character-design-sheet

作者: halt-catch-fire

利用参考图和LoRA技术实现AI生成图像中角色的一致性。涵盖转面视图、表情表、配色方案及风格一致性技巧。适用于:角色设计、游戏美术、插画、动画、漫画、视觉小说。触发词:角色设计、角色表、角色一致性、角色参考、转面图、表情表、角色美术、一致角色、角色概念、参考图、角色创建、OC设计……

npx skills add https://github.com/halt-catch-fire/skills --skill character-design-sheet

Install the belt CLI skill: npx skills add belt-sh/cli

Character Design Sheet

Create consistent characters across multiple AI-generated images via inference.sh CLI.

Quick Start

Requires inference.sh CLI (belt). Install instructions

belt login

# Generate a character concept
belt app run falai/flux-dev-lora --input '{
  "prompt": "character design reference sheet, front view of a young woman with short red hair, green eyes, wearing a blue jacket and white t-shirt, full body, white background, clean lines, concept art style, character turnaround",
  "width": 1024,
  "height": 1024
}'

The Consistency Problem

AI image generation produces different-looking characters every time, even with the same prompt. This is the #1 challenge in AI art for any project requiring the same character across multiple images.

Solutions (Ranked by Effectiveness)

TechniqueConsistencyEffortBest For
FLUX LoRA (trained on character)Very highHigh (requires training data)Ongoing projects, many images
Detailed description anchorMedium-highLowQuick projects, few images
Same seed + similar promptMediumLowVariations of single pose
Image-to-image refinementMediumMediumRefining existing images
Reference image in promptVariesLowWhen model supports it

Reference Sheet Types

1. Turnaround Sheet

Shows the character from multiple angles:

┌────────┬────────┬────────┬────────┐
│        │        │        │        │
│ FRONT  │  3/4   │  SIDE  │  BACK  │
│  VIEW  │  VIEW  │  VIEW  │  VIEW  │
│        │        │        │        │
└────────┴────────┴────────┴────────┘
# Generate front view
belt app run falai/flux-dev-lora --input '{
  "prompt": "character design, front view, young woman with short asymmetric red hair, bright green eyes, wearing navy blue bomber jacket over white graphic tee, dark jeans, red sneakers, standing in neutral pose, full body, clean white background, concept art, sharp details",
  "width": 768,
  "height": 1024
}' --no-wait

# Generate 3/4 view (same description)
belt app run falai/flux-dev-lora --input '{
  "prompt": "character design, three-quarter view, young woman with short asymmetric red hair, bright green eyes, wearing navy blue bomber jacket over white graphic tee, dark jeans, red sneakers, standing, full body, clean white background, concept art, sharp details",
  "width": 768,
  "height": 1024
}' --no-wait

# Generate side view
belt app run falai/flux-dev-lora --input '{
  "prompt": "character design, side profile view, young woman with short asymmetric red hair, bright green eyes, wearing navy blue bomber jacket over white graphic tee, dark jeans, red sneakers, standing, full body, clean white background, concept art, sharp details",
  "width": 768,
  "height": 1024
}' --no-wait

# Generate back view
belt app run falai/flux-dev-lora --input '{
  "prompt": "character design, back view, young woman with short asymmetric red hair, wearing navy blue bomber jacket over white graphic tee, dark jeans, red sneakers, standing, full body, clean white background, concept art, sharp details",
  "width": 768,
  "height": 1024
}' --no-wait

# Stitch into reference sheet (two images per call; repeat to add views)
belt app run infsh/stitch-images --input '{
  "image1": "front.png",
  "image2": "three-quarter.png",
  "orientation": "horizontal"
}'

2. Expression Sheet

Shows the character's face with different emotions:

┌────────┬────────┬────────┐
│NEUTRAL │ HAPPY  │ ANGRY  │
│        │        │        │
├────────┼────────┼────────┤
│  SAD   │SURPRISE│THINKING│
│        │        │        │
└────────┴────────┴────────┘

Minimum 6 expressions: neutral, happy, angry, sad, surprised, thinking.

# Neutral
belt app run falai/flux-dev-lora --input '{
  "prompt": "character portrait, close-up face, young woman with short red hair and green eyes, neutral calm expression, clean white background, concept art, consistent character design",
  "width": 512,
  "height": 512
}' --no-wait

# Happy
belt app run falai/flux-dev-lora --input '{
  "prompt": "character portrait, close-up face, young woman with short red hair and green eyes, warm genuine smile, happy expression, clean white background, concept art, consistent character design",
  "width": 512,
  "height": 512
}' --no-wait

# Angry
belt app run falai/flux-dev-lora --input '{
  "prompt": "character portrait, close-up face, young woman with short red hair and green eyes, furrowed brows, angry determined expression, clean white background, concept art, consistent character design",
  "width": 512,
  "height": 512
}' --no-wait

# (Continue for sad, surprised, thinking...)

3. Outfit/Costume Sheet

Multiple outfits for the same character:

OutfitDescription
CasualBomber jacket, t-shirt, jeans
WorkBlazer, button-down, slacks
AthleticSports bra, leggings, running shoes
FormalEvening dress, heels

4. Color Palette Sheet

Document exact colors for consistency:

CHARACTER: Maya Chen

Skin:    ████ #F5D0A9 (warm beige)
Hair:    ████ #C0392B (auburn red)
Eyes:    ████ #27AE60 (emerald green)
Jacket:  ████ #2C3E50 (navy blue)
T-shirt: ████ #ECF0F1 (off-white)
Jeans:   ████ #34495E (dark slate)
Shoes:   ████ #E74C3C (bright red)

The Description Anchor Technique

The most practical consistency technique: write a 50+ word detailed description and reuse it exactly in every prompt.

Template

[age] [gender] with [hair: color, length, style], [eye color] eyes,
[skin tone], [facial features: any distinctive marks],
wearing [top: specific color and style], [bottom: specific color and style],
[shoes: specific color and style], [accessories: specific items]

Example

young woman in her mid-twenties with short asymmetric auburn red hair
swept to the right side, bright emerald green eyes, light warm skin
with a small beauty mark below her left eye, wearing a fitted navy
blue bomber jacket with silver zipper over a white crew-neck t-shirt,
dark slate slim jeans, and bright red canvas sneakers, small silver
stud earrings

Use this exact block in EVERY prompt for this character, only changing the action/pose/scene.

Proportion Guide

StyleHead-to-Body RatioBest For
Realistic7.5 : 1Film, photorealistic
Heroic8 : 1Superheroes, action
Anime/Manga5-6 : 1Japanese animation style
Stylized4-5 : 1Western animation
Chibi/Super-deformed2-3 : 1Cute, comedic, mascots

Include proportion style in your prompts: "realistic proportions" vs "anime style proportions" vs "chibi proportions"

Using LoRA for Consistency

For projects requiring many images of the same character, train a LoRA:

# Use FLUX with a character LoRA
belt app run falai/flux-dev-lora --input '{
  "prompt": "maya_chen character, sitting at a cafe reading a book, warm afternoon light, candid photography style",
  "loras": [{"path": "path/to/maya-chen-lora.safetensors", "scale": 0.8}]
}'

LoRA Training Tips:

  • Need 10-20 reference images of the character (consistent style)
  • Train on specific trigger word (e.g., "maya_chen")
  • Scale 0.7-0.9 balances consistency with prompt flexibility
  • Lower scale = more creative freedom, higher = more strict matching

Common Consistency Failures

IssueWhy It HappensMitigation
Hair color driftModel interprets "red hair" differently each timeUse specific shade: "auburn red #C0392B"
Eye color changeLow priority in generationMention eye color early in prompt
Outfit inconsistencyModel fills in details creativelyDescribe every clothing item explicitly
Age shiftVague age descriptionUse "mid-twenties" not "young"
Face structure changeDifferent generations = different facesUse LoRA or same seed base
Proportion shiftStyle interpretation variesSpecify "7.5 head proportions"

Character Bible Template

For ongoing projects, maintain a character bible document:

# Character: Maya Chen

## Visual Description (use in all prompts)
young woman in her mid-twenties with short asymmetric auburn red hair...
[full 50+ word anchor description]

## Color Palette
- Skin: #F5D0A9
- Hair: #C0392B
- Eyes: #27AE60
- Primary outfit: Navy #2C3E50
- Accent: Red #E74C3C

## Personality Notes (for expression/pose choices)
- Confident but approachable
- Default expression: slight curious smile
- Gestures: talks with hands, leans forward when interested

## Style Keywords
concept art, clean lines, sharp details, [art style reference]

## LoRA (if trained)
Path: ./loras/maya-chen-v2.safetensors
Trigger: maya_chen
Recommended scale: 0.8

Common Mistakes

MistakeProblemFix
Vague descriptionsDifferent character every time50+ word detailed anchor
Inconsistent prompt structureVarying emphasis = varying resultsSame structure, only change action/scene
Generating one view onlyCan't use character in different contextsCreate full turnaround reference
No color documentationColors drift across generationsRecord exact hex codes
Skipping expression sheetCharacter feels one-dimensionalGenerate 6+ expressions
Not using LoRA for big projectsInconsistency compoundsTrain LoRA for 10+ image projects

Related Skills

npx skills add inference-sh/skills@ai-image-generation
npx skills add inference-sh/skills@flux-image
npx skills add inference-sh/skills@prompt-engineering

Browse all apps: belt app list

来自 halt-catch-fire 的更多技能

agent-browser
halt-catch-fire
通过inference.sh为AI代理提供浏览器自动化功能。可导航网页、使用@e引用与元素交互、截图、录制视频。能力包括:网页抓取、表单填写、点击、输入、拖放、文件上传、JavaScript执行。适用于:网页自动化、数据提取、测试、代理浏览、研究。触发词:浏览器、网页自动化、抓取、导航、点击、填写表单、截图、浏览网页、playwright、无头浏览器、网页代理、上网、录制视频
browser-automationweb-scrapingtesting
agent-tools
halt-catch-fire
We need to translate the given text from English to Simplified Chinese. The instruction says to preserve product names, protocol names, URLs, numbers, and technical terms. Also, do not include the name "agent-tools" unless it appears in the source text. The source text does not contain "agent-tools" explicitly, so we don't include it. We just translate the text inside <text>. The text is a description of an agent skill. We need to output only the translation, no extra commentary. Let's translate: "Run 250+ AI apps via inference.sh CLI - image generation, video creation, LLMs, search, 3D, Twitter automation. Models: FLUX, Veo, Gemini, Grok, Claude, Seedance, OmniHuman, Tavily, Exa, OpenRouter, and many more. Use when running AI apps, generating images/videos, calling LLMs, web search, or automating Twitter. Triggers: inference.sh, infsh, ai model, run ai, serverless ai, ai api, flux, veo
developmentapicreative
ai-avatar-video
halt-catch-fire
通过inference.sh CLI创建AI虚拟形象和说话头像视频。推荐:P-Video-Avatar(最快、最便宜、内置TTS)。其他选项:OmniHuman、Fabric、PixVerse。音频:Inworld TTS-2(支持100+语言、角色情感控制)、ElevenLabs、Kokoro。功能:音频驱动虚拟形象、文本转虚拟形象、唇形同步视频、说话头像生成、虚拟主持人、UGC内容。用途:AI主持人、解说视频、虚拟网红、配音、营销视频、UGC广告、游戏虚拟形象……
videocreativemedia
ai-image-generation
halt-catch-fire
通过 inference.sh CLI 使用 GPT-Image-2、FLUX、Gemini、Grok、Seedream、Reve 及 50 多个模型生成 AI 图像。模型包括:GPT-Image-2、FLUX Dev LoRA、FLUX.2 Klein LoRA、Gemini 3 Pro Image、Grok Imagine、Seedream 4.5、Reve、ImagineArt。功能:文本转图像、图像转图像、图像修复、LoRA、图像编辑、图像放大、文本渲染。用途:AI 艺术、产品模型、概念艺术、社交媒体图形、营销视觉、插图。触发词:flux、图像生成、AI 图像、文本转...
creativemediaimage
ai-video-generation
halt-catch-fire
通过 inference.sh CLI,使用 Google Veo、Seedance 2.0、HappyHorse、Wan、Grok 及 40 多个模型生成 AI 视频。模型包括:Veo 3.1、Veo 3、Seedance 2.0、HappyHorse 1.0、Wan 2.5、Grok Imagine Video、OmniHuman、Fabric、HunyuanVideo。功能涵盖:文生视频、图生视频、参考视频生成、视频编辑、唇形同步、虚拟人动画、视频增强、拟音音效。适用于:社交媒体视频、营销内容、解说视频、产品演示、AI 虚拟人。触发词:视频生成、AI 视频……
creativevideomedia
app-store-screenshots
halt-catch-fire
根据精确的平台规格创建App Store和Google Play截图,涵盖iOS/Android尺寸、图库排序、设备样机和预览视频。用途:应用商店优化、ASO、应用截图、应用预览、Play Store列表。触发词:应用商店截图、ASO、应用商店优化、Play Store截图、应用预览、应用列表、iOS截图、Android截图、应用商店图片、应用样机、设备样机、应用图库、商店列表
designmarketingcreative
competitor-teardown
halt-catch-fire
结构化竞争分析,包含功能矩阵、SWOT分析、定位图谱及用户体验评估。涵盖研究框架、价格对比、评论挖掘与可视化交付物。适用于:市场调研、竞争情报、投资者演示、产品策略、销售赋能。触发词:竞争对手分析、竞争分析、竞品拆解、市场调研、竞争情报、SWOT分析、竞品对比、市场格局、竞品评估、竞争性...
researchdata-analysismarketing
image-to-video
halt-catch-fire
We need to translate the given text from English to Simplified Chinese. The text is a description of an agent skill for image-to-video conversion. We must preserve product names, protocol names, URLs, numbers, and technical terms. The name "image-to-video" is to be preserved but not included unless it appears in the source text. The source text does not contain "image-to-video" as a standalone name; it appears as part of triggers like "image to video". So we should not add it. We translate the entire text inside <text> tags. The text includes: "Still-to-video conversion guide: model selection, motion prompting, and camera movement. Covers Wan 2.5 i2v, Seedance, Fabric, Grok Video with when to use each. Use for: animating images, creating video from stills, adding motion, product animations. Triggers: image to video, i2v, animate image, still to video, add motion to image, image animation, photo to video, animate still, wan i2v,