity: 生成、修图、排版、授权边界是否足以让团队接手。137- Production survivability: 图像经过裁切、放大、压缩和 retouch 后是否仍可用。138139## Production Notes140141必须作为 production handoff 记录:aspect ratio、resolution、safe area、是否需要透明背景、可放文字区域、产品准确性要求、参考限制、授权假设、negative constraints、seed / model notes if available、retouching needs、crop variants、final / concept 状态和人工复核点。142143## Canonical References144145- Visual Semiotics: 支持图像符号、场景意义和品牌语义。146- Brand Systems: 支持图像与品牌识别的一致性。147- Grid & Layout: 支持图像裁切、留白和图文关系。148149## Studio References150151- Gretel: 可验证 motion / image system and campaign visual worlds。152- Collins: 可验证 expressive brand worlds and storytelling。153- Studio Dumbar: 可验证 dynamic visual systems and controlled experimentation。154155## Validation Notes156157- Production validation conclusion: AI 图像进入生产后,最先损耗的是 prompt constraint、产品准确性和 layout safe area。158- Production validation conclusion: model drift 会让同一 art direction 在多轮生成中变成不同视觉世界。159- Production validation conclusion: AI image handoff 必须包含 negative constraints、retouch notes 和 reference safety,否则无法稳定交付。160- Contemporary Validation Sources: D&AD、Creative Review、It’s Nice That 可作为 art direction quality signal;Behance Featured limited 只能作为 weak contemporary signal。161162## Evidence Gaps163164- 缺 prompt-to-retouch deployment proof:从生成到修图到最终版式的损耗链路。165- 缺 model drift proof:不同模型、seed、迭代轮次下 art direction 的稳定性记录。166- 缺 product / logo / text accuracy proof:产品图像在真实交付中的人工修正成本。167168## Production Proof169170First production proof focus: generation stability, model drift, prompt loss and handoff integrity.171172本 Skill 的生产验证目标是判断 AI image art direction 从 prompt 进入多轮生成、修图、裁切、版式和交付后,visual world、subject accuracy、materiality、safe area 和 brand safety 是否稳定保留。173174## Real-world Stress Conditions175176- Model drift: 同一 prompt 在不同模型或多轮迭代中偏离视觉世界。177- Prompt loss: 模型忽略 negative constraints、safe area、材质或产品细节。178- Product accuracy: 包装、logo、文字、结构生成错误。179- Human / skin artifacts: 人物表情、手、皮肤和比例失真。180- Crop adaptation: poster、KV、social 比例变化导致主体或留白丢失。181- Retouch handoff: 修图师不知道哪些内容必须保留或禁止改变。182- Reference risk: 参考图过强,生成结果接近受保护品牌或风格。183184## Production Constraints185186- Prompt must separate subject, world, camera, material, lighting and constraints.187- Negative constraints must include product, text, logo, anatomy and reference safety.188- Safe area and crop variants must be defined before generation.189- Image cannot be considered final until artifact and retouch review passes.190- Product and brand assets may require manual compositing instead of generation.191- Handoff must state final vs concept status.192- Model / seed / iteration notes should be preserved when available.193194## Output Loss Layer195196- What degrades first: fine product details、text accuracy、hands、material surface、brand-specific cues。197- What collapses under pressure: visual world consistency、safe area、lighting continuity、series cohesion。198- What survives: clear subject hierarchy、simple composition、strong material direction、locked crop rule。199- What becomes noise: extra surreal details、fake bokeh、gloss、over-texture、irrelevant props。200- What loses brand signal: off-style lighting、generic AI premium look、incorrect product / logo / color。201202## Failure Under Execution203204- Prompt 写得完整,但模型持续忽略文字安全区,导致无法排版。205- 产品图生成漂亮,但 logo、包装结构或文字错误,无法交付。206- 多张系列图使用同一 direction,却出现不同光线、材质和世界观。207- 修图阶段删除了原本用于品牌识别的关键材质或构图。208- 社媒裁切后主体被截断,poster 版本又缺少留白。209- 参考图过强,结果接近可识别品牌广告或摄影风格。210211## Production Recovery Rules212213- If prompt loss appears, split prompt into subject, composition, material, lighting and constraints blocks.214- If product accuracy fails, generate environment separately and composite approved product assets.215- If safe area fails, regenerate with explicit negative space and crop anchor before retouching.216- If model drift appears, reduce style variables and lock camera, material and lighting.217- If artifacts persist, simplify anatomy / object complexity or switch medium logic.218- If brand signal weakens, reintroduce approved color, material, product or layout cue.219- If reference risk is high, abstract structure and remove identifiable surface features.220221## Injected Cases222223- Case Name: Visual semiotics for image meaning224 Type: canonical225 Why It Matters: 支持图像符号与品牌语义。226 What It Validates: image must communicate meaning。227 What It Corrects: 纠正无概念超现实。228 What It Adds: meaning check。229- Case Name: Photography / art direction composition principles230 Type: canonical231 Why It Matters: 支持镜头、主体、焦点和光线。232 What It Validates: camera language before prompt。233 What It Corrects: 纠正 prompt stuffing。234 What It Adds: shot logic。235- Case Name: Grid & layout principles for image use236 Type: canonical237 Why It Matters: 支持留白、裁切和文字安全区。238 What It Validates: image must fit layout。239 What It Corrects: 纠正 wallpaper image。240 What It Adds: layout usability。241- Case Name: Gretel image / motion systems242 Type: studio243 Why It Matters: 展示图像作为系统的一部分。244 What It Validates: repeatable visual behavior。245 What It Corrects: 纠正单张惊艳。246 What It Adds: image system logic。247- Case Name: Collins expressive brand worlds248 Type: studio249 Why It Matters: 展示视觉世界和品牌语义的结合。250 What It Validates: strong world needs boundaries。251 What It Corrects: 纠正风格泛化。252 What It Adds: brand world logic。253- Case Name: Studio Dumbar controlled visual experimentation254 Type: studio255 Why It Matters: 展示实验图像需要系统规则。256 What It Validates: controlled experimentation。257 What It Corrects: 纠正随机超现实。258 What It Adds: constraint-led art direction。259- Case Name: D&AD / Creative Review art direction patterns260 Type: contemporary261 Why It Matters: 提供 campaign image quality benchmark。262 What It Validates: image must support idea。263 What It Corrects: 纠正漂亮但无任务图像。264 What It Adds: communication benchmark。265- Case Name: Contemporary AI image failure patterns266 Type: contemporary267 Why It Matters: 提醒 plastic skin、fake materials、text errors、layout failure。268 What It Validates: negative constraints must be explicit。269 What It Corrects: 纠正“生成后再修”依赖。270 What It Adds: artifact anticipation。271272## Extracted Heuristics273274- Learn: define visual world before prompt;Strengthens: art direction;Apply: all AI image tasks。275- Learn: shot logic beats style-word stacking;Strengthens: camera language;Apply: photography-like generation。276- Learn: material credibility must be specified;Strengthens: taste and realism;Apply: product and premium imagery。277- Learn: safe area is an image requirement, not layout afterthought;Strengthens: production fit;Apply: posters and KV。278- Learn: negative constraints should include brand and legal risk;Strengthens: safety;Apply: reference-based generation。279- Learn: repeatability matters for campaign images;Strengthens: system thinking;Apply: series generation。280- Learn: surrealism needs message function;Strengthens: semiotic fit;Apply: conceptual imagery。281- Learn: artifact risk is part of evaluation, not technical cleanup;Strengthens: critique;Apply: final image review。282283## Case-backed Refinements284285- Reinforced: art direction must precede prompt.286- Corrected: realism alone is not enough; layout usability and product accuracy matter more.287- Narrowed: contemporary AI inspiration is weak evidence unless tied to production context.288289## Inherits290291- Canonical System: composition、color、hierarchy、image-text layout。292- Trend Overlay: 用于识别 AI aesthetic fatigue 和饱和风格。293- Taste System: materiality、precision、premium vs generic、maturity。294- Anti-Pattern System: generic AI gloss、plastic skin、unsafe imitation。295- Critique System: artifact and production findings。296- Evaluation System: image-specific scoring。297298## Version299300- Version: V1.0 (production-proven)301- Status: production-proven302- Last updated: 2026-05-02