🎨 不要再凭感觉猜图像与视频提示词了。

🎨 STOP GUESSING AT IMAGE & VIDEO PROMPTS. ✍️ TYPE ONE SHORT IDEA — AN AI SKILL EXPANDS IT INTO A MODEL-READY, CINEMATIC PROMPT FOR MIDJOURNEY, FLUX, GPT IMAGE, SEEDANCE, KLING & MORE. Prompts feel thin? Motion looks wrong? Same idea, different model every time? One sentence becomes a production prompt with lighting, camera, style, negatives, and video motion — tuned per engine. It's called the AI Visual Prompt Enhancer. Full code below — copy everything, save as SKILL.md, and start generating. █ █ SKILL CODE START — copy everything below █ █ --- name: ai-visual-prompt-enhancer description: > Transform simple ideas into production-ready prompts for AI text-to-image (Midjourney V8, GPT Image 2, Flux.2, Ideogram 4, Seedream, Stable Diffusion 3.5) and text-to-video (Seedance 2.0, Kling 3.0, Runway Gen-4.5, Google Veo 3.1, Sora 2, Luma Ray 3, MiniMax Hailuo, Pika) generation. Use when the user prefixes a message with /p_img, /p_vid, or asks to enhance/create a prompt for AI image or video generation. --- # AI Visual Prompt Enhancer ## Trigger When the user's message starts with `/p_img` or `/p_vid`, OR when the user asks to "enhance this prompt for Midjourney", "write a GPT Image prompt", "create a prompt for Kling", "Seedance prompt", or any similar request targeting AI image/video generation. - `/p_img` → text-to-image prompt enhancement - `/p_vid` → text-to-video prompt enhancement - If the user doesn't specify, **ask** whether the target is image or video (or both) Extract the **core idea** — everything after the trigger prefix. ## Core Principle The user has a raw visual idea but may not know how to communicate it effectively to an AI model. Your job is to **expand it into a precise, structured prompt** that a specific AI model will interpret well — turning "a dog in a park" into a detailed, cinematic, production-ready prompt. **Every run should offer creative direction** while staying faithful to the user's core vision. **Language:** All skill content and default outputs use **American English** (US spelling and wording). Use `color` not colour, `gray` not grey, `center` not centre, `favor` not favour, `fall` for the season when natural. --- ## Part A: Text-to-Image Prompts (`/p_img`) ### The Core Formula Every strong image prompt follows this structure: ``` [Subject] + [Description/Context] + [Style/Aesthetic] + [Technical/Format] ``` 1. **Subject** — Concrete noun(s), the main focus. Not "happiness" but "a smiling child running through wheat". 2. **Description** — Adjectives, action, environment, mood, time of day, weather. 3. **Style** — Art movement, artist reference, medium, or photographic approach. 4. **Technical** — Camera lens, lighting, aspect ratio, resolution keywords, composition. ### Enhancement Dimensions Expand the user's idea across ALL applicable dimensions: #### 1. Subject Detail & Composition - Specific appearance (age, ethnicity, clothing, colors, textures) - Pose and expression (what is the subject doing? how do they look?) - Composition (centered, rule of thirds, close-up, wide shot, Dutch angle) - Subject count and spatial arrangement (foreground/midground/background) #### 2. Environment & Atmosphere - Location specificity (not "a forest" but "a dense redwood forest with moss-covered trunks") - Time of day (golden hour, blue hour, midnight, overcast noon) - Weather and atmospheric conditions (mist, fog, rain, snow, dust motes) - Season indicators (fall leaves, spring blossoms, bare winter branches) #### 3. Lighting Design - Light source (natural sunlight, studio softbox, neon signs, candlelight, moonlight) - Lighting style (Rembrandt lighting, backlighting, rim light, volumetric light, chiaroscuro, soft diffused, harsh direct) - Color temperature (warm amber, cool blue, mixed lighting) - Shadows (deep shadows, soft shadows, no shadows, dramatic contrast) #### 4. Style & Aesthetic Pick ONE dominant style direction: | Category | Keywords | |----------|----------| | **Photography** | photorealistic, 35mm film, 85mm portrait lens, street photography, documentary style, fashion editorial, astrophotography, macro photography, aerial photography | | **Film Stock** | Kodak Vision3 500T, Fujifilm Velvia, Ilford HP5 black and white, cinematic Kodak Gold, 16mm film grain | | **Art Movements** | impressionist, art nouveau, Bauhaus, abstract expressionism, surrealism, baroque, minimalism, pop art | | **Artist Style** | in the style of Studio Ghibli, Hayao Miyazaki aesthetic, Wes Anderson color palette, Gregory Crewdton atmosphere, Annie Leibovitz portraiture | | **Medium** | oil painting, watercolor, charcoal sketch, pencil drawing, ink wash, acrylic, pastels, colored pencils | | **Digital Art** | 3D render, Unreal Engine 5, Octane render, pixel art, vector illustration, cel-shaded, low-poly | | **Mood** | cinematic, ethereal, dystopian, dreamy, dark and moody, bright and cheerful, melancholic, epic, intimate | #### 5. Technical Specifications - Camera: `shot with 85mm f/1.4 lens`, `wide-angle 24mm`, `50mm standard`, `telephoto 200mm` - Depth of field: `shallow depth of field`, `bokeh background`, `deep focus`, `tilt-shift` - Quality boosters: `8K resolution`, `highly detailed`, `sharp focus`, `professional quality`, `intricate details` - Aspect ratio: specify the target ratio (see model-specific section) #### 6. Negative Prompt (where supported) Elements to exclude (Flux, SD, Midjourney `--no`, some platforms): - **Quality**: `blurry, low quality, pixelated, grainy, artifacts, noise, low resolution` - **Anatomy**: `extra fingers, deformed hands, distorted face, bad anatomy, extra limbs, asymmetrical eyes` - **Unwanted**: `text, watermark, signature, logo, date stamp, cropped, cut off` - **Style conflicts**: When wanting photorealism → `cartoon, anime, 3D render, painting` ### Model-Specific Targeting (Image) When enhancing, adapt the prompt structure to the target model: #### Midjourney V8 / V8.1 - Prefers **concise comma-separated phrases**, 4–10 key descriptors - Append parameters: `--ar 16:9`, `--style raw`, `--v 8`, `--no text watermark` - Example: `sleek black smartwatch, white marble surface, soft studio lighting, shallow depth of field, 8K --ar 16:9 --style raw` - Short, high-signal phrases over long paragraphs - Positional weighting: `desert::2 camel::1` (desert is twice as important) - Best for: artistic exploration, cinematic aesthetics, strong art direction #### GPT Image 2 (OpenAI) - Prefers **conversational, natural language** — full sentences - Excellent prompt adherence, text-in-image, and multi-turn refinement - Example: `Create a photorealistic image of a golden retriever catching a frisbee mid-jump in Central Park. The scene is set during golden hour with warm sunlight filtering through fall trees. Shot with an 85mm lens for a shallow depth of field, with the background beautifully blurred.` - Strong at complex compositions, editing workflows, and readable text inside the image - More forgiving of long, natural descriptions than keyword-style models #### Flux.2 (Pro / Flex / Max — Black Forest Labs) - Superior for **photorealism, color accuracy, and commercial work** - Responds well to **detailed technical descriptions** - Example: `Professional headshot portrait of a confident businesswoman, age 35, shoulder-length brown hair, warm smile, navy blazer, soft studio lighting, blurred office background, shot with 85mm lens, f/2.8, slight film grain, editorial quality` - Clean, descriptive language without excessive quality spam - Strong multi-reference and product consistency on newer Flux.2 variants #### Stable Diffusion 3.5 / SDXL-class open models - Rewards **structured, weighted keywords** - Supports **negative prompts** (dedicated field) - Advanced: `(keyword:1.5)` for emphasis, `(keyword:0.7)` to de-emphasize - Supports LoRAs, ControlNets, IP-Adapter for precise control - Example: `(masterpiece, best quality, ultra-detailed:1.2), portrait of a samurai, cherry blossom garden, cinematic lighting, (dramatic pose:1.3), shot on 35mm film` #### Ideogram 4 / Seedream 4.5 / Recraft - **Ideogram 4**: Excels at **text rendering and typography** — specify text content explicitly - **Seedream 4.5**: Strong product shots, 4K detail, text-heavy marketing creatives - **Recraft**: Design-oriented (logos, brand assets, vector-friendly outputs) - Example (Ideogram): `Coffee shop sign that says "COFFEE" in clear bold sans-serif letters, rustic wooden board, warm interior background` ### Output Format (Image) ``` ## Original Idea > [the core idea extracted from /p_img] ## Enhanced Prompt ### [Model Name] Version ``` [prompt optimized for this specific model] ``` ### Prompt Breakdown - **Subject:** [what's depicted] - **Environment:** [setting and atmosphere] - **Lighting:** [light source and quality] - **Style:** [aesthetic direction] - **Technical:** [camera, lens, format details] - **Negative:** [what to avoid — if applicable] ### Variations 1. [alternative style direction — e.g., "same scene, but oil painting style"] 2. [alternative mood — e.g., "night version with neon lighting"] 3. [alternative composition — e.g., "aerial drone perspective"] --- **Tip:** Change one element at a time between generations to learn what works. ``` ### Variation Engine (Image) When the same idea is enhanced multiple times, vary: | Dimension | Variation Examples | |-----------|-------------------| | **Style** | Photorealistic → Cinematic → Oil Painting → Anime/Cel-shaded → 3D Render → Watercolor | | **Lighting** | Golden hour → Moody overcast → Neon/night → Studio → Volumetric sunbeams → Bioluminescent | | **Camera** | Close-up portrait → Wide landscape → Aerial drone → Macro → Low angle → Over-the-shoulder | | **Mood** | Serene → Dramatic → Playful → Mysterious → Epic → Intimate | | **Era** | Contemporary → Retro/vintage → Futuristic → Historical → Fantasy | --- ## Part B: Text-to-Video Prompts (`/p_vid`) Video prompts differ fundamentally from image prompts: you're not describing a static scene — you're describing a **shot that unfolds through time**. ### The Core Formula ``` [Camera Shot + Movement] + [Subject + Action] + [Environment + Scene Motion] + [Lighting] + [Style] + [Audio cue if supported] ``` **Order matters** — most video models weight earlier concepts more heavily. ### Enhancement Dimensions #### 1. Camera Shot & Movement This is the **first** element — it establishes framing: | Camera Type | Keywords | |-------------|----------| | **Static** | locked camera, tripod shot, still camera | | **Handheld** | handheld camera, shaky handheld, documentary handheld | | **Dolly/Track** | slow dolly forward, tracking shot, lateral dolly, dolly out | | **Pan/Tilt** | slow pan left, tilt up, tilt down, arc shot, orbit | | **Zoom** | slow push-in, pull-back reveal, gentle zoom in | | **Aerial** | aerial drone shot, drone flyover, bird's eye view, drone pullback | | **Special** | crash zoom, whip pan, Steadicam, gimbal shot, overhead top-down | #### 2. Subject Motion (The Action) Describe **what moves and how**: - Keep it **simple and specific** — one or two actions per subject - Refer to subjects generally: `the subject turns`, `she raises her hand`, `he walks away` - For multiple subjects: `The woman on the left walks forward. The man on the right remains still.` - **Good motion descriptors**: walks, turns, smiles, waves, runs, falls, transforms, floats, ripples, blooms - **Avoid complex choreography** — models handle slow, deliberate motion best - For characters: prefer **medium shots** (faces too close = more artifacts) #### 3. Scene / Environment Motion How the world reacts: - **Implied** (via adjectives): `dusty desert` → implies dust kicks up - **Explicit**: `dust trails behind them as they move` - Examples: `gentle waves lap the shore`, `leaves rustle in the breeze`, `steam rises from the mug`, `neon reflections shimmer on wet pavement`, `storm clouds gather` #### 4. Lighting (Temporal) Lighting in video can **change over time**: - Static: `golden hour lighting`, `moonlit scene`, `neon-lit alley` - Dynamic: `lightning flashes illuminate the clouds`, `sun breaks through clouds`, `lights flicker on` #### 5. Style, Technical & Audio - Style: `cinematic`, `live-action`, `documentary style`, `anime`, `stop motion`, `claymation` - Film stock: `Kodak Vision3`, `Fujifilm Velvia`, `16mm film`, `anamorphic lens` - Quality: `smooth motion`, `high quality`, `4K`, `physically accurate motion` - Format: `vertical 9:16`, `seamless loop`, `slow motion`, `time-lapse` - **Native audio** (Seedance, Kling 3, Veo, Sora 2): optionally add short sound cues — `soft rain ambiance`, `distant train horn`, `footsteps on wet pavement` — keep them brief ### Video Prompt Best Practices 1. **Start simple** — Begin with subject + one motion, then add details 2. **Positive phrasing only** — Avoid "no blur", "don't shake" — describe what you WANT 3. **Focus on motion, not description** — In img2vid, the input image handles visuals; the text prompt describes movement 4. **Avoid commands** — Don't write "make the camera move" — write "slow camera pan" 5. **One element at a time** — Add camera motion first, then subject motion, then environment 6. **Keep it direct** — `the subject walks forward` > `I would like to see a person walking` 7. **Match duration** — Prefer motions that fit ~5–15s clips; multi-shot only when the model supports it (Seedance, Kling 3) ### Common Motion Fix-Phrases | Problem | Fix Phrase | |---------|-----------| | Motion too fast/jerky | Add `slow motion`, `smooth camera movement`, `gentle` | | Objects disappear | Reduce clip to 4–5 seconds, simplify scene | | Style is wrong | Add film stock refs or clear aesthetic labels | | Result completely wrong | Radically shorten to subject + one motion, then iterate | | Face artifacts | Use medium shot instead of close-up; avoid faces walking toward camera | | Rubber limbs / bad physics | Add `physically accurate motion`, `believable inertia`, `natural weight` | | Audio mismatch | Keep one clear ambient sound; avoid conflicting dialogue + music | ### Model-Specific Targeting (Video) #### Seedance 2.0 (ByteDance) - **Frontier multi-shot model** — strong for short films, ads, music-video style sequences - Supports multimodal references (characters, products, brand consistency across shots) - Native audio + picture in one pass on current generations - Responds to: director-style natural language — camera, action, cuts, sound - Best for: multi-shot storytelling, synced audio, reference-driven consistency - Example: `Multi-shot cinematic ad. Shot 1: product close-up on wet marble, soft rim light, camera slowly pushes in. Shot 2: wide of a runner on a rainy street at night, handheld tracking from the side, water splashes, neon reflections. Soft rain ambiance, subtle electronic pulse, 4K, photorealistic.` - Tip: Keep each shot's action simple; describe transitions clearly when using multi-shot #### Kling 3.0 (Kuaishou) - **Strongest character-driven motion** and realistic human performance; 4K-capable - Excellent physics, longer coherent clips (up to ~15s on current tiers) - Native audio on recent Kling 3 / Omni generations - Responds to: detailed natural language — subject action first, then camera and environment - Best for: people, product commercials, controlled references, action with believable inertia - Example: `A woman in a gray coat stands at a train platform at night. She looks left then right as a train passes behind her with motion blur. Still camera, moody street lighting, soft ambient city noise, cinematic, 4K.` - Tip: Prefer medium shots for faces; state one primary action per beat #### Runway Gen-4.5 - **Power of simplicity** — short prompts work best - Prefers lowercase, direct motion descriptions - Input image is king in img2vid — text focuses on **motion** - Best for: cinematic live-action, creative production workflows, iteration + editing suite - Example: `a handheld camera tracks the subject as they walk across the desert. dust trails behind. cinematic live-action.` - Avoid negative prompts — use positive descriptions only #### Google Veo 3.1 - Premium cinematic quality; strong outdoor, atmospheric, and brand-safe results - Handles detailed lighting, environment, and physics descriptions well - Native audio on current Veo generations - Example: `Aerial drone shot over a Norwegian fjord at sunset, water reflects pink and orange sky, gentle waves against rocky shores, slow forward movement, soft wind ambiance, photorealistic, 4K` #### Sora 2 (OpenAI) - Advanced multi-shot capability, strong physics and longer coherent narratives - Responds to **detailed descriptive language about lighting, physics, and continuous action** - Example: `Continuous shot: camera orbits a glass of iced coffee on a sunlit table as condensation drips down the glass; background cafe soft bokeh; natural window light shifts slightly; subtle room tone` #### Luma Ray 3 / Dream Machine lineage - Strong for dramatic environmental scenes and abstract motion - Responds to vivid descriptive language and camera choreography - Example: `Time-lapse of storm clouds forming over ocean, lightning flashes illuminate the undersides of dark cumulus clouds, wide shot, dramatic, high contrast` #### MiniMax Hailuo / Pika - **Hailuo**: Fast short-form, solid motion for social clips - **Pika**: Stylized looks, aesthetic labels, creative effects; often supports negative prompts - Example (Pika): `Eye-level tracking shot, woman walks through neon-lit alley in rain, cinematic, moody color grade, cyberpunk` - Example (Hailuo): `Slow dolly in on a steaming bowl of ramen, chopsticks lift noodles, warm kitchen light, shallow depth of field, smooth motion` ### Output Format (Video) ``` ## Original Idea > [the core idea extracted from /p_vid] ## Enhanced Prompt ### [Model Name] Version ``` [prompt optimized for this specific model] ``` ### Shot Breakdown - **Camera:** [shot type + movement] - **Subject Motion:** [what the subject does] - **Scene Motion:** [environmental movement] - **Lighting:** [light source and quality] - **Style:** [aesthetic direction] - **Audio:** [if native audio model — brief cue or "none"] ### Variations 1. [alternative camera angle — e.g., "aerial instead of ground-level"] 2. [different time of day — e.g., "night version with neon lights"] 3. [different mood — e.g., "peaceful morning instead of dramatic storm"] ### Tips for This Prompt - [model-specific tips, common pitfalls, iteration advice] --- **Tip:** Start with the simplest version. Once the basic motion works, add one detail at a time. ``` ### Variation Engine (Video) When the same idea is enhanced multiple times, vary: | Dimension | Variation Examples | |-----------|-------------------| | **Camera** | Static locked → Handheld → Drone aerial → Slow dolly → Orbit/arc → Top-down | | **Speed** | Normal speed → Slow motion → Time-lapse → Ultra slow motion | | **Time** | Golden hour → Blue hour → Night → Overcast → Sunrise → Sunset | | **Weather** | Clear → Rain → Fog → Snow → Wind → Storm | | **Style** | Cinematic → Documentary → Anime → Stop motion → Abstract fluid → Live-action | ### Quick Model Picker (Video) | Goal | Prefer | |------|--------| | Multi-shot story / ad with audio | **Seedance 2.0** | | Realistic humans / character consistency | **Kling 3.0** | | Creative iteration + pro editing tools | **Runway Gen-4.5** | | Cinematic outdoor / premium atmospheric | **Veo 3.1** | | Complex physics / long continuous action | **Sora 2** | | Fast social / stylized short clips | **Hailuo / Pika** | --- ## Part C: Cross-Modal Workflow (Image → Video) When the user wants to turn an image into video: 1. **First generate the image** using an enhanced image prompt 2. **Use that image as input** for the video model (img2vid mode) 3. **Video prompt focuses on motion only** — the image provides all visual information 4. Keep the video prompt short: camera movement + subject motion + any new environmental motion (+ brief audio if supported) Example workflow: ``` Step 1: /p_img → Generate "portrait of a warrior in a cyberpunk alley, neon lighting, rain" Step 2: Upload that image to Seedance 2.0, Kling 3.0, or Runway Gen-4.5 (img2vid) Step 3: Video prompt → "slow dolly forward, rain falls, neon lights flicker, the subject slowly turns to look at camera, soft rain ambiance, cinematic" ``` --- ## Rules 1. **Always preserve the core idea** — enhance, don't replace what the user envisions 2. **Be specific** — no "nice lighting", always "golden hour backlighting with rim light" 3. **Match the model** — adapt syntax, structure, and length to the target AI model 4. **Default language: American English** — write all enhanced prompts, breakdowns, tips, and variations in **US English** (color, gray, center, favor, fall). Only switch language if the user explicitly asks for another language 5. **Provide model-specific versions** — when the user doesn't specify, offer 2–3 variants for different current models (for video prefer Seedance 2.0 + Kling 3.0 + one other) 6. **Offer variations** — give 2–3 alternative directions so the user can choose 7. **Explain the breakdown** — so the user understands WHY each element is there 8. **Flag assumptions** — if making a guess about what the user wants, note it 9. **Don't overcomplicate** — a simple scene stays simple, just detailed 10. **Every run is different** — vary the creative direction on re-runs ## Quick Reference: Common Prompt Patterns ### Portraits (Image) `[pose] portrait of [person description], [expression], [background], [lighting], [lens], [style]` ### Landscapes (Image) `[shot type] of [location], [environmental details], [time/weather], [lighting], [camera specs], [quality]` ### Products (Image) `Product photography of [item], [surface/background], [lighting], [angle], [material details], [quality]` ### Nature (Video) `[camera type], [environmental scene], [motion description], [lighting], [style]` ### Urban (Video) `[camera movement] through [city scene], [subject action], [environmental details], [lighting], [style]` ### Character (Video) `[shot type], [subject description and action], [background], [camera], [lighting], [style]` ### Abstract (Video) `[shot type], [abstract motion description], [colors], [background], [speed], [style]` ### Multi-shot Ad (Video — Seedance / Kling 3) `Multi-shot [genre]. Shot 1: [camera + action + light]. Shot 2: [camera + action + light]. [brief audio], [style], [resolution]` █ █ SKILL CODE END █ █ How to install: Same pattern on every agent that supports the Agent Skills standard (SKILL.md folder): <skills-root>/ai-visual-prompt-enhancer/SKILL.md Copy everything between SKILL CODE START and SKILL CODE END into that file. PI AGENT: ~/.agents/skills/ai-visual-prompt-enhancer/SKILL.md Restart pi → /p_img or /p_vid CLAUDE CODE: ~/.claude/skills/ai-visual-prompt-enhancer/SKILL.md (project: .claude/skills/ai-visual-prompt-enhancer/) New session → /p_img or /p_vid CURSOR: ~/.cursor/skills/ai-visual-prompt-enhancer/SKILL.md (also: ~/.agents/skills/ or .cursor/skills/) New chat → /p_img or /p_vid CODEX CLI: ~/.codex/skills/ai-visual-prompt-enhancer/SKILL.md (project: .agents/skills/ai-visual-prompt-enhancer/) New session → /p_img or /p_vid GROK: ~/.grok/skills/ai-visual-prompt-enhancer/SKILL.md (project: .grok/skills/ai-visual-prompt-enhancer/) Auto-reloads → /p_img or /p_vid HERMES / ANY OTHER: Drop the same folder into your agent's skills directory, or paste SKILL.md into custom instructions / system prompt. Trigger: /p_img, /p_vid, or "enhance this prompt for Midjourney / Kling / Seedance" Windows: use %USERPROFILE% instead of ~ e.g. %USERPROFILE%\.claude\skills\... One short idea in → model-ready visual prompt out. Go.

🎨 不要再凭感觉猜图像与视频提示词了。 ✍️ 输入一句简短想法 — AI 技能会自动将其 扩展为可直接送入模型的、电影感十足的提示词， 适用于 MIDJOURNEY、FLUX、GPT IMAGE、SEEDANCE、KLING 以及更多模型。 提示词总觉得太单薄？画面运动总是不对？ 同一个想法换个模型结果就全变了？ 一句话就能变成一条可直接投产的提示词， 自带光线、镜头、风格、负面词， 以及视频运动描述 — 按每个引擎单独调优。 它叫 AI Visual Prompt Enhancer（AI 视觉提示词增强器）。 完整代码在下方 — 全部复制，另存为 SKILL.md，即可开始生成。 █ █ SKILL 代码开始 — 复制下方全部内容 █ █ --- name: ai-visual-prompt-enhancer description: > 将简单的想法转化为可用于 AI 文本生成图像 （Midjourney V8、GPT Image 2、Flux.2、 Ideogram 4、Seedream、Stable Diffusion 3.5） 以及文本生成视频（Seedance 2.0、Kling 3.0、 Runway Gen-4.5、Google Veo 3.1、Sora 2、 Luma Ray 3、MiniMax Hailuo、Pika） 的生产级提示词。当用户以 /p_img、/p_vid 作为消息前缀，或要求为 AI 图像/视频生成 增强/创建提示词时使用。 --- # AI Visual Prompt Enhancer ## 触发条件 当用户的消息以 `/p_img` 或 `/p_vid` 开头， 或者用户提出类似「为 Midjourney 增强这个 提示词」「写一条 GPT Image 提示词」 「为 Kling 写一条提示词」「Seedance 提示词」 等任何针对 AI 图像/视频生成的要求时触发。 - `/p_img` → 文生图提示词增强 - `/p_vid` → 文生视频提示词增强 - 如果用户未指定，**询问**目标是图像 还是视频（或两者都要） 提取 **核心想法** — 即触发前缀之后的所有 内容。 ## 核心原则 用户拥有原始的视觉想法，但可能并不知道 如何有效地传达给 AI 模型。你的工作是 **将其扩展为精确、结构化的提示词**，让 特定的 AI 模型能够很好地理解 — 把「公园里 一只狗」变成一条细节丰富、电影感、可直接 投产的提示词。**每次运行都应提供创作 方向**，同时忠实于用户的核心构想。 **语言：** 所有技能内容与默认输出使用 **美式英语**（美式拼写与措辞）。使用 `color` 而非 colour，`gray` 而非 grey， `center` 而非 centre，`favor` 而非 favour， 季节时使用自然的 `fall`。 --- ## Part A: 文生图提示词 (`/p_img`) ### 核心公式 每一条强力的图像提示词都遵循以下结构： ``` [主体] + [描述/情境] + [风格/美学] + [技术/格式] ``` 1. **主体** — 具体名词，主要焦点。不是 「幸福」，而是「一个微笑的孩子在麦田里 奔跑」。 2. **描述** — 形容词、动作、环境、氛围、 一天中的时间、天气。 3. **风格** — 艺术流派、艺术家参考、媒介， 或摄影手法。 4. **技术** — 相机镜头、光线、宽高比、 分辨率关键词、构图。 ### 增强维度 在所有适用的维度上扩展用户的想法： #### 1. 主体细节与构图 - 具体外观（年龄、种族、服装、颜色、 材质） - 姿势与表情（主体在做什么？看起来怎样？） - 构图（居中、三分法、特写、远景、 荷兰角） - 主体数量与空间安排（前/中/后景） #### 2. 环境与氛围 - 地点的具体性（不是「一片森林」，而是 「一片浓密的红杉林，树干上覆满苔藓」） - 一天中的时间（黄金时刻、蓝色时刻、 午夜、阴天的正午） - 天气与大气条件（薄雾、雾、雨、雪、 浮尘） - 季节指示（秋叶、春花、冬季光秃的 树枝） #### 3. 光线设计 - 光源（自然阳光、工作室柔光箱、霓虹 灯、烛光、月光） - 光线风格（伦勃朗光、逆光、轮廓光、 体积光、明暗对照法、柔和漫射光、强烈 直射光） - 色温（暖琥珀色、冷蓝色、混合光） - 阴影（深阴影、柔和阴影、无阴影、 强烈反差） #### 4. 风格与美学 选择 ONE 一种主要风格方向： | 类别 | 关键词 | |----------|----------| | **摄影** | 照片级真实感、35mm 胶片、 85mm 人像镜头、街拍、纪录片风格、 时尚编辑、天体摄影、微距摄影、航拍摄影 | | **胶片** | Kodak Vision3 500T、 富士 Velvia、Ilford HP5 黑白、 电影感 Kodak Gold、16mm 胶片颗粒 | | **艺术流派** | 印象派、新艺术运动、 包豪斯、抽象表现主义、超现实主义、 巴洛克、极简主义、波普艺术 | | **艺术家风格** | 吉卜力工作室风格、 宫崎骏美学、韦斯·安德森配色、 格雷戈里·克鲁德森氛围、 安妮·莱博维茨人像 | | **媒介** | 油画、水彩、炭笔速写、 铅笔画、水墨、丙烯、粉彩、彩色铅笔 | | **数字艺术** | 3D 渲染、Unreal Engine 5、 Octane 渲染、像素艺术、矢量插画、 卡通描边、低多边形 | | **氛围** | 电影感、空灵、反乌托邦、 梦幻、阴郁暗黑、明快欢乐、忧郁、 史诗、亲密 | #### 5. 技术规格 - 相机：`shot with 85mm f/1.4 lens`、 `wide-angle 24mm`、`50mm standard`、 `telephoto 200mm` - 景深：`shallow depth of field`、 `bokeh background`、`deep focus`、 `tilt-shift` - 画质增强词：`8K resolution`、 `highly detailed`、`sharp focus`、 `professional quality`、`intricate details` - 宽高比：指定目标比例（见模型特定 章节） #### 6. 负面提示词（在支持时） 需要排除的元素（Flux、SD、Midjourney `--no`，以及部分平台）： - **质量**：`blurry, low quality, pixelated, grainy, artifacts, noise, low resolution` - **解剖**：`extra fingers, deformed hands, distorted face, bad anatomy, extra limbs, asymmetrical eyes` - **不需要的元素**：`text, watermark, signature, logo, date stamp, cropped, cut off` - **风格冲突**：当希望照片级真实感时 → `cartoon, anime, 3D render, painting` ### 针对模型的特定调优（图像） 在增强时，根据目标模型调整提示词结构： #### Midjourney V8 / V8.1 - 偏好 **简洁的逗号分隔短语**，4–10 个 关键描述词 - 附加参数：`--ar 16:9`、`--style raw`、 `--v 8`、`--no text watermark` - 示例：`sleek black smartwatch, white marble surface, soft studio lighting, shallow depth of field, 8K --ar 16:9 --style raw` - 简短高信号的短语优于长段落 - 位置权重：`desert::2 camel::1` （desert 的重要性是 camel 的两倍） - 最适合：艺术探索、电影感美学、强力 艺术指导 #### GPT Image 2（OpenAI） - 偏好 **对话式、自然语言** — 完整的 句子 - 出色的提示词遵循度、图内文字生成， 以及多轮迭代精修 - 示例：`Create a photorealistic image of a golden retriever catching a frisbee mid-jump in Central Park. The scene is set during golden hour with warm sunlight filtering through fall trees. Shot with an 85mm lens for a shallow depth of field, with the background beautifully blurred.` - 擅长复杂构图、编辑工作流，以及图像中 可读的文字 - 比关键词型模型更宽容冗长的自然描述 #### Flux.2（Pro / Flex / Max — Black Forest Labs） - 在 **照片级真实感、色彩准确性、商业 作品** 上表现卓越 - 对 **详尽的技术描述** 响应良好 - 示例：`Professional headshot portrait of a confident businesswoman, age 35, shoulder-length brown hair, warm smile, navy blazer, soft studio lighting, blurred office background, shot with 85mm lens, f/2.8, slight film grain, editorial quality` - 干净、描述性的语言，无需堆砌画质词 - 在较新的 Flux.2 变体上具备强大的 多参考与产品一致性能力 #### Stable Diffusion 3.5 / SDXL 类开源模型 - 奖励 **结构化、加权关键词** - 支持 **负面提示词**（独立字段） - 进阶：`(keyword:1.5)` 用于强调， `(keyword:0.7)` 用于弱化 - 支持 LoRAs、ControlNets、IP-Adapter 以 实现精确控制 - 示例：`(masterpiece, best quality, ultra-detailed:1.2), portrait of a samurai, cherry blossom garden, cinematic lighting, (dramatic pose:1.3), shot on 35mm film` #### Ideogram 4 / Seedream 4.5 / Recraft - **Ideogram 4**：擅长 **文字渲染与字体 排版** — 明确指定文字内容 - **Seedream 4.5**：强力产品图、4K 细节、 文字密集的营销创意 - **Recraft**：面向设计（Logo、品牌资产、 矢量友好的输出） - 示例（Ideogram）：`Coffee shop sign that says "COFFEE" in clear bold sans-serif letters, rustic wooden board, warm interior background` ### 输出格式（图像） ``` ## Original Idea > [从 /p_img 提取的核心想法] ## Enhanced Prompt ### [模型名称] Version ``` [针对此特定模型优化的提示词] ``` ### Prompt Breakdown - **Subject:** [描绘的内容] - **Environment:** [场景与氛围] - **Lighting:** [光源与质感] - **Style:** [美学方向] - **Technical:** [相机、镜头、格式细节] - **Negative:** [需要避免的内容 — 若适用] ### Variations 1. [替代风格方向 — 例「同一场景，但改为 油画风格」] 2. [替代氛围 — 例「夜晚版，霓虹光」] 3. [替代构图 — 例「航拍无人机视角」] --- **Tip:** 每次生成只改动一个元素，以便学习 哪些有效。 ``` ### 变化引擎（图像） 当同一想法被多次增强时，按以下维度变化： | 维度 | 变化示例 | |-----------|-------------------| | **风格** | 照片级真实 → 电影感 → 油画 → 动漫/卡通描边 → 3D 渲染 → 水彩 | | **光线** | 黄金时刻 → 阴郁阴天 → 霓虹/ 夜景 → 工作室 → 体积光光柱 → 生物发光 | | **相机** | 特写人像 → 广阔风景 → 航拍 无人机 → 微距 → 低角度 → 过肩镜头 | | **氛围** | 宁静 → 戏剧化 → 俏皮 → 神秘 → 史诗 → 亲密 | | **年代** | 当代 → 复古/怀旧 → 未来 → 历史 → 奇幻 | --- ## Part B: 文生视频提示词 (`/p_vid`) 视频提示词与图像提示词有着根本性的不同： 你不是在描述一个静态场景 — 而是在描述一个 **随时间展开的镜头**。 ### 核心公式 ``` [镜头 + 运动] + [主体 + 动作] + [环境 + 场景运动] + [光线] + [风格] + [音频线索（若支持）] ``` **顺序很重要** — 大多数视频模型会对排在 前面的概念赋予更高权重。 ### 增强维度 #### 1. 镜头与运动 这是 **第一要素** — 它确立了构图： | 镜头类型 | 关键词 | |-------------|----------| | **静态** | locked camera、tripod shot、 still camera | | **手持** | handheld camera、shaky handheld、 documentary handheld | | **推轨/跟拍** | slow dolly forward、 tracking shot、lateral dolly、dolly out | | **摇/俯仰** | slow pan left、tilt up、 tilt down、arc shot、orbit | | **变焦** | slow push-in、pull-back reveal、 gentle zoom in | | **航拍** | aerial drone shot、drone flyover、 bird's eye view、drone pullback | | **特殊** | crash zoom、whip pan、Steadicam、 gimbal shot、overhead top-down | #### 2. 主体运动（动作） 描述 **什么在动以及如何动**： - 保持 **简单