
Gpt Image 2
- 60 installs
- 10.1k repo stars
- Updated July 12, 2026
- conardli/web-design-skill
Generates and edits images with GPT Image 2 (or OpenAI-compatible endpoints), adapting between local generation, host-tool delegation, and prompt-advisor modes.
About
A prompt-engineering and image-generation pack for GPT Image 2 that runs text-to-image and image-edit tasks across three runtime modes depending on whether an API key or host image tool is available. A developer uses it to produce posters, UI mockups, product visuals, infographics, and other images from structured prompt templates.
- 18 categories and 79 prompt templates for posters, UI, diagrams, and more
- Mode detection picks local, host-native, or advisor behavior automatically
Gpt Image 2 by the numbers
- 60 all-time installs (skills.sh)
- Ranked #865 of 1,335 Generative Media skills by installs in the Skillselion catalog
- Data as of Aug 5, 2026 (Skillselion catalog sync)
npx skills add https://github.com/conardli/web-design-skill --skill gpt-image-2Add your badge
Show developers this skill is listed on Skillselion. Paste this into your README.
| Installs | 60 |
|---|---|
| repo stars | ★ 10.1k |
| Last updated | July 12, 2026 |
| Repository | conardli/web-design-skill ↗ |
What it does
Generates and edits images with GPT Image 2 (or OpenAI-compatible endpoints), adapting between local generation, host-tool delegation, and prompt-advisor modes.
Files
{
"name": "gpt-image-2",
"version": "1.0.4",
"category": "Image Generation / Prompt Engineering",
"description": "Focused image generation and editing skill for GPT Image 2 and OpenAI-compatible image APIs. Supports three runtime modes — Garden local, host-native delegation, and advisor-only — with 18 visual categories and 80+ structured prompt templates.",
"homepage": "https://github.com/ConardLi/garden-skills/tree/main/skills/gpt-image-2",
"compat": [
"claude-code",
"claude-ai",
"cursor",
"codex-cli",
"gemini-cli",
"opencode"
]
}
GPT Image 2 Skill
A focused image-generation / editing skill for GPT Image 2, with a single SKILL definition that adapts to three runtime modes — local generation, host-native delegation, and pure prompt advisor.
中文文档 · Back to collection root

---
What it does
This skill is a structured prompt-engineering and image-generation pack built around the GPT Image 2 model (and OpenAI-compatible image endpoints). It only does two image tasks — POST /images/generations and POST /images/edits — but it does them in three different runtime environments without changing user-facing behavior.
It bundles:
- A mode-aware workflow so the same skill works whether the agent itself owns the image API key, the host has its own image tool, or there is no image tool at all.
- A structured template library of 18 categories and 79 prompt templates covering posters, UI mockups, product visuals, infographics, academic figures, technical diagrams, comics, avatars, and editing workflows.
- Reproducible prompt + image archival under
garden-gpt-image-2/prompt/andgarden-gpt-image-2/image/with task-slug + timestamp naming.
---
The three runtime modes
The very first thing this skill does on any task is run a tiny detection script:
node skills/gpt-image-2/scripts/check-mode.js
# or for structured output:
node skills/gpt-image-2/scripts/check-mode.js --jsonThe output picks one of three modes:
| Mode | Trigger | Behavior |
|---|---|---|
| A — Garden local | ENABLE_GARDEN_IMAGEGEN truthy AND OPENAI_API_KEY present | End-to-end: pick template → render prompt → call generate.js / edit.js → image lands on disk |
| B — Host-native | Garden disabled, but the host agent already has an image tool (image_generation, dalle, nano_banana, image MCP, etc.) | Render the prompt, then delegate image generation to the host's own tool |
| C — Advisor | Garden disabled, host has no image tool | Skill degrades into a high-quality prompt writer — saves the rendered prompt to garden-gpt-image-2/prompt/ and instructs the user to paste it into ChatGPT / Midjourney / DALL·E / Sora / Nano Banana / their own gateway |
In all three modes, prompt files are saved (mode A & C must save, mode B is recommended for reuse). Only mode A produces an image file; mode B leaves that to the host, mode C cannot.
---
Quick start
0. Detect the mode (always step 0)
node skills/gpt-image-2/scripts/check-mode.jsThe commands below (1–4) only apply in Mode A.
1. Text-to-image
node skills/gpt-image-2/scripts/generate.js \
--prompt "A cute baby sea otter" \
--size 1024x1024 \
--quality high2. Generate from a saved prompt file
node skills/gpt-image-2/scripts/generate.js \
--promptfile garden-gpt-image-2/prompt/poster-20260424-153045.md3. Edit an existing image
node skills/gpt-image-2/scripts/edit.js \
--image assets/source.png \
--prompt "Replace the background with a clean studio scene"4. Mask-based local edit
node skills/gpt-image-2/scripts/edit.js \
--image assets/source.png \
--mask assets/mask.png \
--prompt "Replace only the masked area with a glass vase"For Mode B / C there is no CLI entry point — the skill just renders the final prompt and either hands it to the host's image tool (B) or shows it to the user (C).
---
Case Gallery
The public case library covers 18 categories, 79 templates, and 160+ generated / edited results. This gallery is a curated map of the most important capability families: each thumbnail opens the live case page, while the image itself is served from the dedicated ConardLi/gpt-image-2-101 case repository.
UI Mockups
<table> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/ui-mockups%2Flive-commerce-ui%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/ui-mockups/live-commerce-ui/1-thumb.webp" alt="Live commerce UI case" width="100%"></a><br/><strong><code>live-commerce-ui</code></strong><br/><sub>Celebrity livestream commerce interface.</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/ui-mockups%2Fsocial-interface-mockup%2F3"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/ui-mockups/social-interface-mockup/3-thumb.webp" alt="Social interface mockup case" width="100%"></a><br/><strong><code>social-interface-mockup</code></strong><br/><sub>Official product announcement in a social feed.</sub></td> </tr> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/ui-mockups%2Fproduct-card-overlay%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/ui-mockups/product-card-overlay/1-thumb.webp" alt="Product card overlay case" width="100%"></a><br/><strong><code>product-card-overlay</code></strong><br/><sub>Skincare landing-page hero with product, model, and badges.</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/ui-mockups%2Fchat-interface-scene%2F3"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/ui-mockups/chat-interface-scene/3-thumb.webp" alt="Chat interface scene case" width="100%"></a><br/><strong><code>chat-interface-scene</code></strong><br/><sub>Claude-style assistant screenshot with structured conversation.</sub></td> </tr> </table>
Product And Branding
<table> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/product-visuals%2Fexploded-view-poster%2F2"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/product-visuals/exploded-view-poster/2-thumb.webp" alt="Exploded view poster case" width="100%"></a><br/><strong><code>exploded-view-poster</code></strong><br/><sub>Vision Pro 2 optical and compute-module teardown.</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/product-visuals%2Fpremium-studio-product%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/product-visuals/premium-studio-product/1-thumb.webp" alt="Premium studio product case" width="100%"></a><br/><strong><code>premium-studio-product</code></strong><br/><sub>Luxury skincare still life for editorial product pages.</sub></td> </tr> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/branding-and-packaging%2Fcosmetic-packaging%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/branding-and-packaging/cosmetic-packaging/1-thumb.webp" alt="Cosmetic packaging case" width="100%"></a><br/><strong><code>cosmetic-packaging</code></strong><br/><sub>Premium skincare gift box with material polish.</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/branding-and-packaging%2Fbeverage-label-design%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/branding-and-packaging/beverage-label-design/1-thumb.webp" alt="Beverage label design case" width="100%"></a><br/><strong><code>beverage-label-design</code></strong><br/><sub>Guochao sparkling-water bottle label and commercial scene.</sub></td> </tr> </table>
Editing Workflows
<table> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/editing-workflows%2Fbackground-replacement%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/editing-workflows/background-replacement/1-thumb.webp" alt="Background replacement case" width="100%"></a><br/><strong><code>background-replacement</code></strong><br/><sub>Portrait moved into Times Square night ambience.</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/editing-workflows%2Fobject-removal%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/editing-workflows/object-removal/1-thumb.webp" alt="Object removal case" width="100%"></a><br/><strong><code>object-removal</code></strong><br/><sub>Remove unwanted people from a graduation group photo.</sub></td> </tr> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/editing-workflows%2Fproduct-retouching%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/editing-workflows/product-retouching/1-thumb.webp" alt="Product retouching case" width="100%"></a><br/><strong><code>product-retouching</code></strong><br/><sub>Commerce-grade AirPods product cleanup.</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/editing-workflows%2Fportrait-local-edit%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/editing-workflows/portrait-local-edit/1-thumb.webp" alt="Portrait local edit case" width="100%"></a><br/><strong><code>portrait-local-edit</code></strong><br/><sub>Hair color and style edit while preserving identity.</sub></td> </tr> </table>
Infographics And Visual Docs
<table> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/infographics%2Fbento-grid-infographic%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/infographics/bento-grid-infographic/1-thumb.webp" alt="Bento grid infographic case" width="100%"></a><br/><strong><code>bento-grid-infographic</code></strong><br/><sub>iPhone 16 Pro feature breakdown in a compact grid.</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/infographics%2Fcomparison-infographic%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/infographics/comparison-infographic/1-thumb.webp" alt="Comparison infographic case" width="100%"></a><br/><strong><code>comparison-infographic</code></strong><br/><sub>Phone comparison designed for decision support.</sub></td> </tr> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/slides-and-visual-docs%2Fdense-explainer-slides%2F2"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/slides-and-visual-docs/dense-explainer-slides/2-thumb.webp" alt="Dense explainer slide case" width="100%"></a><br/><strong><code>dense-explainer-slides</code></strong><br/><sub>One-page AI Agent mechanism explainer.</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/slides-and-visual-docs%2Fvisual-report-page%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/slides-and-visual-docs/visual-report-page/1-thumb.webp" alt="Visual report page case" width="100%"></a><br/><strong><code>visual-report-page</code></strong><br/><sub>Business summary page with KPI cards and chart rhythm.</sub></td> </tr> </table>
Academic And Technical
<table> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/academic-figures%2Fmethod-pipeline-overview%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/academic-figures/method-pipeline-overview/1-thumb.webp" alt="Method pipeline overview case" width="100%"></a><br/><strong><code>method-pipeline-overview</code></strong><br/><sub>RAG-based long-context QA pipeline for papers.</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/academic-figures%2Fneural-network-architecture%2F2"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/academic-figures/neural-network-architecture/2-thumb.webp" alt="Neural network architecture case" width="100%"></a><br/><strong><code>neural-network-architecture</code></strong><br/><sub>ViT-B/16 architecture figure with tensor flow.</sub></td> </tr> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/technical-diagrams%2Fsystem-architecture%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/technical-diagrams/system-architecture/1-thumb.webp" alt="System architecture case" width="100%"></a><br/><strong><code>system-architecture</code></strong><br/><sub>Multi-tenant AI SaaS production architecture.</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/technical-diagrams%2Fsequence-diagram%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/technical-diagrams/sequence-diagram/1-thumb.webp" alt="Sequence diagram case" width="100%"></a><br/><strong><code>sequence-diagram</code></strong><br/><sub>OAuth 2.0 authorization code + PKCE sequence.</sub></td> </tr> </table>
Story, Maps And Characters
<table> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/storyboards-and-sequences%2Fanime-key-visual%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/storyboards-and-sequences/anime-key-visual/1-thumb.webp" alt="Anime key visual case" width="100%"></a><br/><strong><code>anime-key-visual</code></strong><br/><sub>Fantasy game launch key visual with crop-safe layout.</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/maps%2Ffood-map%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/maps/food-map/1-thumb.webp" alt="Food map case" width="100%"></a><br/><strong><code>food-map</code></strong><br/><sub>Shanghai city-walk food map with illustrated landmarks.</sub></td> </tr> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/maps%2Ftravel-route-map%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/maps/travel-route-map/1-thumb.webp" alt="Travel route map case" width="100%"></a><br/><strong><code>travel-route-map</code></strong><br/><sub>Kyoto three-day route map with illustrated stops.</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/portraits-and-characters%2Fprofessional-portrait%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/portraits-and-characters/professional-portrait/1-thumb.webp" alt="Professional portrait case" width="100%"></a><br/><strong><code>professional-portrait</code></strong><br/><sub>Restrained executive portrait for company and media pages.</sub></td> </tr> </table>
<sub>Full library: <a href="https://gpt-image2.mmh1.top/#/case"><b>live case browser</b></a> · <a href="https://github.com/ConardLi/gpt-image-2-101/tree/main/public/case">case source repository</a> · local index at <code>website/gpt-image2-website/public/case/INDEX.md</code>.</sub>
---
Skill structure
skills/gpt-image-2/
├── SKILL.md Main skill definition
├── scripts/
│ ├── check-mode.js Mode A/B/C detector (run this first)
│ ├── generate.js Text-to-image (Mode A only)
│ ├── edit.js Image edit / inpaint (Mode A only)
│ ├── shared.js Shared request, save, env-resolution logic
│ └── package.json
└── references/
├── prompt-writing.md Methodology: how to design templates & ask for missing fields
├── ui-mockups/ Live commerce, social, product card, chat, video cover
├── product-visuals/ Exploded view, white-bg, premium studio, packaging, lifestyle
├── infographics/ Information graphics
├── poster-and-campaigns/ Brand poster, campaign KV, banner, editorial cover
├── slides-and-visual-docs/ Dense explainer, policy slide, visual report, educational
├── portraits-and-characters/ Pro portrait, founder portrait, virtual host, character sheet
├── scenes-and-illustrations/ Healing, concept, picture book, minimalist mood
├── editing-workflows/ Background replace, local replace, removal, retouch, portrait
├── avatars-and-profile/ Style transfer, character grid, 3D icon, sticker, cultural series
├── storyboards-and-sequences/ 4-panel, manga spread, anime KV, character relations, recipe
├── grids-and-collages/ 2×2 banner grid, lookbook, mixed-style, anime pitch board
├── branding-and-packaging/ Identity board, mascot kit, cosmetic, beverage label
├── typography-and-text-layout/ Title-safe poster, bilingual layout
├── assets-and-props/ Skeuomorphic icons, game screenshot mockup
├── academic-figures/ Method pipeline, NN architecture, qualitative comparison
├── technical-diagrams/ Architecture, flow, sequence diagrams
└── maps/ Food map, travel route, illustrated city, store distribution---
Environment variables
Read in this order: CLI args → process.env → <cwd>/.env → <cwd>/.gateway.env → ~/.gateway.env.
| Variable | Required | Purpose |
|---|---|---|
ENABLE_GARDEN_IMAGEGEN | Mode A | Master switch for Mode A (1 / true / yes / on) |
OPENAI_API_KEY | Mode A | Required for actual image API calls |
OPENAI_BASE_URL | optional | Default https://api.openai.com/v1; can point to any OpenAI-compatible gateway |
OPENAI_IMAGE_MODEL | optional | Default gpt-image-2; can be swapped for gpt-image-1 / dall-e-3 / etc. |
The skill is wire-compatible with the OpenAI image API and is not hard-coded to any third-party gateway.
---
Output convention
Unless the user specifies otherwise:
| What | Where | Used in |
|---|---|---|
| Rendered prompts | garden-gpt-image-2/prompt/<task-slug>-<timestamp>.md | A / B / C |
| Generated images | garden-gpt-image-2/image/<task-slug>-<timestamp>.png | A only (B = host decides, C = none) |
<task-slug> is auto-derived from the user's request; <timestamp> is YYYYMMDD-HHMMSS.
Examples:
garden-gpt-image-2/prompt/live-commerce-ui-20260424-153045.mdgarden-gpt-image-2/image/vr-headset-exploded-view-20260424-153102.png
---
Design principles
1. Mode-aware first. The same skill never silently fails because the host doesn't have an API key — it degrades cleanly into B or C and tells the user what happened. 2. Templates over freeform prompts. 18 categories of pre-validated structured templates with explicit {argument ...} slots and default markers — much higher quality than asking "describe what you want." 3. Ask precisely, not vaguely. When a template field is missing, the skill asks per field (e.g. "Who is the host? real photo, named celebrity, free description, or random?") instead of "what style do you want?" 4. Always archive prompts. Even in advisor mode, the rendered prompt is saved so the work is reusable. 5. OpenAI-compatible by default. No vendor lock-in to any specific gateway.
---
License
MIT
GPT Image 2 Skill
面向 GPT Image 2 的聚焦型图像生成 / 编辑技能。一份 SKILL 定义,自动适配三种运行环境——本地直接出图、宿主原生图像工具、纯提示词顾问。
English · 返回集合首页

---
这个 Skill 干什么
围绕 GPT Image 2(以及任何 OpenAI 兼容的图像接口)做的结构化提示词工程 + 图像生成包。只做两件事——POST /images/generations 和 POST /images/edits,但能在三种完全不同的运行环境下做到对用户无感。
它内置了:
- 模式感知工作流:无论 Agent 自己持有 API key、宿主带原生图像工具、还是完全没有图像工具,同一份 Skill 都能用。
- 结构化模板库:18 大类、79 个提示词模板,覆盖海报、UI 样机、产品图、信息图、学术图、技术架构图、漫画、头像、编辑工作流。
- 可复用的 prompt + 图片归档:默认落盘到
garden-gpt-image-2/prompt/和garden-gpt-image-2/image/,按<task-slug>-<timestamp>命名。
---
三种运行模式
任何任务的第一步都是跑这个探测脚本:
node skills/gpt-image-2/scripts/check-mode.js
# 想拿结构化结果:
node skills/gpt-image-2/scripts/check-mode.js --json输出会判定为以下三种之一:
| 模式 | 触发条件 | 行为 |
|---|---|---|
| A · Garden 本地生图 | ENABLE_GARDEN_IMAGEGEN 为真 且 有 OPENAI_API_KEY | 端到端:选模板 → 渲染 prompt → 调用 generate.js / edit.js → 图片落盘 |
| B · Host-Native 委托宿主出图 | 未启用 Garden,但宿主 Agent 自带图像工具(image_generation / dalle / nano_banana / 图像 MCP 等) | 渲染好 prompt 后交给宿主自带的图像工具出图 |
| C · Advisor 纯提示词顾问 | 未启用 Garden,宿主也没有图像工具 | 退化成"高质量 prompt 撰写顾问"——把 prompt 落盘到 garden-gpt-image-2/prompt/,告诉用户去 ChatGPT / Midjourney / DALL·E / Sora / Nano Banana / 自己的网关里执行 |
三种模式都建议落盘 prompt 文件(A、C 必须,B 推荐),但只有 A 会产出图片文件——B 由宿主决定,C 不可能。
---
快速上手
0. 检测运行模式(永远是第一步)
node skills/gpt-image-2/scripts/check-mode.js下面 1~4 仅在 Mode A 下使用。
1. 文本生图
node skills/gpt-image-2/scripts/generate.js \
--prompt "A cute baby sea otter" \
--size 1024x1024 \
--quality high2. 用提示词文件生图
node skills/gpt-image-2/scripts/generate.js \
--promptfile garden-gpt-image-2/prompt/poster-20260424-153045.md3. 编辑已有图片
node skills/gpt-image-2/scripts/edit.js \
--image assets/source.png \
--prompt "Replace the background with a clean studio scene"4. 带遮罩的局部编辑
node skills/gpt-image-2/scripts/edit.js \
--image assets/source.png \
--mask assets/mask.png \
--prompt "Replace only the masked area with a glass vase"Mode B / C 没有 CLI 入口——Skill 只负责把最终 prompt 渲染好,然后交给宿主图像工具(B)或直接呈现给用户(C)。
---
案例画廊
公开案例库目前覆盖 18 大类、79 个模板、160+ 个生成 / 编辑结果。这里不是完整索引,而是挑出最能代表能力边界的关键案例:每张缩略图都会跳到线上案例页,图片本身来自独立的 ConardLi/gpt-image-2-101 案例仓库。
UI 样机
<table> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/ui-mockups%2Flive-commerce-ui%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/ui-mockups/live-commerce-ui/1-thumb.webp" alt="直播带货 UI 案例" width="100%"></a><br/><strong><code>live-commerce-ui</code></strong><br/><sub>明星直播带货界面,含商品、弹幕、礼物和状态层。</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/ui-mockups%2Fsocial-interface-mockup%2F3"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/ui-mockups/social-interface-mockup/3-thumb.webp" alt="社交界面样机案例" width="100%"></a><br/><strong><code>social-interface-mockup</code></strong><br/><sub>科技品牌官方账号发布产品更新公告。</sub></td> </tr> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/ui-mockups%2Fproduct-card-overlay%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/ui-mockups/product-card-overlay/1-thumb.webp" alt="产品落地页叠层案例" width="100%"></a><br/><strong><code>product-card-overlay</code></strong><br/><sub>护肤落地页 hero,包含模特、产品和卖点徽章。</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/ui-mockups%2Fchat-interface-scene%2F3"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/ui-mockups/chat-interface-scene/3-thumb.webp" alt="聊天界面案例" width="100%"></a><br/><strong><code>chat-interface-scene</code></strong><br/><sub>Claude 风格 AI 助手截图,强调对话层级和结构化回答。</sub></td> </tr> </table>
产品与品牌
<table> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/product-visuals%2Fexploded-view-poster%2F2"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/product-visuals/exploded-view-poster/2-thumb.webp" alt="产品爆炸图案例" width="100%"></a><br/><strong><code>exploded-view-poster</code></strong><br/><sub>Vision Pro 2 光机与算力模块拆解主视觉。</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/product-visuals%2Fpremium-studio-product%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/product-visuals/premium-studio-product/1-thumb.webp" alt="高端影棚产品图案例" width="100%"></a><br/><strong><code>premium-studio-product</code></strong><br/><sub>高端护肤静物,适合官网 hero 和杂志跨页。</sub></td> </tr> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/branding-and-packaging%2Fcosmetic-packaging%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/branding-and-packaging/cosmetic-packaging/1-thumb.webp" alt="化妆品包装案例" width="100%"></a><br/><strong><code>cosmetic-packaging</code></strong><br/><sub>国货高端护肤礼盒,兼顾材质和品牌感。</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/branding-and-packaging%2Fbeverage-label-design%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/branding-and-packaging/beverage-label-design/1-thumb.webp" alt="饮料标签设计案例" width="100%"></a><br/><strong><code>beverage-label-design</code></strong><br/><sub>国潮气泡水酒标 / 瓶标与商拍场景。</sub></td> </tr> </table>
图像编辑工作流
<table> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/editing-workflows%2Fbackground-replacement%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/editing-workflows/background-replacement/1-thumb.webp" alt="背景替换案例" width="100%"></a><br/><strong><code>background-replacement</code></strong><br/><sub>把日间人像替换到时代广场夜景并重新布光。</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/editing-workflows%2Fobject-removal%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/editing-workflows/object-removal/1-thumb.webp" alt="杂物去除案例" width="100%"></a><br/><strong><code>object-removal</code></strong><br/><sub>毕业合影去除边缘误入人物并修补背景。</sub></td> </tr> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/editing-workflows%2Fproduct-retouching%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/editing-workflows/product-retouching/1-thumb.webp" alt="产品精修案例" width="100%"></a><br/><strong><code>product-retouching</code></strong><br/><sub>AirPods 电商主图质感、边缘与标签锐化。</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/editing-workflows%2Fportrait-local-edit%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/editing-workflows/portrait-local-edit/1-thumb.webp" alt="人像局部编辑案例" width="100%"></a><br/><strong><code>portrait-local-edit</code></strong><br/><sub>在保留身份的前提下调整发色与发型。</sub></td> </tr> </table>
信息图与视觉文档
<table> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/infographics%2Fbento-grid-infographic%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/infographics/bento-grid-infographic/1-thumb.webp" alt="便当格信息图案例" width="100%"></a><br/><strong><code>bento-grid-infographic</code></strong><br/><sub>iPhone 16 Pro 功能拆解,以便当格组织高密度信息。</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/infographics%2Fcomparison-infographic%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/infographics/comparison-infographic/1-thumb.webp" alt="对比信息图案例" width="100%"></a><br/><strong><code>comparison-infographic</code></strong><br/><sub>手机选购对比图,围绕决策维度组织信息。</sub></td> </tr> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/slides-and-visual-docs%2Fdense-explainer-slides%2F2"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/slides-and-visual-docs/dense-explainer-slides/2-thumb.webp" alt="高密度讲解单页案例" width="100%"></a><br/><strong><code>dense-explainer-slides</code></strong><br/><sub>AI Agent 工作机制一页讲清,适合技术培训。</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/slides-and-visual-docs%2Fvisual-report-page%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/slides-and-visual-docs/visual-report-page/1-thumb.webp" alt="视觉报告页案例" width="100%"></a><br/><strong><code>visual-report-page</code></strong><br/><sub>商业执行摘要页,结合 KPI 卡片与趋势图节奏。</sub></td> </tr> </table>
学术与技术图
<table> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/academic-figures%2Fmethod-pipeline-overview%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/academic-figures/method-pipeline-overview/1-thumb.webp" alt="方法流程图案例" width="100%"></a><br/><strong><code>method-pipeline-overview</code></strong><br/><sub>RAG 长上下文问答方法流程,适合论文 overview。</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/academic-figures%2Fneural-network-architecture%2F2"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/academic-figures/neural-network-architecture/2-thumb.webp" alt="神经网络架构图案例" width="100%"></a><br/><strong><code>neural-network-architecture</code></strong><br/><sub>ViT-B/16 架构图,包含 Patch Embedding 与张量流向。</sub></td> </tr> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/technical-diagrams%2Fsystem-architecture%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/technical-diagrams/system-architecture/1-thumb.webp" alt="系统架构图案例" width="100%"></a><br/><strong><code>system-architecture</code></strong><br/><sub>多租户 AI 客服 SaaS 生产架构总览。</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/technical-diagrams%2Fsequence-diagram%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/technical-diagrams/sequence-diagram/1-thumb.webp" alt="时序图案例" width="100%"></a><br/><strong><code>sequence-diagram</code></strong><br/><sub>OAuth 2.0 授权码 + PKCE 标准时序。</sub></td> </tr> </table>
故事、地图与角色
<table> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/storyboards-and-sequences%2Fanime-key-visual%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/storyboards-and-sequences/anime-key-visual/1-thumb.webp" alt="动漫主视觉案例" width="100%"></a><br/><strong><code>anime-key-visual</code></strong><br/><sub>东方幻想游戏首发 KV,兼顾多比例裁切。</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/maps%2Ffood-map%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/maps/food-map/1-thumb.webp" alt="美食地图案例" width="100%"></a><br/><strong><code>food-map</code></strong><br/><sub>上海武康路 City Walk 美食地图,带插画地标。</sub></td> </tr> <tr> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/maps%2Ftravel-route-map%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/maps/travel-route-map/1-thumb.webp" alt="旅行路线图案例" width="100%"></a><br/><strong><code>travel-route-map</code></strong><br/><sub>京都三日慢走路线图,带站点插画与路线节奏。</sub></td> <td width="50%" align="center"><a href="https://gpt-image2.mmh1.top/#/case/portraits-and-characters%2Fprofessional-portrait%2F1"><img src="https://cdn.jsdelivr.net/gh/ConardLi/gpt-image-2-101@main/public/case/portraits-and-characters/professional-portrait/1-thumb.webp" alt="职业肖像案例" width="100%"></a><br/><strong><code>professional-portrait</code></strong><br/><sub>克制的企业领袖肖像,适合官网 About 与媒体页。</sub></td> </tr> </table>
<sub>完整案例库:<a href="https://gpt-image2.mmh1.top/#/case"><b>线上案例浏览器</b></a> · <a href="https://github.com/ConardLi/gpt-image-2-101/tree/main/public/case">案例资源仓库</a> · 本地索引 <code>website/gpt-image2-website/public/case/INDEX.md</code>。</sub>
---
Skill 结构
skills/gpt-image-2/
├── SKILL.md 主技能定义
├── scripts/
│ ├── check-mode.js 模式 A/B/C 探测器(先跑这个)
│ ├── generate.js 文本生图(仅 Mode A)
│ ├── edit.js 图像编辑 / 局部编辑(仅 Mode A)
│ ├── shared.js 共享请求 / 落盘 / 环境变量解析
│ └── package.json
└── references/
├── prompt-writing.md 方法论:模板怎么设计、缺字段怎么问
├── ui-mockups/ 直播带货、社交、产品卡、聊天、短视频封面
├── product-visuals/ 爆炸图、纯白底、影棚、包装、生活方式
├── infographics/ 信息图
├── poster-and-campaigns/ 品牌主海报、Campaign KV、banner、杂志封面
├── slides-and-visual-docs/ 高密度讲解、政策风、商业报告、教学示意
├── portraits-and-characters/ 职业肖像、创始人肖像、虚拟主播、角色设定
├── scenes-and-illustrations/ 治愈系、概念大场景、绘本、极简留白
├── editing-workflows/ 背景替换、局部替换、去除、产品精修、人像编辑
├── avatars-and-profile/ 风格化自拍、角色网格、3D 图标、贴纸、文化系列
├── storyboards-and-sequences/ 4 格漫画、漫画分镜、动漫 KV、角色关系图、流程图
├── grids-and-collages/ 2×2 banner、lookbook、混风格拼贴、动漫 pitch board
├── branding-and-packaging/ 品牌识别系统、吉祥物、化妆品包装、饮料标签
├── typography-and-text-layout/ 大字海报、双语版式
├── assets-and-props/ 拟物图标、游戏截图样机
├── academic-figures/ 方法 pipeline、神经网络架构、定性对比
├── technical-diagrams/ 架构图、流程图、时序图
└── maps/ 美食地图、旅行路线图、城市插画、门店分布---
环境变量
按以下顺序读取:CLI 参数 → process.env → <cwd>/.env → <cwd>/.gateway.env → ~/.gateway.env。
| 变量 | 必需性 | 说明 |
|---|---|---|
ENABLE_GARDEN_IMAGEGEN | Mode A 必需 | 模式开关:1 / true / yes / on 启用 Mode A |
OPENAI_API_KEY | Mode A 必需 | 真正调图像 API 用 |
OPENAI_BASE_URL | 可选 | 默认 https://api.openai.com/v1,可指向任意 OpenAI 兼容网关 |
OPENAI_IMAGE_MODEL | 可选 | 默认 gpt-image-2,也可换成 gpt-image-1 / dall-e-3 等 |
默认实现严格按 OpenAI 兼容接口工作,不绑定任何第三方网关。
---
输出约定
如果用户没有明确指定输出路径:
| 内容 | 落盘位置 | 适用模式 |
|---|---|---|
| 渲染好的 prompt | garden-gpt-image-2/prompt/<task-slug>-<timestamp>.md | A / B / C |
| 生成的图片 | garden-gpt-image-2/image/<task-slug>-<timestamp>.png | 仅 A(B 由宿主决定,C 不产出) |
<task-slug> 由用户请求自动派生,<timestamp> 是 YYYYMMDD-HHMMSS。
示例:
garden-gpt-image-2/prompt/live-commerce-ui-20260424-153045.mdgarden-gpt-image-2/image/vr-headset-exploded-view-20260424-153102.png
---
设计原则
1. 先判模式,再干活。 不会因为宿主没 API key 就静默失败,而是优雅地降级到 B / C 并明确告知用户当前状态。 2. 模板优于自由提示。 18 大类预校验过的结构化模板,带显式 {argument ...} 参数槽和 default 标记,质量远高于"你说说想要啥"。 3. 精确提问,不要笼统提问。 模板字段缺失时按字段精确问("主播是谁?真人照片 / 名人名字 / 自由描述 / 随机生成?"),不要笼统问"想要什么风格"。 4. 永远归档 prompt。 即使在顾问模式,渲染好的 prompt 也会落盘,方便复用。 5. 默认 OpenAI 兼容。 不锁定任何特定网关。
---
许可证
MIT
Graphical Abstract / 图形摘要模板
本文件用于生成「期刊投稿 Graphical Abstract / 论文图形摘要 / 投稿封面图」:
- 期刊投稿要求附带的 Graphical Abstract
- 论文一图概览("一图讲清主贡献")
- 答辩首页 / 组会汇报首页里的研究亮点图
特征:
- 极简、紧凑、4 部分核心叙事(问题 → 方法 → 关键过程 → 结果)
- 横向左→右 或 中心展开布局
- 白底、低饱和工程色、≤3 主色,像高质量期刊图形摘要,绝不像营销海报
- 文字精炼到短语,禁止段落式说明
适用范围
- Elsevier / ACS / Wiley / Springer / IEEE 等期刊投稿要求的 Graphical Abstract
- arXiv / 预印本 README 顶部的"研究一图"
- 论文 supplementary 或 highlight figure
- 答辩 / 汇报"研究亮点"页
何时使用
- 用户提到「graphical abstract / 图形摘要 / 投稿摘要图 / 一图讲清 / highlight figure」
- 用户希望视觉「期刊封面级摘要图,简洁克制学术风」
- 用户已能用 1-2 句话讲清"这篇论文做了什么、得到了什么"
不要使用:
- 用户要的是「方法 pipeline 总览」 → 用
academic-figures/method-pipeline-overview.md - 用户要的是「开题 / 答辩首页总览图」 → 用
academic-figures/research-overview-poster.md - 用户要的是「机制 / 机理图」 → 用
academic-figures/mechanism-diagram.md - 用户要的是「营销 / 品牌 / 杂志封面感」 → 用
poster-and-campaigns/editorial-cover.md
缺失信息优先提问顺序
1. 研究主题(一句话;写在标题或图注里) 2. 目标期刊或目标场景(决定纵横比 + 主色调;不同期刊偏好不同) 3. 4 个核心要素:研究问题 / 方法或系统 / 关键过程或机制 / 主要结果 4. 是否有"研究对象"的简化示意(颗粒 / 分子 / 器件 / 流程) 5. 标签语言(中文 / 英文 / 双语;多数期刊要求英文) 6. 比例(默认横向 16:9 / 2:1;部分期刊要求方形 1:1,要先确认)
主模板:横向 4 段式 Graphical Abstract
📖 描述
整张图横向流动:从最左边的「研究问题 / 研究对象」开始,依次到「方法 / 系统」、「关键过程 / 机制」、「主要结果」。四个区域比例均匀,文字精炼到短语,视觉层级清晰,整体像高质量工程类期刊摘要图。
📝 提示词
{
"type": "学术期刊图形摘要(Graphical Abstract)",
"goal": "生成一张可直接用于期刊投稿的 Graphical Abstract,要求极简、白底、工程化克制配色、几秒内可读、绝无营销海报感",
"canvas": {
"aspect_ratio": "{argument name=\"aspect_ratio\" default=\"2:1\"}",
"background": "pure white #FFFFFF",
"outer_padding": "60px around the diagram",
"render_quality": "vector-clean look, anti-aliased edges, sharp text, suitable for grayscale print"
},
"title_block": {
"enabled": "{argument name=\"title_block_enabled\" default=\"false\"}",
"title": "{argument name=\"title\" default=\"\"}",
"rule": "most journals do not allow titles inside the graphical abstract; enable only when user explicitly requested a title"
},
"sections": [
{
"id": "P1",
"role": "Problem",
"label": "{argument name=\"problem_label\" default=\"Research Problem\"}",
"summary": "{argument name=\"problem_summary\" default=\"a short phrase stating the gap, e.g. 'unstable combustion under variable moisture'\"}",
"depiction": "{argument name=\"problem_depiction\" default=\"a minimal line-art sketch of the studied object or scenario\"}"
},
{
"id": "P2",
"role": "Method",
"label": "{argument name=\"method_label\" default=\"Method\"}",
"summary": "{argument name=\"method_summary\" default=\"a short phrase, e.g. 'thermogravimetric + kinetics analysis'\"}",
"depiction": "{argument name=\"method_depiction\" default=\"a minimal schematic of the analytical or experimental setup\"}"
},
{
"id": "P3",
"role": "Process",
"label": "{argument name=\"process_label\" default=\"Key Mechanism\"}",
"summary": "{argument name=\"process_summary\" default=\"a short phrase, e.g. 'two-stage volatile combustion'\"}",
"depiction": "{argument name=\"process_depiction\" default=\"a small mechanism strip with 2-3 sub-steps, line-art style\"}"
},
{
"id": "P4",
"role": "Result",
"label": "{argument name=\"result_label\" default=\"Outcome\"}",
"summary": "{argument name=\"result_summary\" default=\"a short phrase, e.g. 'optimized excess-air ratio reduces NOx by ~X%'\"}",
"depiction": "{argument name=\"result_depiction\" default=\"a minimal qualitative chart sketch (no fabricated numbers) or a result icon (gauge / bar)\"}"
}
],
"section_block_style": {
"shape": "implicit columns separated by generous whitespace, NOT four heavy rectangles in a row",
"header_text": "section label in bold sans-serif, 12-13pt, top-aligned",
"summary_text": "single phrase, 10pt regular, max 2 lines, no period",
"depiction_size": "around 35-50% of column height, vertically centered"
},
"connectors": {
"style": "thin arrows (1.2px) with simple triangle arrowheads, dark gray #334155, between adjacent sections only",
"rule": "no crossing, no curved decorative arcs; arrows convey 'leads to' / 'analyzed by' relationships",
"label_arrows": "false by default; only add label when the relationship is non-trivial"
},
"color_palette": {
"rule": "≤ 3 main colors total, drawn from a low-saturation engineering set: deep blue #1E3A8A / slate blue #3B82F6 / charcoal #1F2937; allow ONE low-saturation accent (e.g. amber #F59E0B for a heat / risk highlight) only if the user signaled a thermal or risk emphasis",
"must_print_grayscale_readable": true
},
"typography": {
"language": "{argument name=\"language\" default=\"english\"}",
"rule": "english → Inter / Helvetica / Arial; chinese → PingFang SC / Source Han Sans; bilingual → english as primary, chinese as smaller secondary line",
"consistency": "all section headers identical size; all summaries identical size; never mix serif and sans-serif"
},
"constraints": {
"must_keep": [
"all four sections visually equal-weight, no section dominates",
"white background, no gradient, no decorative pattern, no photographic background",
"language matches the target journal (default english)",
"summaries are short phrases, never full sentences with periods",
"the figure must look like it could appear on an Elsevier / ACS / IEEE table of contents page",
"every numerical claim must come from the user; if absent, render qualitatively"
],
"avoid": [
"marketing-poster aesthetics, brand campaign aesthetics, magazine cover aesthetics",
"3D effects, drop shadows, gradients, glossy fills, lens flare, motion blur",
"exaggerated flames, smoke, sparks (even when the topic is combustion)",
"cartoon mascots, emoji, decorative icons, hand-drawn wobble",
"stock-photo-style realistic backgrounds",
"fabricated numbers, percentages, equations, or chart data not provided by the user",
"saturated colors (no neon, no vivid), more than 3 main colors",
"watermarks, copyright stamps, vendor logos"
]
}
}参数策略
- 必问:4 个
*_summary(问题 / 方法 / 关键过程 / 结果)至少能给出短语 - 可默认:
aspect_ratio(2:1)、background(白色)、color_palette(深蓝/灰蓝/黑灰) - 可随机:每个 section 的
*_depiction具体造型(用户给了对象/方法名时可推断;否则反问)
自动补全策略
- 用户只给主题但没给 4 段 → 反问 4 个 summary,禁止编造研究内容
- 用户给了定性贡献但没数 → 用
qualitatively shows/consistently reduces这类无数字表达 - 用户给了数(如"NOx 降低 18%")→ 直接写
~18%,不要伪造其他指标 - 用户说"中文期刊 / 中文摘要图" → 切换中文 + 字体 PingFang / 思源黑
变体 1:中心展开式(Hub-and-spoke)
{
"type": "中心展开式 Graphical Abstract",
"modify": {
"layout": "中心放置研究对象 / 核心系统的简化示意,向外辐射出 3-4 个扇区,每个扇区代表一个核心要素(问题、方法、机制、结果之一)",
"rule": "扇区在视觉上等权,使用细线条分隔;中央对象占画面 30-40%",
"use_case": "适合系统型研究、平台型研究,或难以线性叙事的多模态贡献"
}
}适用:综合性研究、系统性贡献(如新平台、新框架)。
变体 2:方形 1:1(部分期刊要求)
{
"type": "方形 Graphical Abstract",
"modify": {
"aspect_ratio": "1:1",
"layout": "2×2 网格,左上 = 问题 / 对象,右上 = 方法,左下 = 关键过程,右下 = 结果",
"rule": "四象限严格等大、对齐;象限间留出统一间距;箭头沿 Z 字型走 P1 → P2 → P3 → P4",
"use_case": "ACS / Wiley 等部分期刊要求方形 Graphical Abstract"
}
}适用:投稿要求方形比例的期刊。
变体 3:竖版(社交媒体 / 预印本卡片)
{
"type": "竖版 Graphical Abstract",
"modify": {
"aspect_ratio": "3:4",
"layout": "上 → 下 四段式:Problem → Method → Mechanism → Outcome",
"rule": "宽度紧凑,每段保留呼吸空间;适合手机端或 Twitter / LinkedIn 卡片预览",
"use_case": "用于社交媒体推广预印本、Lab 主页 highlight 卡"
}
}适用:投稿之外的科研宣传,但仍保持学术克制风格。
避免事项
- 把 Graphical Abstract 画成"全文压缩版"——塞进所有方法步骤、所有公式、所有结果
- 用任何形式的渐变 / 玻璃质感 / 光晕 / 3D → 立刻像营销图
- 中英文标签随意混用(除非显式要求双语)
- 在没有真实数据时画出带具体数值的柱图 / 折线(严格禁止虚构数据;只能定性展示)
- 用饱和 brand 色或霓虹色——期刊摘要图应保持低饱和工程色
- 把研究对象画成超现实 3D 渲染(学术风需要的是简化线稿)
- 加期刊 logo / 水印 / "submitted to ..." 等标签
机理示意图模板
本文件用于生成「学术机理示意图 / 因果链路 / 转化路径 / 演化机制图」:
- 论文正文里的机制 / 机理分析图
- 反应 / 转化 / 退化路径图
- 因果链路 / 多阶段演化图
- 答辩 PPT 的机制说明页
特征:
- 中心对象 + 多阶段转化路径 + 结果区域
- 阶段化标注(干燥 → 热解 → 燃烧 → 氧化 → 排放,或类似的因果序列)
- 白底 + 工程化低饱和配色(深蓝 / 灰蓝 / 黑灰为主,可加 ≤1 种低饱和暖色作为高温/风险强调)
- 学术克制风格,绝对不是营销插画或科普海报
适用范围
- 燃烧 / 化学反应 / 催化 / 退化 / 老化 / 腐蚀 / 衰减 等机制示意
- 生物 / 医药 / 药物作用 / 分子互作 等通路图(学术风,非科普插画)
- 材料相变 / 损伤演化 / 失效路径
- 因果链分析图 / 演化路径图
何时使用
- 用户提到「机理 / 机制 / 反应路径 / 转化 / 演化 / 因果 / 通路 / 失效路径」
- 用户希望视觉「论文里的机制图,不是科普插画也不是营销图」
- 用户已能给出阶段顺序或转化关系
不要使用:
- 用户要的是「方法 pipeline / 系统总览」 → 用
academic-figures/method-pipeline-overview.md - 用户要的是「实验装置 / 测试系统」 → 用
academic-figures/scientific-schematic.md - 用户要的是「业务流程 / 决策图」 → 用
technical-diagrams/flowchart-decision.md - 用户要的是「教学步骤、温暖插画感」 → 用
infographics/step-by-step-infographic.md
缺失信息优先提问顺序
1. 机制 / 现象总名称(写在标题或图注里) 2. 中心研究对象是什么(颗粒 / 分子 / 器件 / 组织 / 反应体系) 3. 阶段顺序(建议 3-6 个阶段;超过 6 个考虑分组) 4. 每个阶段:阶段名 + 主导过程的极简描述(短语化) 5. 是否有分支 / 平行路径 / 反馈环 6. 是否需要标注高温区 / 风险区 / 关键反应区等局部强调 7. 标签语言(中文 / 英文 / 双语;论文图通常英文) 8. 比例(默认横向 16:9;机制图也常见 4:3)
主模板:中心对象 + 多阶段转化 + 结果区
📖 描述
中心是研究对象的简化示意(颗粒 / 分子结构 / 器件 / 反应体系),周围以"阶段化转化路径"展开:从初始态经过若干中间机制阶段到达最终结果区。所有连接以学术克制风格的箭头表达,禁止戏剧化效果(无火焰、无浓烟、无炫光)。
📝 提示词
{
"type": "学术机理示意图(mechanism / pathway figure)",
"goal": "生成一张可直接放进工程类或自然科学论文正文的机制示意图,强调因果路径清晰、学术克制、可单色印刷可读",
"canvas": {
"aspect_ratio": "{argument name=\"aspect_ratio\" default=\"16:9\"}",
"background": "pure white #FFFFFF",
"outer_padding": "60px around the diagram",
"render_quality": "vector-clean look, anti-aliased edges, sharp text"
},
"title_caption": {
"figure_label": "{argument name=\"figure_label\" default=\"Figure X.\"}",
"caption": "{argument name=\"caption\" default=\"Schematic of the proposed mechanism.\"}",
"position": "bottom-center, italic serif or compact sans-serif, smaller font size"
},
"central_object": {
"label": "{argument name=\"object_label\" default=\"Biomass particle\"}",
"depiction": "{argument name=\"object_depiction\" default=\"a simplified cross-sectional sketch of a porous biomass particle, line-art style, no photo realism\"}",
"position": "horizontally centered, occupying roughly 25-35% of canvas width",
"style": "thin line-art / engineering schematic, no 3D, no shading, no hyperreal texture"
},
"stages": {
"count": "{argument name=\"stage_count\" default=\"5\"}",
"items": [
{
"id": "M1",
"name": "{argument name=\"stage_1_name\" default=\"Drying\"}",
"summary": "{argument name=\"stage_1_summary\" default=\"moisture evaporation under heating\"}",
"highlight": "{argument name=\"stage_1_highlight\" default=\"none\"}"
},
{
"id": "M2",
"name": "{argument name=\"stage_2_name\" default=\"Pyrolysis\"}",
"summary": "{argument name=\"stage_2_summary\" default=\"thermal decomposition releasing volatiles\"}",
"highlight": "{argument name=\"stage_2_highlight\" default=\"reaction zone\"}"
},
{
"id": "M3",
"name": "{argument name=\"stage_3_name\" default=\"Volatile Combustion\"}",
"summary": "{argument name=\"stage_3_summary\" default=\"gas-phase combustion of released volatiles\"}",
"highlight": "{argument name=\"stage_3_highlight\" default=\"high-temperature region\"}"
},
{
"id": "M4",
"name": "{argument name=\"stage_4_name\" default=\"Char Oxidation\"}",
"summary": "{argument name=\"stage_4_summary\" default=\"surface oxidation of the remaining char\"}",
"highlight": "{argument name=\"stage_4_highlight\" default=\"none\"}"
},
{
"id": "M5",
"name": "{argument name=\"stage_5_name\" default=\"Emission Formation\"}",
"summary": "{argument name=\"stage_5_summary\" default=\"formation of NOx, CO, particulate matter\"}",
"highlight": "{argument name=\"stage_5_highlight\" default=\"emission risk region\"}"
}
]
},
"result_region": {
"enabled": "{argument name=\"result_region_enabled\" default=\"true\"}",
"label": "{argument name=\"result_region_label\" default=\"Outcome\"}",
"items": "{argument name=\"result_region_items\" default=\"temperature distribution, combustion efficiency, emission characteristics\"}",
"position": "rightmost block or bottom-right region, visually separated from stages but stylistically consistent"
},
"stage_block_style": {
"shape": "rounded rectangle (corner radius ~6px) OR stage label + leader line directly attached to the central object",
"size_per_stage": "consistent across all stages",
"fill": "very light tint (e.g. #F1F5F9, #ECFEFF) — at most 2 different tints; use a low-saturation warm tint (e.g. #FEF3C7) only for stages whose 'highlight' is non-none",
"border": "1.2px solid dark gray #334155",
"title_text": "stage name in bold sans-serif (Helvetica / Inter / Arial / PingFang / Source Han Sans for CJK), 11-12pt",
"summary_text": "single phrase, 9-10pt regular, no full sentence, no period"
},
"connectors": {
"style": "thin arrows (1.2px) with simple triangle arrowheads, dark gray #334155",
"rule": "connect stages in causal / temporal order, no crossing, no decorative curves; only label arrows when carrying a named quantity (e.g. 'heat flux', 'O2', 'volatiles')",
"feedback_loop": {
"enabled": "{argument name=\"feedback_loop\" default=\"false\"}",
"rule": "if true, add one curved dashed arrow looping back, labeled e.g. 'self-propagating heat'"
}
},
"highlight_strategy": {
"rule": "for stages whose 'highlight' is non-none, apply ONLY a subtle low-saturation tint background (e.g. #FEF3C7 for high-temperature; #FEE2E2 for emission risk). NEVER use flames, smoke, glow, lens flare, or 3D heat-map effects",
"max_highlighted_stages": 2
},
"constraints": {
"must_keep": [
"central object visually anchors the figure; stages radiate or flow outward in a stable reading order",
"white background, no gradient, no decorative pattern",
"color palette ≤ 3 main colors, must remain readable in grayscale print",
"only sans-serif typography, no script / handwritten / display fonts",
"stage labels are short phrases, never full sentences",
"the figure must look like it came from a journal article, not a popular-science illustration",
"all arrows aligned, no crossings unless the mechanism genuinely requires it"
],
"avoid": [
"exaggerated flames, smoke, sparks, glow, lens flare, motion blur",
"3D rendering, metallic highlights, glossy fills",
"cartoon mascots, emoji, decorative icons, hand-drawn wobble",
"photo-realistic photography of equipment, products, or scenery",
"marketing poster aesthetics, magazine cover aesthetics",
"fabricated numbers, equations, or chemical formulas not provided by the user",
"saturated brand-style colors (no neon, no vivid)",
"watermarks, copyright stamps, vendor logos"
]
}
}参数策略
- 必问:
object_label/object_depiction、阶段名、阶段顺序 - 可默认:
aspect_ratio(16:9)、background(白色)、figure_label/caption、配色 tint - 可随机:每个 stage 的
summary措辞(用户给了大意可学术化润色)、highlight是否启用(无明确说明时默认 none)
自动补全策略
- 用户给出现象名 + 阶段数但没说每阶段细节 → 反问,禁止编造不存在的物理 / 化学过程
- 用户给出阶段名但没给摘要 → 用学术化短语补全(保持 ≤6 词)
- 用户没说有没有反馈环 → 默认
feedback_loop: false - 用户说"中文论文 / 答辩" → 切换标签为中文 + 字体 PingFang / 思源黑
变体 1:左 → 中 → 右 三段式因果链
{
"type": "三段式因果链机制图",
"modify": {
"layout": "左侧 = 初始条件 / 触发因素;中间 = 多阶段转化机制;右侧 = 最终结果 / 表征",
"rule": "三段之间用粗一些的分隔留白(视觉分组),但保持统一描边和字体;左右两侧文字精炼到 ≤4 项",
"use_case": "需要清晰区分'起因 → 过程 → 结果'的机制图,例如'生物质燃烧 → 多阶段反应 → 排放与残炭'"
}
}适用:燃烧 / 反应工程、退化老化、损伤演化、临床因果通路(学术风)。
变体 2:循环 / 自激发机制
{
"type": "循环自激发机制图",
"modify": {
"layout": "阶段排成环形,箭头沿环顺时针方向;中央写出循环驱动力或关键中间产物",
"annotation": "环上选 1-2 个箭头加 dashed 样式标注 'positive feedback' / 'self-propagating'",
"use_case": "正反馈机制、自催化反应、慢性退化循环"
}
}适用:自催化、链式反应、热失控、慢性炎症通路。
变体 3:多分支竞争路径
{
"type": "多分支竞争机制图",
"modify": {
"layout": "中心对象向外分出 2-3 条平行路径,每条代表一种竞争性机制;末端各自连到不同的结果区",
"annotation": "每条路径起点处标注控制条件(temperature / O2 partial pressure / pH 等)",
"use_case": "需要表达'相同前体在不同条件下走不同机制'的对比型机理图"
}
}适用:路径选择性反应、相分离、不同温度区间下的反应主导机制。
避免事项
- 用渲染感火焰 / 浓烟 / 爆炸 / 炫光来"装专业" → 立刻沦为营销插画
- 阶段块大小不一、字号混乱、字体混用衬线 + 无衬线
- 用 emoji 或卡通图标当阶段图示
- 用饱和 / 霓虹 / 渐变背景代替克制工程色
- 把不存在的化学方程、物理常数、温度数值塞进图里(严格禁止虚构数据)
- 把"机制示意图"画成完整设备剖视图(应该用
scientific-schematic.md) - 把对比 / 多工况结果(应该用
multi-condition-comparison.md)混进机制图
论文方法 Pipeline 总览图模板
本文件用于生成"论文 method 章节首页那张总览图":
- 顶会论文 method 章节首图(CVPR / NeurIPS / ICLR / ACL / SIGGRAPH 等)
- 系统总览 / pipeline figure
- 综述论文 framework 概念图
- 实验装置 / 数据流总览
- 答辩 PPT 方法概览
特征:
- 横向 3-6 个 stage 块
- 每个 stage 之间有清晰的有向数据流
- 每个 stage 有:阶段名称 + 简化插图 + 输入 / 输出小标
- 整体白底 / 浅灰底,黑色或深灰主线条
- 出版物字体(Helvetica / Inter / Arial),克制的辅助色
- 极简、几何精确、可单色印刷可读
适用范围
- 论文 method overview / framework figure
- 综述论文 pipeline 总览
- 系统总览图("我们的方法分 4 步:...")
- 数据流 / 信号流总览
- 实验流程总览
何时使用
- 用户提到 "论文 / paper / method / pipeline / framework / overview / 综述 / 顶会 / arXiv"
- 用户希望视觉「极简、白底、黑线、几何精确、像 CVPR 论文那种总览图」
- 用户已有具体的 stage 描述
不要使用:
- 用户要的是「神经网络架构图」(layer 块 + tensor shape)→ 用
academic-figures/neural-network-architecture.md - 用户要的是「概念 / 原理示意图」(自由度高的科学示意)→ 用
academic-figures/scientific-schematic.md - 用户要的是「步骤教程」(插画感、温暖)→ 用
infographics/step-by-step-infographic.md - 用户要的是「工程系统架构图」(暗色 + 半透明色块)→ 用
technical-diagrams/system-architecture.md - 用户要的是「业务流程图」 → 用
technical-diagrams/flowchart-decision.md
缺失信息优先提问顺序
1. 方法 / 系统的总名称(写在图标题或图注里) 2. 阶段数(建议 3-6 个,超过 6 个考虑分层) 3. 每个阶段的:名称 + 主操作 + 输入 + 输出 4. 数据形态(图像 / 文本 / 点云 / 音频 / 多模态)—— 决定 stage 内的简化插图 5. 是否有跳连 / 反馈环 / 多分支 6. 比例(横向 16:9 或 2:1,符合论文双栏格式) 7. 是否需要英文标签(论文图通常英文)
主模板:横向 N 阶段方法 pipeline 图
📖 描述
整张图横向流动:从最左边的输入开始,依次经过 3-6 个矩形 / 圆角矩形阶段块,每个块内有简化插图 + 阶段名 + 输入输出小标,箭头串联,最右边输出结果。整体克制、对齐严格、几何精确。
📝 提示词
{
"type": "学术论文方法 Pipeline 总览图(method overview figure)",
"goal": "生成一张可直接放进顶会论文 method 章节首页的 pipeline 总览图,要求极简、白底、几何精确、出版物级可读",
"canvas": {
"aspect_ratio": "{argument name=\"aspect_ratio\" default=\"16:9\"}",
"background": "pure white #FFFFFF or very light gray #FAFAFA",
"outer_padding": "60px around the diagram",
"render_quality": "vector-clean look, anti-aliased edges, sharp text"
},
"title_caption": {
"figure_label": "{argument name=\"figure_label\" default=\"Figure 1.\"}",
"caption": "{argument name=\"caption\" default=\"Overview of our proposed pipeline.\"}",
"position": "bottom-center, italic serif or compact sans-serif, smaller font size"
},
"input": {
"label": "{argument name=\"input_label\" default=\"Input Image\"}",
"thumbnail": "{argument name=\"input_thumbnail\" default=\"a small representative thumbnail (e.g. an RGB image, a text snippet, a point cloud)\"}",
"position": "leftmost, vertically centered"
},
"stages": {
"count": "{argument name=\"stage_count\" default=\"4\"}",
"items": [
{
"id": "S1",
"name": "{argument name=\"stage_1_name\" default=\"Feature Extractor\"}",
"icon_or_glyph": "{argument name=\"stage_1_glyph\" default=\"a stack of 3 small horizontal bars representing CNN feature maps\"}",
"sub_label": "{argument name=\"stage_1_sub\" default=\"ResNet-50\"}"
},
{
"id": "S2",
"name": "{argument name=\"stage_2_name\" default=\"Multi-scale Encoder\"}",
"icon_or_glyph": "{argument name=\"stage_2_glyph\" default=\"a small triangle / pyramid representing multi-scale\"}",
"sub_label": "{argument name=\"stage_2_sub\" default=\"FPN-style\"}"
},
{
"id": "S3",
"name": "{argument name=\"stage_3_name\" default=\"Cross-attention Decoder\"}",
"icon_or_glyph": "{argument name=\"stage_3_glyph\" default=\"two interleaved arrows representing cross-attention\"}",
"sub_label": "{argument name=\"stage_3_sub\" default=\"Transformer\"}"
},
{
"id": "S4",
"name": "{argument name=\"stage_4_name\" default=\"Prediction Head\"}",
"icon_or_glyph": "{argument name=\"stage_4_glyph\" default=\"a small grid representing dense prediction\"}",
"sub_label": "{argument name=\"stage_4_sub\" default=\"MLP × 2\"}"
}
]
},
"output": {
"label": "{argument name=\"output_label\" default=\"Predicted Mask\"}",
"thumbnail": "{argument name=\"output_thumbnail\" default=\"a small representative output (e.g. a segmentation mask, a 3D model, a generated image)\"}",
"position": "rightmost, vertically centered"
},
"stage_block_style": {
"shape": "rounded rectangle (corner radius ~6px)",
"size_per_stage": "around 120px wide × 80px tall, all stages identical size",
"fill": "very light tint (e.g. #F1F5F9, #ECFEFF, #FEF9C3) — at most 2 different tints used to group stages by category",
"border": "1.2px solid dark gray #334155",
"title_text": "stage name in bold sans-serif (Helvetica / Inter / Arial), 11-12pt, top-center inside block",
"icon_position": "centered inside block, takes ~50% of block height",
"sub_label_text": "sub-label in italic gray, below stage name"
},
"connectors": {
"style": "thin black arrows (1.2px) with simple triangle arrowheads",
"rule": "horizontal flow left → right; small label above arrow only when carrying intermediate data type (e.g. 'feature map H/4 × W/4 × 256')",
"skip_connections": {
"enabled": "{argument name=\"skip_connections\" default=\"false\"}",
"rule": "if true, draw curved arrows that arc above the main flow with dashed style, label them 'skip' / 'residual'"
}
},
"extras": {
"loss_branch": {
"enabled": "{argument name=\"loss_branch_enabled\" default=\"false\"}",
"label": "{argument name=\"loss_branch_label\" default=\"L = L_cls + λ L_reg\"}",
"rule": "if enabled, draw a small dashed branch from output back to a 'Loss' box, formula in italic"
},
"color_legend": {
"enabled": "{argument name=\"color_legend_enabled\" default=\"false\"}",
"rule": "if multiple stage tints are used, add a tiny legend bottom-right explaining each color group"
}
},
"constraints": {
"must_keep": [
"all stage blocks identical size and vertically aligned",
"white or near-white background, no gradient, no decoration",
"only sans-serif typography, no script / handwritten / display fonts",
"color palette ≤ 4 colors total, must remain readable in grayscale print",
"input thumbnail and output thumbnail same size, both have a thin border",
"arrows must not overlap stage blocks; labels must not collide with arrows",
"use English labels by default unless user requested otherwise",
"the figure should look like it came directly from a CVPR / NeurIPS PDF"
],
"avoid": [
"3D effects, drop shadows, gradients, glossy fills",
"cartoon icons, emoji, hand-drawn wobble",
"saturated colors (no neon, no vivid)",
"Helvetica + serif mixed in same diagram",
"decorative background patterns / textures",
"illustrative photo backgrounds inside stage blocks",
"stage blocks of unequal size or unaligned baselines",
"Chinese mixed with English labels unless explicitly bilingual"
]
}
}参数策略
- 必问:
stage_count、每个 stage 的名称 - 可默认:
aspect_ratio(16:9)、background(白色)、figure_label/caption、stage 块尺寸 / 颜色 - 可随机:每个 stage 内的
icon_or_glyph具体造型(用户没指定时可推断)
自动补全策略
- 用户给出方法名和"我有 4 个 stage"但没说每 stage 是什么 → 反问(不能瞎编算法细节)
- 用户给出 stage 名但没给 sub_label → 留空或自动推断(可推断时填上 "ResNet-50" 这种典型选项)
- 用户没说有没有跳连 → 默认
skip_connections: false - 用户没说有没有 loss → 默认
loss_branch: false(只在用户明确要 training pipeline 时才加) - 用户说"中文论文" / "答辩" → 切换标签为中文 + 字体 PingFang / 思源黑
变体 1:双行多分支 pipeline
{
"type": "双行多分支 pipeline 图",
"modify": {
"layout": "上下两行 stages 平行流动;中间用 fusion block 汇合",
"use_case": "多模态融合方法(如 visual + text,或 RGB + depth)",
"rule": "上行处理一种模态、下行处理另一种,最后中央汇合到 fusion block 再到输出"
}
}适用:多模态、双流网络、teacher-student 方法。
变体 2:训练 + 推理两套 pipeline 对照
{
"type": "Training vs Inference 对照 pipeline 图",
"modify": {
"layout": "上下两行:上行 'Training Phase'(含 loss、ground truth 输入、梯度回流),下行 'Inference Phase'(仅前向、轻量化)",
"annotation": "左侧用大括号标 'Training' / 'Inference'",
"use_case": "需要明确区分训练和推理流程的方法"
}
}适用:知识蒸馏、自监督预训练、半监督方法。
变体 3:迭代 / Recurrent pipeline
{
"type": "迭代式 / 循环 pipeline 图",
"modify": {
"layout": "stages 横向,但最后一个 stage 有一条曲线箭头回到第二个 stage,形成循环",
"annotation": "在循环箭头上标 'iterate × N' 或 'until convergence'",
"use_case": "迭代优化、扩散去噪、Diffusion model timestep 流"
}
}适用:扩散模型、迭代细化方法、能量模型。
变体 4:工程类技术路线图(左 / 中 / 右 三段式)
{
"type": "工程类技术路线图(engineering research roadmap)",
"modify": {
"layout": "左 / 中 / 右 三段式:左侧 = 研究对象与背景(简化线稿示意),中间 = 多步骤分析路径(4-7 个学术化模块),右侧 = 输出与结果导向(3-4 个短语化结论方向)",
"rule": "三段宽度比约 2:5:2;左右两侧用学术化短语 + 简化线稿,禁止商业图标 / 写实渲染 / 火焰浓烟特效;中间分析路径模块大小统一、对齐严格、连接关系简洁",
"tone": "更接近高质量 Graphical Abstract 与方法路线图融合的工程论文图,不是 office 流程框图,也不是商业海报",
"stage_naming_examples_for_engineering": [
"fuel / material characterization",
"kinetics / thermodynamics analysis",
"experimental setup OR numerical model",
"boundary / operating condition design",
"process simulation or experiment",
"field / behavior evaluation",
"emission / performance analysis"
],
"color_palette": "deep blue / slate blue / charcoal as main; one low-saturation amber accent for high-temperature or risk modules ONLY when user signaled it; ≤ 3 main colors total",
"data_authenticity": "if no real data is provided, do NOT invent equations, kinetic constants, temperature values, emission factors, or chart numbers; render module summaries as qualitative phrases only",
"use_case": "能源动力 / 燃烧 / 热能工程 / 环境工程 / 材料 / 化工 等工程方向的开题答辩、综述论文、Methods 章节首图;区别于 CS/CV pipeline 的横向 stage 块结构"
}
}适用:能源动力、燃烧、热能工程、环境工程、化工、材料等工程方向的研究路线图与高质量 Graphical Abstract 融合需求;CS/CV/ML 类首选主模板。
避免事项
- 用渐变 / drop shadow / 玻璃质感 → 立刻 "PPT 风" 而不是论文风
- stage 块大小不一 / 高度不齐
- 用 emoji / 卡通图标当 stage glyph
- 用 Comic Sans / 手写体当标题字体
- 颜色超过 4 种或饱和度过高
- 输入输出缩略图分辨率明显不同
- 箭头穿过 stage 块或标签碰撞
- 中英文标签混用(除非显式双语)
- 把"对比方法"也画在同一 pipeline 上(应该用
qualitative-comparison-grid.md) - 把网络层细节(卷积核大小、激活函数)塞进 pipeline 图(这属于
neural-network-architecture.md的范畴)
多工况 / 多条件结果对比图模板
本文件用于生成「同一研究对象在不同工况 / 条件 / 组别下的多面板结果对比图」:
- 不同温度 / 压力 / 浓度 / 配比 / 时间下的实验或仿真结果
- 不同处理组 / 对照组 / 工艺方案的并列结果
- 多面板 (a)(b)(c)(d) 形式的论文 result figure
特征:
- 2×2 / 1×3 / 1×4 等统一网格布局
- 所有 panel 严格统一:相同尺寸、相同色彩逻辑、相同图例、相同字体层级、相同边距
- 白底、低饱和工程色,论文结果图风格
- 无真实数据时只做定性表达,禁止虚构数值 / 等值线 / 色标范围
适用范围
- 工程 / 物理 / 化学 / 能源 / 材料 / 环境方向的多工况结果对比
- 燃烧 / 流场 / 温度场 / 应力场 / 浓度场 等场图对比
- 不同处理组 / 不同剂量 / 不同时间点 的实验对照
- 同一指标在多个 condition 下的多面板可视化
何时使用
- 用户提到「多工况 / 多条件 / 不同 X 下的对比 / panel (a)(b)(c)(d) / 结果对比图」
- 用户希望视觉「论文 result figure,不是营销信息图」
- 比较的是同一对象在不同条件下的同类结果
不要使用:
- 用户要的是「不同方法在同一样本上的输出对比」(行=样本,列=方法) → 用
academic-figures/qualitative-comparison-grid.md - 用户要的是「单个 publication-ready 图表」(bar / line / scatter) → 用
academic-figures/publication-chart.md - 用户要的是「营销 / 信息图风格的二元对比」 → 用
infographics/comparison-infographic.md
缺失信息优先提问顺序
1. 比较对象是什么(同一现象 / 同一指标) 2. 比较的是哪些工况 / 条件(建议 2-6 个;超过 6 个考虑分两张图) 3. 每个 panel 显示的是什么(场图 / 折线 / 柱图 / 等值线 / 显微图)—— 必须所有 panel 同类型 4. 是否有真实数据(关键:决定是定性图还是定量图) 5. 网格布局(2×2 / 1×3 / 1×4 / 2×3) 6. 标签语言(中文 / 英文 / 双语) 7. 共享图例 / 共享色标(强烈建议共享)
主模板:N panel 多工况对比(统一规格)
📖 描述
整张图按统一网格分割成 N 个 panel,每个 panel 展示同一类结果在不同工况下的表现。所有 panel 共享色标 / 图例 / 字体层级 / 边距。子图标记为 (a)(b)(c)(d),标签简短克制。绝对不允许每个 panel 自成一套风格。
📝 提示词
{
"type": "学术多工况结果对比图(multi-condition comparison figure)",
"goal": "生成一张可直接放进论文 results 章节的多面板对比图,要求所有 panel 严格统一、白底、低饱和工程色、可单色印刷可读",
"canvas": {
"aspect_ratio": "{argument name=\"aspect_ratio\" default=\"4:3\"}",
"background": "pure white #FFFFFF",
"outer_padding": "50px around the grid",
"inter_panel_gap": "16-20px, identical horizontal and vertical",
"render_quality": "vector-clean look, anti-aliased, sharp text"
},
"title_caption": {
"figure_label": "{argument name=\"figure_label\" default=\"Figure X.\"}",
"caption": "{argument name=\"caption\" default=\"Comparison of results under varying conditions.\"}",
"position": "bottom-center, italic serif or compact sans-serif, smaller font size"
},
"grid_layout": {
"rows": "{argument name=\"rows\" default=\"2\"}",
"cols": "{argument name=\"cols\" default=\"2\"}",
"panel_count": "{argument name=\"panel_count\" default=\"4\"}",
"rule": "rows × cols == panel_count; all panels identical size; consistent vertical and horizontal alignment"
},
"panels": {
"panel_type": "{argument name=\"panel_type\" default=\"contour-field\"}",
"panel_type_options": "contour-field | line-chart | bar-chart | heatmap | micrograph | flow-field | bubble-chart",
"rule": "ALL panels MUST share the same panel_type; never mix bar with line within the same comparison figure",
"items": [
{
"id": "(a)",
"condition_label": "{argument name=\"panel_a_label\" default=\"Condition A\"}",
"condition_detail": "{argument name=\"panel_a_detail\" default=\"e.g. excess-air ratio λ = 1.0\"}"
},
{
"id": "(b)",
"condition_label": "{argument name=\"panel_b_label\" default=\"Condition B\"}",
"condition_detail": "{argument name=\"panel_b_detail\" default=\"e.g. excess-air ratio λ = 1.2\"}"
},
{
"id": "(c)",
"condition_label": "{argument name=\"panel_c_label\" default=\"Condition C\"}",
"condition_detail": "{argument name=\"panel_c_detail\" default=\"e.g. excess-air ratio λ = 1.4\"}"
},
{
"id": "(d)",
"condition_label": "{argument name=\"panel_d_label\" default=\"Condition D\"}",
"condition_detail": "{argument name=\"panel_d_detail\" default=\"e.g. excess-air ratio λ = 1.6\"}"
}
]
},
"panel_style": {
"frame": "thin border 1px #1F2937 OR clean axis lines without outer frame, applied identically to all panels",
"label_position": "(a) (b) (c) (d) at top-left of each panel, bold sans-serif, 11pt",
"condition_label_position": "centered above each panel OR inside each panel top-right, identical position across all panels",
"axis_labels": "shared if possible; if shown, identical font size, identical tick density across panels",
"internal_titles": "AVOID per-panel decorative titles; rely on (a)(b)(c)(d) + condition label only"
},
"shared_legend": {
"enabled": "{argument name=\"shared_legend_enabled\" default=\"true\"}",
"position": "{argument name=\"shared_legend_position\" default=\"right-of-grid\"}",
"rule": "single legend / colorbar shared across ALL panels; never give each panel its own legend with different range",
"colorbar_range": "{argument name=\"colorbar_range\" default=\"qualitative-low-to-high\"}",
"colorbar_range_rule": "if user provided a numerical range, use it; otherwise render as a qualitative gradient labeled 'low → high' with NO fabricated numerical ticks"
},
"color_logic": {
"rule": "≤ 3 main colors total; if a sequential colormap is used, choose a perceptually uniform low-saturation engineering colormap (e.g. viridis-like, blue-to-orange, gray-to-deep-blue); apply the SAME colormap and SAME range to every panel",
"must_print_grayscale_readable": true
},
"data_authenticity": {
"user_provided_real_data": "{argument name=\"has_real_data\" default=\"false\"}",
"rule_when_false": "render the panels as QUALITATIVE schematics: smooth gradient fields, generic shapes, no numerical tick labels on the colorbar, no specific values in axes; explicitly avoid the visual impression of a real dataset",
"rule_when_true": "use the user-provided values; never extrapolate, interpolate, or invent additional values"
},
"constraints": {
"must_keep": [
"all panels identical size, identical aspect, identical position scheme",
"shared color logic and shared legend across all panels",
"white background, no gradient backdrop, no decorative pattern",
"(a)(b)(c)(d) labels in identical position and identical style across all panels",
"only sans-serif typography, identical font family across all panels",
"the figure should look like it came from a results section of an engineering or science journal"
],
"avoid": [
"different colormap or different color range per panel",
"different chart type per panel (e.g. mixing bar and line)",
"decorative panel titles, hero panel that visually dominates the rest",
"saturated brand colors, neon, vivid gradients",
"3D effects, drop shadows, glossy fills, lens flare",
"fabricated numerical tick values, fabricated colorbar ranges, fabricated isolines",
"marketing-poster aesthetics, infographic-collage aesthetics",
"watermarks, copyright stamps"
]
}
}参数策略
- 必问:
panel_count、panel_type(所有 panel 同一类型)、has_real_data - 可默认:
aspect_ratio、grid_layout(2×2 是最常见)、shared_legend_enabled(true) - 可随机:每个 condition 的
*_detail措辞(用户给了控制变量名时可学术化)
自动补全策略
- 用户没说有没有真实数据 → 必须先确认:
has_real_data= false 时全图走定性渲染 - 用户给了不同 condition 但没说每 panel 的具体取值 → 在 condition_detail 里用占位短语(如
λ = X1),不要编造数字 - 用户给了 panel 数但 row × col 不匹配 → 自动选最接近正方的网格(2×2 / 2×3 / 3×3)
- 用户说"中文论文 / 答辩" → 切换标签为中文 + 字体 PingFang / 思源黑
变体 1:横向 1×N(适合窄 panel 比较)
{
"type": "横向 1×N 多工况对比",
"modify": {
"layout": "rows = 1, cols = N(建议 N ≤ 4)",
"use_case": "panel 内部是窄柱图 / 窄折线,更适合横向铺开;或论文双栏排版需要横向单行"
}
}适用:单栏 / 双栏论文格式中的横向比较。
变体 2:行列双因子矩阵(M×N)
{
"type": "双因子矩阵对比",
"modify": {
"layout": "rows = M(一种因子的不同水平),cols = N(另一种因子的不同水平)",
"rule": "顶部一行写列因子标签,最左一列写行因子标签;panel 内部样式严格统一",
"use_case": "需要同时变化两个独立变量(如温度 × 含水率,或时间 × 浓度)"
}
}适用:双因子实验设计的结果展示,正交试验结果可视化。
变体 3:定性场图渲染(无真实数据)
{
"type": "定性场图多工况对比",
"modify": {
"panel_type": "contour-field",
"data_authenticity": {
"user_provided_real_data": false,
"rule": "render smooth qualitative gradient fields with NO numerical tick labels and NO specific isoline values; the colorbar shows 'low → high' as a qualitative scale only",
"intent": "visually communicate 'higher temperature in panel (b)' without claiming any specific value"
},
"use_case": "答辩 / 开题阶段尚未拿到数据,需要先讲清研究思路时使用"
}
}适用:示意性结果对比、方法论说明阶段。
避免事项
- 给每个 panel 用不同 colormap / 不同 range → 直接破坏可比性
- 在没有真实数据时画出带具体数值的等值线 / 色标刻度(严格禁止虚构数据)
- 让某一 panel 视觉权重明显大于其他 panel(不允许"主图 + 辅图"的结构)
- 在每个 panel 加独立的装饰性标题
- 把不同类型的图(bar / line / contour)混排在同一对比图里
- 使用饱和 brand 色或霓虹渐变
- 把"对比方法"的逻辑(行=样本×列=方法)误用到本模板(请改用
qualitative-comparison-grid.md) - 加水印 / 期刊 logo / 设备品牌标
神经网络架构图模板
本文件用于生成"论文中那种神经网络架构图":
- Transformer / Encoder-Decoder 架构图
- U-Net / FPN / 多尺度网络架构
- GAN / Diffusion / VAE 架构
- Attention 机制示意
- 自定义模型架构图
特征:
- 多个 layer 块按数据流方向排布(横向或竖向)
- 每个 layer 块有:层名 + tensor shape 标注(H × W × C)
- 跳连 / residual / attention 连线清晰
- 颜色编码不同 layer 类型(Conv / Attention / FC / Norm)
- 出版物级,白底克制
适用范围
- 论文中的 model architecture figure
- 综述论文 framework
- 答辩 PPT 模型介绍页
- 教学 slide 中的网络示意
何时使用
- 用户提到 "网络架构 / network architecture / model architecture / Transformer / U-Net / GAN / Diffusion / VAE"
- 用户希望「层级清晰、tensor shape 标准、跳连一目了然」
- 用户希望视觉「论文风、白底、彩色编码 layer 类型」
不要使用:
- 用户要的是「方法 pipeline 总览」(多 stage 业务流)→ 用
academic-figures/method-pipeline-overview.md - 用户要的是「系统架构图」(前端 + 后端 + DB)→ 用
technical-diagrams/system-architecture.md - 用户要的是「数据流向 / ER 图」 → 用
technical-diagrams/er-diagram.md - 用户要的是「概念示意 / 注意力可视化」(自由度高)→ 用
academic-figures/scientific-schematic.md
缺失信息优先提问顺序
1. 模型类型(Encoder-Decoder / U-Net / Transformer / GAN / Diffusion / 自定义) 2. 主干网络层数 / 每层类型(如「6 层 Transformer encoder + 6 层 decoder + 8 头 attention」) 3. Tensor shape(输入分辨率 / 通道数 / 序列长度) 4. 是否有跳连 / residual / cross-attention 5. 是否有 multi-task / multi-head 输出 6. 是否要中文标签(论文图通常英文) 7. 比例(横向 16:9 / 2:1,符合论文双栏)
主模板:Transformer / Encoder-Decoder 架构图
📖 描述
整张图横向流动:左输入 embedding → 多层 encoder 块 → cross-attention → 多层 decoder 块 → 右输出 head。每个 layer 块标注层类型与 tensor shape,跳连用弧形虚线。
📝 提示词
{
"type": "神经网络架构图(neural network architecture diagram)",
"goal": "生成论文级别的网络架构图:层级清晰、tensor shape 标注、跳连分明、可单色印刷可读",
"canvas": {
"aspect_ratio": "{argument name=\"aspect_ratio\" default=\"16:9\"}",
"background": "white #FFFFFF",
"outer_padding": "60px"
},
"model_meta": {
"name": "{argument name=\"model_name\" default=\"Our Transformer\"}",
"input_spec": "{argument name=\"input_spec\" default=\"Input Image: 224×224×3\"}",
"output_spec": "{argument name=\"output_spec\" default=\"Class Logits: 1000\"}"
},
"layer_groups": {
"rule": "use color-coded blocks per layer type — keep palette ≤ 5 muted academic colors",
"color_legend": [
{ "type": "Embedding / PatchEmbed", "fill": "#E0E7FF", "border": "#6366F1" },
{ "type": "Self-Attention", "fill": "#FEE2E2", "border": "#DC2626" },
{ "type": "Cross-Attention", "fill": "#FEF3C7", "border": "#D97706" },
{ "type": "Feed Forward / MLP", "fill": "#D1FAE5", "border": "#059669" },
{ "type": "Norm / Residual", "fill": "#F3F4F6", "border": "#6B7280" }
]
},
"layers": {
"count": "{argument name=\"layer_count\" default=\"8\"}",
"items": [
{ "id": "L1", "type": "Embedding / PatchEmbed", "name": "Patch Embed", "shape": "196×768" },
{ "id": "L2", "type": "Norm / Residual", "name": "LayerNorm", "shape": "196×768" },
{ "id": "L3", "type": "Self-Attention", "name": "Multi-head Self-Attn (×8)", "shape": "196×768", "annotation": "× N=6 (encoder)" },
{ "id": "L4", "type": "Feed Forward / MLP", "name": "FFN", "shape": "196×768" },
{ "id": "L5", "type": "Cross-Attention", "name": "Cross-Attn", "shape": "K×768" },
{ "id": "L6", "type": "Self-Attention", "name": "Decoder Self-Attn", "shape": "K×768", "annotation": "× N=6 (decoder)" },
{ "id": "L7", "type": "Feed Forward / MLP", "name": "FFN", "shape": "K×768" },
{ "id": "L8", "type": "Norm / Residual", "name": "Output Head (Linear)", "shape": "K×C" }
]
},
"block_style": {
"shape": "rounded rectangle (corner radius 4-6px)",
"size_rule": "blocks of same layer type share identical width and height; visually grouped",
"border": "1.2px solid (use the type's border color)",
"fill": "use the type's fill color (very light tint)",
"label_text": "layer name on first line (sans-serif bold 10-11pt) + tensor shape on second line (monospace italic 9pt)",
"annotation_text": "if 'annotation' present (e.g. '× N=6'), draw it as a curly brace with label on the right side of the repeated block"
},
"connections": {
"main_flow": {
"style": "thin black solid arrows (1.2px), horizontal left → right",
"arrowhead": "small filled triangle"
},
"residual": {
"enabled": "{argument name=\"residual_enabled\" default=\"true\"}",
"style": "curved dashed arrow arcing above the main flow, label '+' near join",
"rule": "draw residual from input of attention block to its output"
},
"cross_attention": {
"enabled": "{argument name=\"cross_attention_enabled\" default=\"true\"}",
"style": "horizontal arrow from encoder side feeding into decoder cross-attn, label 'K, V'",
"rule": "encoder output is shown as K, V input to decoder cross-attn"
}
},
"extras": {
"show_param_count": {
"enabled": "{argument name=\"show_params\" default=\"false\"}",
"rule": "if true, add parameter count below each major group (e.g. '85M params')"
},
"highlight_novelty": {
"enabled": "{argument name=\"highlight_novelty\" default=\"true\"}",
"rule": "if true, surround the user's contributed module with a thicker dashed orange border + label 'Ours' / 'Novel'"
}
},
"constraints": {
"must_keep": [
"tensor shapes are accurate and labeled in monospace font",
"color encodes layer type consistently across the figure",
"all layers of the same type have identical block size",
"white background, no gradient, no decoration",
"all labels in English by default (or all Chinese if explicitly requested), no mixing",
"must remain readable when printed in grayscale (rely on shape and label, not color alone)",
"novel contribution (if any) is clearly marked"
],
"avoid": [
"3D extruded blocks, drop shadows, glossy fills",
"rainbow palette (>5 colors)",
"cartoon icons, emoji",
"freeform 'art-style' blobs instead of crisp rectangles",
"tensor shapes typeset in proportional font",
"arrows crossing through blocks",
"missing tensor shape labels (the figure is then useless for paper review)",
"unlabeled cross-attention (must say K, V)"
]
}
}参数策略
- 必问:
layer_count、每层的type和shape - 可默认:
aspect_ratio(16:9)、background(白)、color_legend(默认 5 类配色)、block_style - 可随机:blocks 内每行的精确字号 / padding,annotation 摆放位置
自动补全策略
- 用户给「我用 Transformer」但没给细节 → 反问关键参数(层数、头数、隐藏维度、序列长度);不要瞎编模型规模
- 用户给「U-Net」 → 自动用 contracting + expansive 双臂布局变体(见变体 2)
- 用户没说有没有 residual → 默认
residual_enabled: true(绝大多数现代网络都有) - 用户没说有没有 novelty → 默认
highlight_novelty: true(论文图一般要标自己的贡献) - 用户没说参数量 → 默认
show_params: false(除非用户提到模型规模对比)
变体 1:U-Net / FPN 双臂架构
{
"type": "U-Net / FPN 双臂架构图",
"modify": {
"layout": "U 形:左臂下采样(contracting path)+ 中央 bottleneck + 右臂上采样(expansive path),每层之间有水平 skip connection",
"annotation": "skip 用横向虚线箭头标注,特征图用渐窄 / 渐宽的矩形示意 spatial 维度变化"
}
}适用:U-Net、FPN、HRNet、所有 encoder-decoder 分割网络。
变体 2:GAN / Diffusion 双网络对抗 / 多步推理
{
"type": "GAN / Diffusion 架构图",
"modify": {
"layout_gan": "上方 Generator(noise → image)+ 下方 Discriminator(image → real/fake),中间共享生成图像作为 D 的输入",
"layout_diffusion": "横向 timestep 序列 t=T → t=0,每个 timestep 是同一个 U-Net 实例,标 't' 嵌入条件"
}
}适用:GAN 系列、扩散模型、Score-based 模型。
变体 3:Multi-task / Multi-head 输出
{
"type": "多任务 / 多头输出架构图",
"modify": {
"layout": "共享 backbone 在中央 → 右侧分叉成 2-4 个 task head(如 classification head / regression head / segmentation head)",
"annotation": "每个 head 旁边标对应 loss 函数和权重 λ"
}
}适用:多任务学习、检测 + 分割、辅助监督。
避免事项
- tensor shape 缺失或随便写 → 论文图核心信息没了
- 用渐变 / 3D 立方体堆叠 → 像 PPT 不像论文
- 颜色 ≥ 6 种 → 失去 layer 类型语义
- 没有 residual / cross-attention 标注(如果架构里有)→ 误导读者
- 用 Comic Sans / 手写字体
- 跳连箭头穿过 layer 块
- 同一类 layer 块大小不一致
- 中英文标签混用
- 把"训练 loss"画进结构图(应该单独一张 training figure 或 caption 里说明)
- 在结构图里塞具体超参数表(应该走 table,不进 figure)
Publication-Ready 数据图表模板
本文件用于生成"论文 / 报告里出现的标准数据图表":
- Bar chart / grouped bar chart(消融实验、方法对比)
- Line chart / 训练曲线(loss / accuracy 随 epoch)
- Scatter plot(性能-效率 trade-off)
- Box plot / Violin plot(统计分布)
- Heatmap(confusion matrix / attention map / 相关性矩阵)
特征:
- matplotlib / seaborn / R ggplot2 出版物风
- 含坐标轴 + 标签 + 单位 + 图例 + 误差棒 + 显著性标记
- 字体 ≥ 10pt(确保打印可读)
- 配色克制(≤ 6 色),可单色印刷
- 网格线极淡或无
⚠️ 重要免责声明:本模板生成的是"出版级图表的视觉呈现",不是真实数据可视化。GPT Image 2 不能保证坐标和数据的精确对应。
>
- 如果你需要"展示一张论文图表的样子" / "做封面 / hero 配图" → 用本模板
- 如果你需要"用真实数据生成可发表的图表" → 请用 matplotlib / seaborn / ggplot2 / Plotly
适用范围
- 论文方法对比 chart 的视觉示例
- 教学 slide 中"看一眼这个图就懂"的演示图
- Blog / 公众号配图 — "我们的方法在这个 chart 上表现"
- 投资人 deck 中的"数据 mock"
- 演示用、可视化教学用的图表
何时使用
- 用户提到 "publication chart / matplotlib 风 / seaborn 风 / 论文图表 / bar chart / line chart / scatter / heatmap / confusion matrix"
- 用户希望「白底、克制、可单色、像 NeurIPS 论文那种图表」
- 用户明确知道这只是视觉呈现,不依赖坐标精度
不要使用:
- 用户要的是「真实数据可视化产出」 → 推荐 matplotlib / seaborn / Plotly
- 用户要的是「KPI 仪表盘 / 数据回顾」 → 用
infographics/kpi-dashboard-infographic.md - 用户要的是「商业 PPT 数据页」 → 用
slides-and-visual-docs/visual-report-page.md - 用户要的是「手绘风信息图」 → 用
infographics/hand-drawn-infographic.md
缺失信息优先提问顺序
1. 图表类型(bar / line / scatter / box / violin / heatmap / pie) 2. 主题("我们方法在 ImageNet 上的 accuracy vs baselines") 3. X 轴和 Y 轴名称 + 单位 4. 数据系列数量(单一系列 / 多系列) 5. 是否有误差棒、显著性标记 * 6. 配色基调(学术克制 / 强调对比 / 黑白单色) 7. 图标题 + caption(论文 figure 一般有 caption)
主模板:Publication-Ready Bar Chart(默认)
📖 描述
整张图是一张标准学术 bar chart:横轴为方法 / 类别,纵轴为指标,多个方法对比,含误差棒、显著性 *、图例。整体白底,sans-serif 字体,限定配色。
📝 提示词
{
"type": "Publication-Ready Bar Chart(学术出版级条形图)",
"goal": "生成视觉呈现一张论文 / 报告中的 bar chart,要求白底、克制、专业、可单色印刷",
"canvas": {
"aspect_ratio": "{argument name=\"aspect_ratio\" default=\"4:3\"}",
"background": "white #FFFFFF",
"outer_padding": "60px"
},
"title": {
"text": "{argument name=\"title\" default=\"Accuracy on ImageNet-1K\"}",
"position": "top-center, sans-serif bold 13pt",
"subtitle": "{argument name=\"subtitle\" default=\"\"}"
},
"axes": {
"x_axis": {
"label": "{argument name=\"x_label\" default=\"Method\"}",
"categories": [
"{argument name=\"cat1\" default=\"ResNet-50\"}",
"{argument name=\"cat2\" default=\"ViT-B\"}",
"{argument name=\"cat3\" default=\"Swin-B\"}",
"{argument name=\"cat4\" default=\"ConvNeXt-B\"}",
"{argument name=\"cat5\" default=\"Ours\"}"
],
"tick_label_rotation": "0deg or 30deg if labels are long"
},
"y_axis": {
"label": "{argument name=\"y_label\" default=\"Top-1 Accuracy (%)\"}",
"range": "{argument name=\"y_range\" default=\"75 to 86\"}",
"tick_format": "decimal or percent",
"gridlines": "very faint horizontal gridlines (light gray dashed, low opacity)"
}
},
"bars": {
"style": "vertical bars, ~30-40% width of category slot, gap between bars",
"color_rule": {
"default": "use a single muted color for all baselines (e.g. slate blue #64748B), highlight 'Ours' bar in accent color (e.g. orange #D97706 or red #DC2626)",
"alternative": "if comparing methods grouped by family, use 2-3 muted colors to encode family"
},
"value_labels": {
"enabled": "{argument name=\"value_labels_enabled\" default=\"true\"}",
"rule": "show numeric value above each bar, sans-serif 9pt bold, e.g. '82.3'"
}
},
"error_bars": {
"enabled": "{argument name=\"error_bars_enabled\" default=\"true\"}",
"style": "thin black T-bar at top of each bar, ±std or ±95% CI",
"annotation": "mention what the error represents in caption (e.g. 'error bars show ±1 std over 5 runs')"
},
"significance_markers": {
"enabled": "{argument name=\"significance_enabled\" default=\"false\"}",
"rule": "if true, draw thin horizontal brackets between compared bars, with * / ** / *** annotation above (p<0.05 / p<0.01 / p<0.001)"
},
"legend": {
"enabled": "{argument name=\"legend_enabled\" default=\"false\"}",
"rule": "only show legend if multiple colors / groups used; place top-right inside or outside the plot area",
"items": ["Baselines", "Ours"]
},
"caption": {
"enabled": "{argument name=\"caption_enabled\" default=\"true\"}",
"label": "{argument name=\"figure_label\" default=\"Figure 3.\"}",
"text": "{argument name=\"caption_text\" default=\"Top-1 accuracy on ImageNet-1K. Our method outperforms all baselines while using fewer parameters. Error bars show ±1 std over 5 runs.\"}",
"style": "below the chart, italic serif or compact sans-serif, justified, smaller font"
},
"constraints": {
"must_keep": [
"white background, no gradient, no pattern fills",
"sans-serif fonts only (Helvetica / Inter / Arial); axis tick labels ≥ 9pt, axis labels ≥ 11pt, title ≥ 13pt",
"color palette ≤ 6 colors, must remain readable in grayscale",
"all axes have labels and units",
"no 3D bar effects, no perspective tilt",
"if multiple bars per category, group them with consistent spacing",
"Ours bar is visually distinguishable (color or annotation)"
],
"avoid": [
"rainbow colors / saturated palette",
"3D extruded bars / pie charts (3D distorts perception)",
"missing axis labels or units",
"unreadable tick labels (too small or rotated awkwardly)",
"decorative background images / textures",
"emoji / cartoon icons inside or around bars",
"value labels overlapping bars or each other",
"random / irrelevant accent colors",
"fake precision: don't render bar heights to imply real numbers — keep it clearly illustrative"
]
}
}参数策略
- 必问:图表类型(如果不是 bar)、
title、x_label/y_label/ 单位、categories - 可默认:
aspect_ratio(4:3)、background(白)、error_bars_enabled(true)、value_labels_enabled(true)、legend_enabled(false 单系列时) - 可随机:bar 宽度、tick 数量、网格线密度(在合理范围)
自动补全策略
- 用户给"我有 5 个方法的 accuracy 对比" → 自动用 default 5 categories,highlight 最后一个为 Ours
- 用户没指定 y_range → 推断(基于数值范围 ± 5%)
- 用户没说 error → 默认开启 error_bars(论文标准做法)
- 用户没说 significance → 默认关闭(除非是统计学论文)
- 用户说"不是 bar" → 切换到对应变体
变体 1:Line Chart(训练曲线 / 时间序列)
{
"type": "Publication-Ready Line Chart(学术出版级折线图)",
"modify": {
"x_axis_typical": "epoch / step / time / iteration",
"y_axis_typical": "loss / accuracy / metric",
"lines_count": "1-5 series, each a different muted color",
"line_style": "solid 1.5px main line + optional shaded area (semi-transparent same color) for std band",
"markers": "optional small markers at sparse intervals (circles / triangles), not on every point",
"legend": "always enabled for multi-series, top-right or below",
"rule_extra": "axes can be log-scale if data spans orders of magnitude (label as 'log scale')"
}
}适用:训练曲线、time series 趋势、ablation 随超参变化、scaling laws。
变体 2:Scatter Plot(trade-off 图)
{
"type": "Publication-Ready Scatter Plot(学术出版级散点图)",
"modify": {
"typical_use": "performance vs efficiency trade-off (e.g. accuracy vs FLOPs / latency / params)",
"x_axis_typical": "compute / params / latency (often log scale)",
"y_axis_typical": "accuracy / metric",
"point_style": "filled circles, size encodes a third dimension (e.g. model size), color encodes a category (e.g. method family)",
"label_each_point": "small text label next to each point with method name (no leader lines unless crowded)",
"ours_emphasis": "Our method points are larger and use accent color + black border",
"frontier_line": "optional: draw a Pareto frontier curve to show 'we push the frontier'"
}
}适用:性能-效率 trade-off、参数 vs 准确率、Pareto frontier。
变体 3:Heatmap(confusion matrix / attention map / 相关性)
{
"type": "Publication-Ready Heatmap(学术出版级热力图)",
"modify": {
"grid": "N × N(默认 5×5 至 10×10)",
"color_map": "sequential — viridis / Blues / Reds / 灰阶;diverging(如相关矩阵)— RdBu_r 红蓝双向",
"cell_annotation": "show numeric value inside each cell in monospace, color flips for readability on dark cells",
"axes_label": "row labels = ground truth, column labels = predicted(confusion matrix 场景)",
"colorbar": "right side vertical colorbar with label and ticks",
"rule_extra": "always include colorbar; never use rainbow colormap for sequential data (jet 已被学界淘汰)"
}
}适用:confusion matrix、attention 权重可视化、相关性矩阵、ablation grid。
变体 4:Box Plot / Violin Plot(统计分布)
{
"type": "Publication-Ready Box / Violin Plot(学术出版级分布图)",
"modify": {
"typical_use": "compare distributions across methods / conditions / groups",
"elements": "box (Q1, median, Q3) + whiskers (1.5 IQR) + outlier dots; violin 形状叠加显示密度",
"median_line_emphasis": "median 线粗实线,颜色区分 group",
"annotation": "可叠加 swarm / strip plot 显示每个数据点",
"rule_extra": "如果用 violin,violin 内部仍画 box;不要纯 violin(损失中位数信息)"
}
}适用:实验重复结果分布、跨数据集 / 跨用户 / 跨条件分布对比。
避免事项
- 用 3D 柱 / 3D 饼 → 严重不专业
- 用彩虹 / jet colormap 表示连续值(学界已抛弃)
- 漏掉单位 / 漏掉坐标轴标签
- value 标签过小读不清
- 没有 caption 或 caption 没解释 error bar
- 把 4-5 个不相关 chart 拼一张(应该用 multi-panel figure 模板,每个 sub 图独立)
- 假装精确(暗示这是真数据但其实是 illustrative)
- 用花哨字体(Comic Sans / 手写体)
- 加水印 / 装饰背景
- 漏掉图例(多系列必须有)
- 多 series 但配色完全相同
- 漏掉 Ours 高亮(论文图通常要让 reviewer 一眼看出你的)
多方法 Qualitative 对比网格模板
本文件用于生成"论文 qualitative results 对比网格":
- CV 论文:多方法分割 / 检测 / 生成结果对比
- NLP 论文:多方法生成文本对比(截图式)
- 3D / 重建论文:多方法重建结果对比
- Diffusion / 图像生成论文:不同 prompt × 不同方法的网格
- Ablation study 的视觉对比
特征:
- 严格的网格:行 = 样本 / 输入,列 = 方法(含 GT 和 Ours)
- 列首行有方法名(带 citation)
- Ours 列通常加边框 / 高亮
- 单元格内容统一(图片 / 文本片段 / heatmap)
- 网格之间留细 gap,整体白底
- 可附 caption 解释
适用范围
- 论文 qualitative results section
- Ablation study 的视觉对比
- 顶会 supplementary 大网格图
- 综述论文 method gallery
- 答辩 PPT 对比页
何时使用
- 用户提到 "qualitative / 对比图 / comparison grid / methods comparison / ablation visual"
- 用户希望「行=样本、列=方法的标准论文对比网格」
不要使用:
- 用户要的是「双产品消费对比」 → 用
infographics/comparison-infographic.md - 用户要的是「多人头像网格」 → 用
avatars-and-profile/character-grid-portrait.md - 用户要的是「数据图表」 → 用
academic-figures/publication-chart.md - 用户要的是「视频帧序列」 → 用
storyboards-and-sequences/
缺失信息优先提问顺序
1. 行数(样本数,建议 3-6 行) 2. 列数(方法数,建议 3-6 列,含 Input/GT 和 Ours) 3. 每列的方法名(含 citation 引用,如 "Method A [12]") 4. 单元格内容类型(RGB 图 / mask / heatmap / 文本片段 / 3D 渲染) 5. 是否要 row labels(左侧标"Sample 1 / 2 / ..."或"Easy / Medium / Hard") 6. 是否要在某些位置加红框 zoom-in(focus area) 7. 是否要 caption 注释
主模板:Qualitative comparison grid (M rows × N cols)
📖 描述
整张图是严格的 M×N 网格:每一行是一个样本,每一列是一个方法。最左可加 row labels,最上一行是列首(方法名 + citation)。Ours 列加边框高亮,可在某些 cell 内画红色 zoom-in 框。
📝 提示词
{
"type": "Qualitative Comparison Grid(论文级多方法多样本对比网格)",
"goal": "生成一张可直接放进论文 qualitative results 章节的网格对比图,要求严格对齐、清晰列首、Ours 高亮、可单色印刷可读",
"canvas": {
"aspect_ratio": "{argument name=\"aspect_ratio\" default=\"4:3\"}",
"background": "white #FFFFFF",
"outer_padding": "40px"
},
"grid": {
"rows": "{argument name=\"rows\" default=\"4\"}",
"cols": "{argument name=\"cols\" default=\"5\"}",
"cell_size_rule": "all cells identical size; gap between cells 4-6px",
"cell_aspect": "{argument name=\"cell_aspect\" default=\"square\"}"
},
"headers": {
"column_headers": {
"enabled": true,
"items": [
{ "id": "C1", "label": "{argument name=\"col1_name\" default=\"Input\"}" },
{ "id": "C2", "label": "{argument name=\"col2_name\" default=\"Method A [12]\"}" },
{ "id": "C3", "label": "{argument name=\"col3_name\" default=\"Method B [34]\"}" },
{ "id": "C4", "label": "{argument name=\"col4_name\" default=\"Method C [56]\"}" },
{ "id": "C5", "label": "{argument name=\"col5_name\" default=\"Ours\"}", "highlight": true }
],
"style": "centered above each column, sans-serif bold 11pt, citations in smaller superscript or in [brackets]"
},
"row_labels": {
"enabled": "{argument name=\"row_labels_enabled\" default=\"true\"}",
"items": [
"{argument name=\"row1_label\" default=\"Sample 1\"}",
"{argument name=\"row2_label\" default=\"Sample 2\"}",
"{argument name=\"row3_label\" default=\"Sample 3\"}",
"{argument name=\"row4_label\" default=\"Sample 4\"}"
],
"style": "rotated 90° on the left margin OR placed above each row in italic 10pt"
}
},
"cell_content": {
"type": "{argument name=\"content_type\" default=\"rgb_image\"}",
"options_explained": {
"rgb_image": "natural images / photos",
"segmentation_mask": "color-coded mask overlays",
"heatmap": "viridis / jet style heatmap",
"depth_map": "grayscale or turbo colormap",
"text_snippet": "rendered text block in a code-like box",
"3d_render": "rendered 3D mesh from a fixed viewpoint",
"side_by_side": "two halves: input | result"
},
"consistency_rule": "all cells in the same row should depict the SAME underlying sample so the comparison is fair"
},
"highlights": {
"ours_column": {
"enabled": true,
"style": "thicker border 1.5px in deep red / accent color (e.g. #DC2626) around each Ours cell"
},
"zoom_in_boxes": {
"enabled": "{argument name=\"zoom_in_enabled\" default=\"false\"}",
"rule": "if true, draw small red rectangles inside cells highlighting interesting regions; same red box appears at the same coordinate across the row to make comparison fair",
"callout_style": "optional zoomed crop placed below the row, connected by thin lines"
}
},
"caption": {
"enabled": "{argument name=\"caption_enabled\" default=\"true\"}",
"label": "{argument name=\"figure_label\" default=\"Figure 4.\"}",
"text": "{argument name=\"caption_text\" default=\"Qualitative comparison with state-of-the-art methods. Our method (last column) preserves fine details and reduces artifacts.\"}",
"style": "below the grid, italic serif or compact sans-serif, justified, smaller font"
},
"constraints": {
"must_keep": [
"all cells identical size and tightly aligned",
"white or near-white background, no gradient",
"column headers clearly above each column with citation",
"Ours column visually distinguished (border / shaded header)",
"row content depicts the same sample across all methods",
"if zoom-in boxes used, position is identical across the row",
"labels in English by default, no mixing with Chinese unless requested",
"must remain interpretable in grayscale print"
],
"avoid": [
"different cell sizes between rows / columns",
"random colors as cell backgrounds (cells are content, not decoration)",
"missing citations on baseline methods",
"ours column hidden or unmarked",
"rotated cells / tilted layouts (must be axis-aligned)",
"decorative emoji / cartoon icons inside cells",
"varying content type per row (e.g. one row mask, next row RGB) without explicit row label",
"more than 6 cols (becomes unreadable in two-column paper format)"
]
}
}参数策略
- 必问:
rows、cols、每列方法名(含 citation)、content_type - 可默认:
aspect_ratio(4:3)、row_labels_enabled(true)、caption_enabled(true) - 可随机:列间 gap 精确像素、字体大小(在合理范围内)
自动补全策略
- 用户给 "我有 4 个方法 + ours" → 自动加上 Input 列(成为 5 列:Input / M1 / M2 / M3 / M4 / Ours,共 6 列)
- 用户没给 row labels → 默认用 "Sample 1, 2, 3, ..." 或反问是否要分难易度
- 用户没给 citation → 提示 "建议加 [n] 引用占位" 而不是擅自编造
- 用户说 "ablation study" → 列名改为 "w/o A", "w/o B", "Full" 等消融变体
- 用户说 "需要 zoom-in" → 启用
zoom_in_enabled并提示需要标 region 坐标
变体 1:纯文本 NLP qualitative 对比
{
"type": "NLP qualitative comparison grid",
"modify": {
"content_type": "text_snippet",
"cell_aspect": "tall rectangle (e.g. 2:3 portrait)",
"cell_styling": "monospace font in cell, black text on white, with key tokens highlighted in colored boxes",
"row_labels": "input prompt / question 显示在每一行最左",
"use_case": "对比多个 LLM / 翻译 / summarization 输出"
}
}适用:NLP 论文生成结果对比、机器翻译质量对比。
变体 2:分割 mask 多列对比(含彩色 overlay)
{
"type": "Segmentation mask comparison grid",
"modify": {
"content_type": "segmentation_mask",
"cell_styling": "RGB image base + 半透明 mask 叠加;每类颜色一致;GT 列与 Ours 列容易对比",
"extras": "在 cells 下方可加 'mIoU: 0.78' 等定量指标小字",
"color_legend": "图右下角附小图例:颜色 → 类别名"
}
}适用:语义分割、实例分割、医学影像分割论文。
变体 3:Diffusion / 生成模型 prompt × method 矩阵
{
"type": "Generation prompt × method matrix",
"modify": {
"rows": "different text prompts (left labels show prompt text)",
"cols": "different generation methods or different sampling steps",
"cell_content": "generated images, all from same prompt across the row",
"extras": "可在 ours 列加 '↑ +0.3 CLIP score' 小标"
}
}适用:扩散模型、文本到图像生成、图像编辑方法对比。
避免事项
- 单元格大小不一致 → 完全失去对比意义
- 缺 citation → 同行评审会扣分
- Ours 列没有标记 → 读者不知道哪个是你的
- 同一行的样本不一致(这一行第一列是猫,第二列是狗)→ 对比不成立
- 添加渐变 / 阴影 / 圆角过大 → 不像论文
- 用 emoji 或 cartoon 装饰 → 严重不专业
- 列数 > 6 → 论文双栏排版下看不清
- 没有 caption → 读者不知道这张图想说什么
- zoom-in 框位置在不同 cell 不一致 → 对比不公平
开题 / 答辩 / 汇报研究总览图模板
本文件用于生成「开题答辩首页 / 论文汇报首页 / 组会引导页的研究总览图」:
- 硕博开题答辩首页的研究框架图
- 中期 / 终期答辩首页的总览图
- 组会 / 学术汇报 PPT 的引导页
- Lab 主页 / 课题介绍的研究总览
特征:
- 高层级、易读、适合 PPT 一页展示
- 5 个核心模块:背景 / 目标 / 研究模块 1 / 研究模块 2 / 预期结果
- 白底、低饱和工程色、≤3 主色,论文图感而非商业咨询路演图
- 文字精炼到短语,禁止文字墙
适用范围
- 开题 / 中期 / 答辩首页(学术汇报)
- 组会引导页 / 课题汇报 / Lab meeting cover
- 项目立项书的研究框架图(学术风)
- Faculty 个人主页 / Lab 主页的"current research"区块
何时使用
- 用户提到「开题 / 答辩 / 总览图 / 研究框架 / PPT 首页 / 引导页 / lab 主页」
- 用户希望视觉「学术答辩 PPT 首页风,正式克制工程化,不要咨询路演风」
- 用户已能给出 5 个左右的核心模块
不要使用:
- 用户要的是「期刊投稿 Graphical Abstract」 → 用
academic-figures/graphical-abstract.md - 用户要的是「方法 pipeline」 → 用
academic-figures/method-pipeline-overview.md - 用户要的是「商业 / 投资人路演封面」 → 用
slides-and-visual-docs/visual-report-page.md - 用户要的是「品牌主视觉海报」 → 用
poster-and-campaigns/brand-poster.md
缺失信息优先提问顺序
1. 课题 / 研究主题(写在标题区) 2. 答辩类型(开题 / 中期 / 终期 / 组会 / 立项)—— 决定语气与模块构成 3. 5 个核心模块的命名(默认是:背景 / 目标 / 研究内容 1 / 研究内容 2 / 预期结果) 4. 是否需要主观点 / 关键问题(强烈建议有,写在背景模块下方) 5. 是否需要研究对象简化示意(颗粒 / 器件 / 流程 / 系统) 6. 标签语言(中文 / 英文 / 双语;中文答辩通常中文为主,可英文副标题) 7. 比例(默认 16:9 适配 PPT;4:3 适配旧 PPT 模板)
主模板:上中下三层 + 五模块研究总览
📖 描述
整张图按"上方主题 + 中间核心模块 + 下方结果导向"分成三层。中间层包含 4-5 个研究内容模块,呈现层级清晰、对齐严格的学术布局,绝对不像商业路演 PPT。
📝 提示词
{
"type": "学术研究总览图(research overview / framework figure for thesis defense)",
"goal": "生成一张可直接放进开题答辩 / 论文汇报 PPT 首页的研究总览图,要求正式克制、白底、工程化配色、明显论文图感、绝无商业路演感",
"canvas": {
"aspect_ratio": "{argument name=\"aspect_ratio\" default=\"16:9\"}",
"background": "pure white #FFFFFF",
"outer_padding": "60px around the diagram",
"render_quality": "vector-clean look, anti-aliased edges, sharp text"
},
"title_block": {
"main_title": "{argument name=\"main_title\" default=\"Research Overview\"}",
"subtitle": "{argument name=\"subtitle\" default=\"e.g. thesis topic in one short phrase\"}",
"occasion_label": "{argument name=\"occasion_label\" default=\"Thesis Proposal Defense\"}",
"position": "top-center, main_title in bold sans-serif 18-22pt, subtitle in regular 12-14pt below, occasion_label in italic gray 10pt at top-right"
},
"background_section": {
"label": "{argument name=\"background_label\" default=\"Background & Problem\"}",
"summary": "{argument name=\"background_summary\" default=\"a short phrase stating why this matters and what gap exists\"}",
"key_question": "{argument name=\"key_question\" default=\"a single research question, expressed as one sentence ≤ 18 words\"}",
"position": "upper-middle band, full width, visually anchored as 'context'"
},
"objective_section": {
"label": "{argument name=\"objective_label\" default=\"Objective\"}",
"summary": "{argument name=\"objective_summary\" default=\"a short phrase stating the research goal\"}",
"position": "directly below background, narrower than background, centered"
},
"research_modules": {
"count": "{argument name=\"module_count\" default=\"3\"}",
"items": [
{
"id": "RM1",
"name": "{argument name=\"module_1_name\" default=\"Characterization\"}",
"summary": "{argument name=\"module_1_summary\" default=\"a short phrase stating what is studied / measured\"}",
"method_hint": "{argument name=\"module_1_method\" default=\"thermogravimetric analysis\"}"
},
{
"id": "RM2",
"name": "{argument name=\"module_2_name\" default=\"Modeling\"}",
"summary": "{argument name=\"module_2_summary\" default=\"a short phrase stating the modeling / simulation focus\"}",
"method_hint": "{argument name=\"module_2_method\" default=\"CFD combustion model\"}"
},
{
"id": "RM3",
"name": "{argument name=\"module_3_name\" default=\"Optimization\"}",
"summary": "{argument name=\"module_3_summary\" default=\"a short phrase stating optimization or application focus\"}",
"method_hint": "{argument name=\"module_3_method\" default=\"parameter sweep + emission analysis\"}"
}
],
"layout": "horizontal row of equal-width modules in the central band, all modules identical size and identical style"
},
"expected_outcome_section": {
"label": "{argument name=\"outcome_label\" default=\"Expected Outcomes\"}",
"items": "{argument name=\"outcome_items\" default=\"3-4 short phrases listing deliverables, e.g. 'kinetics database', 'optimized operating window', 'engineering recommendations'\"}",
"position": "bottom band, full width, visually distinct from research_modules but stylistically consistent"
},
"module_block_style": {
"shape": "rounded rectangle (corner radius ~8px) OR stage label + thin underline",
"size_per_module": "all modules identical size, vertically aligned",
"fill": "very light tint (e.g. #F1F5F9, #ECFEFF) — at most 2 different tints; modules of the same role share the same tint",
"border": "1.2px solid #334155",
"title_text": "module name in bold sans-serif (PingFang SC / Source Han Sans for CJK; Inter / Helvetica / Arial for english), 13-14pt",
"summary_text": "single phrase, 10-11pt regular, 1-2 lines max, no period",
"method_hint_text": "italic gray 9-10pt, below summary"
},
"connectors": {
"style": "thin arrows (1.2px) with simple triangle arrowheads, dark gray #334155",
"rule": "vertical flow background → objective → modules → outcomes; modules are horizontally parallel (no inter-module arrows unless logically required)",
"decoration": "none; no curved arcs, no dashed unless explicitly indicating a feedback loop"
},
"color_palette": {
"rule": "≤ 3 main colors total, drawn from a low-saturation engineering set: deep blue #1E3A8A / slate blue #3B82F6 / charcoal #1F2937; allow ONE low-saturation accent (e.g. amber #F59E0B) for the outcome band only if user signaled emphasis",
"must_print_grayscale_readable": true
},
"typography": {
"language": "{argument name=\"language\" default=\"chinese\"}",
"rule": "chinese → PingFang SC / Source Han Sans; english → Inter / Helvetica / Arial; bilingual → primary line larger, secondary line smaller and gray",
"consistency": "all module titles identical size; all summaries identical size; never mix serif and sans-serif"
},
"constraints": {
"must_keep": [
"all research modules identical size, vertically aligned, equal weight",
"white background, no gradient, no decorative pattern",
"language and font consistent across the entire figure",
"summaries are short phrases, never full paragraphs",
"the figure must look like the cover slide of an academic defense, not a corporate roadmap or pitch deck",
"color palette ≤ 3 main colors, must remain readable in grayscale print"
],
"avoid": [
"consulting / pitch-deck aesthetics, brand campaign aesthetics",
"decorative icons, emoji, mascots, hand-drawn wobble",
"3D rendering, glossy fills, lens flare, drop shadow blocks",
"stock-photo backgrounds, photographic hero images",
"fabricated quantitative claims (no '+30% efficiency', '150 samples' unless user provided them)",
"saturated brand colors, neon, vivid gradients",
"dense text walls; no module summary should exceed 2 lines",
"watermarks, copyright stamps, university / lab logos unless explicitly requested"
]
}
}参数策略
- 必问:
main_title、5 个核心模块(背景 / 目标 / 研究内容 1-N / 预期结果)的命名 - 可默认:
aspect_ratio(16:9)、background(白色)、color_palette(深蓝/灰蓝/黑灰) - 可随机:
method_hint措辞(用户给出方法名时可学术化);occasion_label(开题 / 中期 / 终期 / 组会)
自动补全策略
- 用户给出主题但没给模块 → 反问 3-4 个研究模块,禁止编造研究内容
- 用户给出
key_question超过 18 词 → 主动建议精简或拆成 2 个子问题 - 用户没给 expected outcomes → 用占位短语(如 "deliverable 1: ...")并标注待用户补充
- 用户说"中文答辩 / 中文 PPT" →
language默认中文,主标题中文 + 英文副标题(小一号 + 灰色)
变体 1:中心主题 + 周围模块(辐射式)
{
"type": "中心主题 + 周围模块的研究总览",
"modify": {
"layout": "中央放置研究主题 / 研究对象的简化示意;周围呈环形或四象限放置 4 个研究模块;下方留出预期结果带",
"rule": "中央对象占画面 25-30%;周围模块等大、等距、对齐严格",
"use_case": "适合系统型 / 平台型课题,研究模块之间是平行而非前后依赖关系"
}
}适用:平台型课题、综合性课题、研究方向多支并行的总览。
变体 2:左右双栏(左 = 研究内容,右 = 路线 / 时间表)
{
"type": "左右双栏研究总览",
"modify": {
"layout": "左栏 = 研究内容模块(垂直堆叠 3-4 个);右栏 = 时间表 / 路线 / 里程碑(gantt 风极简)",
"rule": "左右栏宽度比约 3:2;右栏时间轴用细线 + 节点圆,节点旁标月份或学期",
"use_case": "开题答辩需要明确"做什么 + 什么时候做"的项目计划"
}
}适用:开题答辩、项目立项书需要附进度计划的场景。
变体 3:极简版(只显示研究模块,无 timeline / 无 outcome 带)
{
"type": "极简研究总览",
"modify": {
"layout": "去掉 expected_outcome_section,去掉时间轴;只保留 title + background/key_question + 3-4 个研究模块",
"use_case": "组会汇报引导页或 lab meeting cover,只需快速点出'这次要讲什么'"
}
}适用:组会 / Lab meeting / 课程汇报。
避免事项
- 把研究总览图画成商业咨询路演风(深色背景 + 大色块 + brand 色) → 立刻"非学术"
- 用 emoji / 商业图标 / 卡通插画装饰研究模块
- 把研究模块写成完整段落,每个模块超过 2 行 → 视觉拥挤
- 在"预期结果"中编造具体百分比 / 数据指标(严格禁止虚构数据)
- 让某一研究模块明显大于其他(应该等权)
- 用饱和 / 渐变 / 玻璃质感装饰背景
- 强加学校 / 实验室 logo 或 watermark(除非用户明确要求)
- 把方法 pipeline 详细图(应该用
method-pipeline-overview.md)塞进总览图中
局部对象替换工作流模板
本文件用于“将原图中某一对象替换为另一对象”的编辑任务:
- 把图中咖啡杯换成保温杯
- 把图中人物 T 恤换成卫衣
- 把车 logo 换为另一品牌
- 把宠物从猫换成狗
- 把背景里某物换为另一物
特征:
- 主体大部分保留
- 仅局部精确替换
- 周围光线 / 阴影需要适配
- 必要时配合 mask(蒙版)
适用范围
- 单一对象替换
- 多对象替换
- 配合蒙版的精确替换
何时使用
- 用户提供原图(REFERENCE_0)+ 想换某物
- 用户希望除被替换对象外其他都不动
不要使用:
- 整张换背景(用
background-replacement.md) - 仅删除某物(用
object-removal.md) - 产品 / 人像精修(用
product-retouching.md/portrait-local-edit.md)
缺失信息优先提问顺序
1. 原图中要替换的对象 2. 要替换为什么 3. 是否提供蒙版 4. 替换后是否需要重新阴影 / 反光 5. 是否保留替换对象的尺寸 / 位置
主模板:单对象替换
📖 描述
精确替换一个对象,其余画面尽量保留。
📝 提示词
以 REFERENCE_0 为基础,将 {argument name="original object" default="桌上的白色陶瓷咖啡杯"} 替换为 {argument name="replacement object" default="同尺寸的不锈钢保温杯,哑光银色,瓶身有简洁品牌字 'AURORA'"}。
保留原图中其他所有元素的位置、光线、阴影与构图;只对被替换对象本身做修改。
为新对象重新生成与原图光线方向一致的阴影、反光与材质。
不要改变其它人物、桌面、背景。参数策略
- 必问:原对象、替换对象
- 可默认:是否需要重新阴影
- 可随机:替换对象的次要细节
自动补全策略
- 默认保留原对象尺寸与位置
- 默认重新生成阴影
- 用户没指定材质时,按合理类比选
变体 1:配合蒙版的精确替换
📝 提示词
以 REFERENCE_0 为基础,使用 REFERENCE_1(蒙版)所标记的区域,精确替换 {argument name="object to replace" default="人物的白色 T 恤"} 为 {argument name="new object" default="深蓝色长袖卫衣,胸前印有 'AURORA' 字样"}。
仅对蒙版区域做修改,其余区域必须像素级保留;
为新衣服生成与原图灯光一致的褶皱与阴影;
保持人物身材与姿势完全不变。变体 2:批量对象替换
📝 提示词
以 REFERENCE_0 为基础,将画面中所有 {argument name="original objects" default="木质椅子"} 替换为 {argument name="replacement objects" default="米色塑胶椅"}。
保持每把椅子的位置、角度与摆放不变;
为新椅子生成与原图光线方向一致的阴影;
不要修改桌子、墙面、灯具、人物。变体 3:自动补全模式
📝 提示词
以 REFERENCE_0 为基础,将原图中 {argument name="object" default="主要前景物体"} 替换为视觉风格更现代的同功能版本,自动决定材质与配色,但保持位置与尺寸一致。避免事项
- 不要替换后改变原对象的位置 / 比例(除非用户允许)
- 不要让替换对象的灯光方向与原图不一致
- 不要顺便修改其它无关元素
- 不要让替换对象的材质显得"贴上去"
- 没有蒙版时,不要假装精确(说明边缘可能略微变化)