Back to Details

paper-illustration-image2

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Paper Illustration Image2

Paper Illustration Image2

Generate publication-quality paper figures using Claude as the planner/reviewer and a local Codex app-server MCP bridge as the raster renderer.
通过将Claude作为规划师/审核者、本地Codex应用服务器MCP桥接工具作为光栅渲染器，生成达到出版级质量的论文插图。

Core Design Philosophy

核心设计理念

text
┌──────────────────────────────────────────────────────────────────────────┐
│                    MULTI-STAGE ITERATIVE WORKFLOW                        │
├──────────────────────────────────────────────────────────────────────────┤
│                                                                          │
│   User Request                                                           │
│       │                                                                  │
│       ▼                                                                  │
│   ┌─────────────┐                                                        │
│   │   Claude    │ ◄─── Step 1: Parse request, create initial prompt     │
│   │  (Planner)  │      - Extract components, labels, and data flow       │
│   │             │      - Write a paper-ready figure brief                │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   ┌─────────────┐                                                        │
│   │Claude/Codex │ ◄─── Step 2: Optimize layout description               │
│   │   Layout    │      - Refine component positioning                    │
│   │   Review    │      - Optimize spacing and grouping                   │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   ┌─────────────┐                                                        │
│   │Claude/Codex │ ◄─── Step 3: CVPR/NeurIPS style verification           │
│   │   Style     │      - Check palette, arrows, and label standards      │
│   │   Check     │      - Tighten the prompt before rendering             │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   ┌─────────────┐                                                        │
│   │ codex-image2│ ◄─── Step 4: Native image generation via bridge        │
│   │ MCP bridge  │      - Call generate_start / generate_status           │
│   │ + app-server│      - Accept only native imageGeneration output       │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   ┌─────────────┐                                                        │
│   │   Claude    │ ◄─── Step 5: STRICT visual review + SCORE (1-10)      │
│   │  (Reviewer) │      - Verify logic, labels, arrows, and aesthetics    │
│   │   STRICT!   │      - Reject unclear or non-paper-ready figures       │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   Score ≥ 9? ──YES──► Accept & Output                                    │
│          │                                                               │
│          NO                                                              │
│          │                                                               │
│          ▼                                                               │
│   Generate SPECIFIC improvement feedback ──► Loop back to Step 2        │
│                                                                          │
└──────────────────────────────────────────────────────────────────────────┘
text
┌──────────────────────────────────────────────────────────────────────────┐
│                    MULTI-STAGE ITERATIVE WORKFLOW                        │
├──────────────────────────────────────────────────────────────────────────┤
│                                                                          │
│   用户请求                                                               │
│       │                                                                  │
│       ▼                                                                  │
│   ┌─────────────┐                                                        │
│   │   Claude    │ ◄─── 步骤1：解析请求，创建初始提示词                  │
│   │  (规划师)   │      - 提取组件、标签和数据流                          │
│   │             │      - 撰写符合论文要求的插图简要说明                  │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   ┌─────────────┐                                                        │
│   │Claude/Codex │ ◄─── 步骤2：优化布局描述                              │
│   │   布局审核  │      - 细化组件定位                                    │
│   │             │      - 优化间距与分组方式                              │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   ┌─────────────┐                                                        │
│   │Claude/Codex │ ◄─── 步骤3：CVPR/NeurIPS风格验证                      │
│   │   风格检查  │      - 检查配色、箭头和标签标准                        │
│   │             │      - 渲染前优化提示词                                │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   ┌─────────────┐                                                        │
│   │ codex-image2│ ◄─── 步骤4：通过桥接工具进行原生图像生成                │
│   │ MCP桥接工具 │      - 调用generate_start / generate_status接口        │
│   │ + app-server│      - 仅接受原生imageGeneration输出                  │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   ┌─────────────┐                                                        │
│   │   Claude    │ ◄─── 步骤5：严格视觉审核 + 评分（1-10分）              │
│   │  (审核者)   │      - 验证逻辑、标签、箭头和美学设计                  │
│   │   严格执行！│      - 拒绝模糊或不符合论文要求的插图                  │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   评分≥9？──是──► 接受并输出                                            │
│          │                                                               │
│          否                                                              │
│          │                                                               │
│          ▼                                                               │
│   生成具体改进反馈 ──► 回到步骤2循环优化                                │
│                                                                          │
└──────────────────────────────────────────────────────────────────────────┘

Constants

常量定义

  • RENDERER = 
    codex-image2
     — Native image generation bridge exposed through local Codex app-server
  • OPTIONAL_TEXT_CRITIC = 
    mcp__codex__codex
     — Optional text-only second opinion for layout/style checks
  • MAX_ITERATIONS = 5 — Maximum refinement rounds
  • TARGET_SCORE = 9 — Minimum acceptable score (1-10)
  • OUTPUT_DIR = 
    figures/ai_generated/
     — Output directory
  • TEXT_LANGUAGE = 
    English
     — Default figure text language unless the user requests otherwise
  • NATIVE_IMAGE_REQUIREMENT = 
    strict
     — Accept only native 
    imageGeneration
    output; reject shell/Python fallbacks
  • CANONICAL_HELPER = 
    python3 tools/paper_illustration_image2.py
     — Preflight, finalize, verify, and repair path for this integration
  • RENDERER = 
    codex-image2
     — 通过本地Codex app-server暴露的原生图像生成桥接工具
  • OPTIONAL_TEXT_CRITIC = 
    mcp__codex__codex
     — 用于布局/风格检查的可选纯文本二次审核工具
  • MAX_ITERATIONS = 5 — 最大优化迭代次数
  • TARGET_SCORE = 9 — 最低可接受评分（1-10分）
  • OUTPUT_DIR = 
    figures/ai_generated/
     — 输出目录
  • TEXT_LANGUAGE = 
    English
     — 默认插图文本语言，除非用户另有要求
  • NATIVE_IMAGE_REQUIREMENT = 
    strict
     — 仅接受原生
    imageGeneration
    输出；拒绝shell/Python替代方案
  • CANONICAL_HELPER = 
    python3 tools/paper_illustration_image2.py
     — 用于该集成的预检、定稿、验证和修复脚本

CVPR/ICLR/NeurIPS Top-Tier Conference Style Guide

CVPR/ICLR/NeurIPS顶级会议风格指南

What "CVPR Style" Actually Means:
“CVPR风格”的实际含义：

Visual Standards

视觉标准

  • Clean white background — No decorative patterns or gradients unless extremely subtle
  • Sans-serif fonts — Arial, Helvetica, or similarly clean paper-friendly typography
  • Subtle color palette — Use 3-5 coordinated colors, not rainbow colors
  • Print-friendly — Must remain understandable in grayscale
  • Professional borders — Thin to medium, clean, and consistent
  • 纯白色背景 — 除非极其淡雅，否则不使用装饰性图案或渐变
  • 无衬线字体 — Arial、Helvetica或类似的适合论文的简洁字体
  • 淡雅配色方案 — 使用3-5种协调颜色，避免彩虹色
  • 适合印刷 — 转为灰度后仍需清晰可读
  • 专业边框 — 粗细适中、干净一致

Layout Standards

布局标准

  • Horizontal flow — Left-to-right is the default for pipelines
  • Clear grouping — Use spacing or subtle grouping boxes for related modules
  • Consistent sizing — Similar components should have similar sizes
  • Balanced whitespace — Avoid both cramped and overly sparse layouts
  • 横向流向 — 流水线默认采用从左到右的布局
  • 清晰分组 — 使用间距或淡雅分组框区分相关模块
  • 尺寸一致 — 同类组件应保持相似尺寸
  • 留白均衡 — 避免过于拥挤或过于稀疏的布局

Arrow Standards (MOST CRITICAL)

箭头标准（最关键）

  • Thick strokes — Arrows must remain visible after paper scaling
  • Clear arrowheads — Large, unmistakable arrowheads
  • Dark colors — Prefer black or dark gray arrows
  • Labeled — Important arrows should show what flows through them
  • No crossings — Reorganize the figure to avoid crossings where possible
  • CORRECT DIRECTION — Arrows must point to the right target
  • 粗线条 — 箭头在论文缩放后仍需清晰可见
  • 清晰箭头头 — 大而明确的箭头头
  • 深色 — 优先使用黑色或深灰色箭头
  • 带标签 — 重要箭头应标注流经内容
  • 避免交叉 — 尽可能调整插图布局以避免箭头交叉
  • 方向正确 — 箭头必须指向正确目标

Visual Appeal (Academic Professional Style)

视觉吸引力（学术专业风格）

目标：既不保守也不花哨，找到平衡点
目标：既不保守也不花哨，找到平衡点

✅ Should have

✅ 建议具备

  • Subtle gradients — Gentle same-family gradients are acceptable
  • Rounded corners — Modern but restrained rounded blocks
  • Clear hierarchy — Main modules larger, secondary modules smaller
  • Consistent color coding — Stable mapping between module types and colors
  • Professional typography — Clean labels with readable size hierarchy
  • 淡雅渐变 — 同色系的柔和渐变是可接受的
  • 圆角 — 现代但克制的圆角模块
  • 清晰层级 — 主模块更大，次要模块更小
  • 一致颜色编码 — 模块类型与颜色保持稳定映射
  • 专业排版 — 清晰的标签与可读的尺寸层级

❌ Avoid

❌ 需避免

  • ❌ Rainbow gradients
  • ❌ Heavy drop shadows
  • ❌ 3D perspective effects
  • ❌ Glowing effects
  • ❌ Decorative clip-art icons
  • ❌ Slide-deck styling that feels flashy rather than paper-ready
  • ❌ 彩虹渐变
  • ❌ 厚重阴影
  • ❌ 3D透视效果
  • ❌ 发光效果
  • ❌ 装饰性剪贴画图标
  • ❌ 过于花哨的幻灯片风格，而非符合论文要求的风格

✓ Ideal effect

✓ 理想效果

  • Looks intentional, professional, and immediately readable
  • Has moderate visual appeal without becoming decorative
  • Feels appropriate for a top-tier conference paper figure
  • Survives PDF scaling and grayscale printing
  • 看起来严谨、专业且易于快速阅读
  • 具备适度视觉吸引力但不过度装饰
  • 适合顶级会议论文插图
  • 经PDF缩放和灰度打印后仍清晰可读

What to AVOID (CRITICAL)

必须避免的问题（关键）

  • ❌ Thin, hairline arrows
  • ❌ Unlabeled or ambiguous connections
  • ❌ Tiny unreadable text
  • ❌ Flat, boring box soup with no hierarchy
  • ❌ Over-decorated figures with shadows/glows/icons
  • ❌ Wrong arrow directions
  • ❌ 纤细的发丝状箭头
  • ❌ 未标注或模糊的连接
  • ❌ 过小无法阅读的文本
  • ❌ 无层级的单调方块堆砌
  • ❌ 带有阴影/发光/图标的过度装饰插图
  • ❌ 箭头方向错误

Scope

适用范围

Figure TypeQualityExamples
Architecture diagramsExcellentModel architecture, pipeline, encoder-decoder
Method illustrationsExcellentConceptual diagrams, algorithm flowcharts
Conceptual figuresGoodComparison diagrams, taxonomy trees
Not for: Statistical plots (use 
/paper-figure
), deterministic vector topology figures (prefer 
/figure-spec
), photo-realistic scenes
插图类型质量表现示例
架构图优秀模型架构、流水线、编码器-解码器
方法示意图优秀概念图、算法流程图
概念性插图良好对比图、分类树
不适用场景： 统计图表（使用
/paper-figure
）、确定性矢量拓扑图（优先使用
/figure-spec
）、照片级写实场景

Workflow: MUST EXECUTE ALL STEPS

工作流程：必须执行所有步骤

Step 0: Pre-flight Check

步骤0：预检检查

Render this checklist explicitly before starting:
text
📋 paper-illustration-image2 integration checklist:
   [ ] 1. python3 tools/paper_illustration_image2.py preflight --workspace <cwd> --json-out figures/ai_generated/preflight.json
   [ ] 2. Confirm preflight JSON says ok=true before rendering
   [ ] 3. Render via mcp__codex-image2__generate_start + generate_status
   [ ] 4. Finalize via python3 tools/paper_illustration_image2.py finalize --workspace <cwd> --best-image <best_png>
   [ ] 5. Verify artifacts via python3 tools/paper_illustration_image2.py verify --workspace <cwd> --json-out figures/ai_generated/verify.json
  1. Create 
    figures/ai_generated/
    if it does not exist.
  2. Confirm the request is suitable for a raster illustration:
    • architecture diagram
    • conceptual method figure
    • workflow illustration
  3. Prefer English figure text unless the user asked otherwise.
  4. Run:
bash
python3 tools/paper_illustration_image2.py preflight \
  --workspace <cwd> \
  --json-out figures/ai_generated/preflight.json
  1. If preflight is not 
    ok=true
    , stop and say so clearly.
开始前需明确列出以下检查清单：
text
📋 paper-illustration-image2集成检查清单:
   [ ] 1. python3 tools/paper_illustration_image2.py preflight --workspace <cwd> --json-out figures/ai_generated/preflight.json
   [ ] 2. 确认预检JSON显示ok=true后再开始渲染
   [ ] 3. 通过mcp__codex-image2__generate_start + generate_status进行渲染
   [ ] 4. 通过python3 tools/paper_illustration_image2.py finalize --workspace <cwd> --best-image <best_png>完成定稿
   [ ] 5. 通过python3 tools/paper_illustration_image2.py verify --workspace <cwd> --json-out figures/ai_generated/verify.json验证产物
  1. 若
    figures/ai_generated/
    目录不存在则创建该目录。
  2. 确认请求适合光栅插图：
    • 架构图
    • 概念性方法图
    • 工作流示意图
  3. 优先使用英文插图文本，除非用户另有要求。
  4. 运行以下命令：
bash
python3 tools/paper_illustration_image2.py preflight \
  --workspace <cwd> \
  --json-out figures/ai_generated/preflight.json
  1. 若预检结果不是
    ok=true
    ，则停止操作并明确告知用户。

Step 1: Claude Plans the Figure

步骤1：Claude规划插图

Turn the user request into a fully specified image prompt. Include:
  • figure type
  • exact modules / stages
  • flow direction
  • labels to show
  • data-flow arrows
  • style constraints
  • what to avoid
When the input is a method note or a paper section, summarize it first into a clean figure brief before writing the final image prompt.
将用户请求转化为完全指定的图像提示词，包含：
  • 插图类型
  • 具体模块/阶段
  • 流向
  • 需显示的标签
  • 数据流箭头
  • 风格约束
  • 需避免的内容
当输入是方法说明或论文章节时，先将其总结为清晰的插图简要说明，再撰写最终图像提示词。

Step 2: Layout Optimization

步骤2：布局优化

This step is required. Before rendering, refine the prompt into a concrete layout plan:
  • exact module order
  • spacing and grouping
  • relative module prominence
  • arrow routing and likely collision points
If 
mcp__codex__codex
is available, you may ask it for a short second-opinion layout critique here, but Claude should still complete this step even without Codex.
Use Codex layout critique for:
  • missing components
  • confusing layout
  • weak flow hierarchy
  • likely arrow-direction ambiguity or clutter
此步骤为必填项。渲染前需将提示词细化为具体的布局方案：
  • 具体模块顺序
  • 间距与分组
  • 模块相对重要性
  • 箭头路径与可能的碰撞点
若
mcp__codex__codex
可用，可在此步骤请求其提供简短的布局审核二次意见，但即使没有Codex，Claude仍需完成此步骤。
使用Codex布局审核检查：
  • 缺失的组件
  • 易混淆的布局
  • 薄弱的流向层级
  • 可能存在的箭头方向模糊或杂乱问题

Step 3: Style Verification

步骤3：风格验证

This step is also required. Check the prompt against the intended paper style before rendering:
  • palette is restrained and academic
  • arrows are thick, dark, and readable
  • labels are concise and in English unless requested otherwise
  • the figure will read clearly in grayscale / print
  • no glow, rainbow gradient, or slide-deck decoration slips in
If 
mcp__codex__codex
is available, you may ask it for a short text-only style audit, but do not block on it.
此步骤同样为必填项。渲染前需检查提示词是否符合目标论文风格：
  • 配色克制且符合学术规范
  • 箭头粗壮、深色且可读
  • 标签简洁，除非另有要求否则使用英文
  • 插图转为灰度/印刷后仍清晰可读
  • 未混入发光、彩虹渐变或幻灯片风格装饰
若
mcp__codex__codex
可用，可请求其提供简短的纯文本风格审核，但无需等待其结果。

Step 4: Generate Through the Bridge

步骤4：通过桥接工具生成图像

Call 
mcp__codex-image2__generate_start
with:
  • prompt
    : the final image prompt
  • cwd
    : current project root or paper workspace
  • outputPath
    : 
    figures/ai_generated/figure_v1.png
  • system
    : a short instruction like 
    Academic paper figure. Prefer crisp English labels.
  • timeoutSeconds
    : a bounded render timeout such as 
    180
Then call 
mcp__codex-image2__generate_status
with bounded waits until:
  • done=true
    and 
    status=completed
    , or
  • done=true
    and 
    status=failed
If generation fails, report the bridge error directly instead of hiding it.
调用
mcp__codex-image2__generate_start
，参数包括：
  • prompt
    : 最终图像提示词
  • cwd
    : 当前项目根目录或论文工作区
  • outputPath
    : 
    figures/ai_generated/figure_v1.png
  • system
    : 简短指令，如
    Academic paper figure. Prefer crisp English labels.
  • timeoutSeconds
    : 有限的渲染超时时间，如
    180
然后调用
mcp__codex-image2__generate_status
并等待，直到：
  • done=true
    且
    status=completed
    ，或
  • done=true
    且
    status=failed
若生成失败，直接报告桥接工具错误，不得隐瞒。

Step 5: Review the Output

步骤5：审核输出结果

Review the generated image with a strict checklist:
  • are all major components present?
  • is the logical flow obvious?
  • are labels readable?
  • do arrows point the right way?
  • does the figure look paper-ready rather than like a slide?
Score it from 1-10.
使用严格的检查清单审核生成的图像：
  • 所有主要组件是否齐全？
  • 逻辑流向是否清晰？
  • 标签是否可读？
  • 箭头方向是否正确？
  • 插图是否符合论文要求而非幻灯片风格？
为插图评分（1-10分）。

Step 6: Refine if Needed

步骤6：必要时优化

If score < 9, write a targeted refinement prompt:
  • say exactly what was wrong
  • say what to preserve
  • regenerate to 
    figure_v2.png
    , 
    figure_v3.png
    , etc.
Keep refinement feedback concrete:
  • Increase spacing between genome scan and scoring modules
  • Make the off-target branch thinner and secondary
  • Use cleaner English labels: "Candidate sgRNA library", not "sgRNA library 23 bp"
若评分<9分，撰写针对性的优化提示词：
  • 明确说明存在的问题
  • 说明需要保留的内容
  • 重新生成至
    figure_v2.png
    、
    figure_v3.png
    等版本
优化反馈需具体明确：
  • Increase spacing between genome scan and scoring modules
  • Make the off-target branch thinner and secondary
  • Use cleaner English labels: "Candidate sgRNA library", not "sgRNA library 23 bp"

Step 7: Finalize And Verify

步骤7：定稿与验证

When accepted:
  • run the canonical helper to promote the best image to 
    figure_final.png
  • let the helper write 
    latex_include.tex
  • let the helper write 
    review_log.json
  • run helper verification before claiming success
bash
python3 tools/paper_illustration_image2.py finalize \
  --workspace <cwd> \
  --best-image figures/ai_generated/figure_vN.png \
  --score 9 \
  --review-summary "Accepted after strict review; labels and arrows are paper-ready."

python3 tools/paper_illustration_image2.py verify \
  --workspace <cwd> \
  --json-out figures/ai_generated/verify.json
Suggested LaTeX:
latex
\begin{figure*}[t]
    \centering
    \includegraphics[width=0.95\textwidth]{figures/ai_generated/figure_final.png}
    \caption{[Replace with a paper-ready caption].}
    \label{fig:[replace-me]}
\end{figure*}
当插图被接受后：
  • 运行标准辅助脚本将最佳图像命名为
    figure_final.png
  • 让辅助脚本生成
    latex_include.tex
  • 让辅助脚本生成
    review_log.json
  • 在宣告成功前运行辅助脚本验证
bash
python3 tools/paper_illustration_image2.py finalize \
  --workspace <cwd> \
  --best-image figures/ai_generated/figure_vN.png \
  --score 9 \
  --review-summary "Accepted after strict review; labels and arrows are paper-ready."

python3 tools/paper_illustration_image2.py verify \
  --workspace <cwd> \
  --json-out figures/ai_generated/verify.json
推荐的LaTeX代码：
latex
\begin{figure*}[t]
    \centering
    \includegraphics[width=0.95\textwidth]{figures/ai_generated/figure_final.png}
    \caption{[Replace with a paper-ready caption].}
    \label{fig:[replace-me]}
\end{figure*}

Key Rules

核心规则

  1. Never skip Step 2 or Step 3; layout and style checks are required.
  2. Never skip the final visual review.
  3. Never accept a figure that is logically wrong just because it looks attractive.
  4. Use the 
    codex-image2
    bridge only for native image generation.
  5. If the bridge says native image generation is unavailable, surface that honestly.
  6. Reject any shell/Python/manual bitmap fallback masquerading as image generation.
  7. Keep figure text in English unless the user requested another language.
  8. Prefer 1-3 strong refinement rounds over many shallow ones.
  9. Use specific, actionable refinement feedback instead of vague comments.
  10. Review arrow direction, label clarity, and visual hierarchy every round.
  11. Accept only figures that look paper-ready, not slide-ready.
  12. Always use 
    tools/paper_illustration_image2.py finalize
    to emit the final artifacts.
  13. Always use 
    tools/paper_illustration_image2.py verify
    before claiming success.
  1. 不得跳过步骤2或步骤3；布局与风格检查为必填项。
  2. 不得跳过最终视觉审核。
  3. 不得因插图美观而接受逻辑错误的内容。
  4. 仅将
    codex-image2
    桥接工具用于原生图像生成。
  5. 若桥接工具提示原生图像生成不可用，需如实告知用户。
  6. 拒绝任何伪装成图像生成的shell/Python/手动位图替代方案。
  7. 除非用户要求，否则插图文本使用英文。
  8. 优先进行1-3轮针对性强的优化，而非多次浅层次优化。
  9. 使用具体、可执行的优化反馈，而非模糊评论。
  10. 每一轮都需检查箭头方向、标签清晰度和视觉层级。
  11. 仅接受符合论文要求的插图，而非幻灯片风格的插图。
  12. 必须使用
    tools/paper_illustration_image2.py finalize
    生成最终产物。
  13. 宣告成功前必须使用
    tools/paper_illustration_image2.py verify
    进行验证。

Repair Path

修复流程

If rendering succeeded but final artifacts were skipped, repair the integration explicitly:
bash
python3 tools/paper_illustration_image2.py finalize \
  --workspace <cwd> \
  --best-image figures/ai_generated/figure_vN.png

python3 tools/paper_illustration_image2.py verify \
  --workspace <cwd> \
  --json-out figures/ai_generated/verify.json
若渲染成功但跳过了最终产物生成，需明确执行以下集成修复：
bash
python3 tools/paper_illustration_image2.py finalize \
  --workspace <cwd> \
  --best-image figures/ai_generated/figure_vN.png

python3 tools/paper_illustration_image2.py verify \
  --workspace <cwd> \
  --json-out figures/ai_generated/verify.json

Output Structure

输出结构

text
figures/ai_generated/
├── preflight.json         # Helper preflight receipt
├── figure_v1.png          # Iteration 1
├── figure_v2.png          # Iteration 2
├── figure_v3.png          # Iteration 3
├── figure_final.png       # Accepted version (copy of best, score ≥ 9)
├── latex_include.tex      # LaTeX snippet
├── review_log.json        # Review notes and refinement history
└── verify.json            # Helper verification diagnostic
text
figures/ai_generated/
├── preflight.json         # 辅助预检回执
├── figure_v1.png          # 迭代版本1
├── figure_v2.png          # 迭代版本2
├── figure_v3.png          # 迭代版本3
├── figure_final.png       # 接受版本（最佳版本副本，评分≥9）
├── latex_include.tex      # LaTeX代码片段
├── review_log.json        # 审核记录与优化历史
└── verify.json            # 辅助验证诊断文件

Model Summary

模型总结

StageAgent / ToolPurpose
Step 0
python3 tools/paper_illustration_image2.py preflight
Observable activation predicate and preflight receipt
Step 1ClaudeParse request and create the initial figure prompt
Step 2Claude (+ optional Codex critique)Refine layout, grouping, spacing, and arrow routing
Step 3Claude (+ optional Codex critique)Verify academic visual style before rendering
Step 4
mcp__codex-image2__generate_start
+ 
generate_status
Native raster image generation through Codex app-server
Step 5ClaudeStrict visual review and scoring
Step 7
python3 tools/paper_illustration_image2.py finalize
+ 
verify
Emit canonical artifacts and external verification receipt
阶段智能体/工具用途
步骤0
python3 tools/paper_illustration_image2.py preflight
可观测的激活条件与预检回执
步骤1Claude解析请求并创建初始插图提示词
步骤2Claude（可选Codex审核）优化布局、分组、间距与箭头路径
步骤3Claude（可选Codex审核）渲染前验证学术视觉风格
步骤4
mcp__codex-image2__generate_start
+ 
generate_status
通过Codex app-server进行原生光栅图像生成
步骤5Claude严格视觉审核与评分
步骤7
python3 tools/paper_illustration_image2.py finalize
+ 
verify
生成标准产物与外部验证回执