paper-illustration

name	paper-illustration
description	Generate publication-quality AI illustrations for academic papers via pluggable image backends (Gemini / OpenAI gpt-image / any OpenAI-compatible endpoint, with mermaid fallback). Creates architecture diagrams, method illustrations with Claude-supervised iterative refinement loop. Use when user says "生成图表", "画架构图", "AI绘图", "paper illustration", "generate diagram", or needs visual figures for papers.
argument-hint	[description-or-method-file] [— style-ref: <source>]
allowed-tools	Bash(*), Read, Write, Edit, Grep, Glob, mcp__codex__codex, mcp__codex__codex-reply, WebSearch

Paper Illustration: Multi-Stage Claude-Supervised Figure Generation

Generate publication-quality illustrations using a multi-stage workflow with Claude as the STRICT supervisor/reviewer.

Core Design Philosophy

┌──────────────────────────────────────────────────────────────────────────┐
│                    MULTI-STAGE ITERATIVE WORKFLOW                        │
├──────────────────────────────────────────────────────────────────────────┤
│                                                                          │
│   User Request                                                           │
│       │                                                                  │
│       ▼                                                                  │
│   ┌─────────────┐                                                        │
│   │   Claude    │ ◄─── Step 1: Parse request, create initial prompt     │
│   │  (Planner)  │                                                        │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   ┌─────────────┐                                                        │
│   │   Gemini    │ ◄─── Step 2: Optimize layout description               │
│   │ (REASONING_MODEL)│   - Refine component positioning                    │
│   │  Layout     │      - Optimize spacing and grouping                   │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   ┌─────────────┐                                                        │
│   │   Gemini    │ ◄─── Step 3: CVPR/NeurIPS style verification          │
│   │ (REASONING_MODEL)│   - Check color palette compliance                  │
│   │  Style      │      - Verify arrow and font standards                 │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   ┌─────────────┐                                                        │
│   │ Paperbanana │ ◄─── Step 4: Render final image                       │
│   │ (gemini-3-  │      - High-quality image generation                   │
│   │ pro-image)  │      - Internal codename: Nano Banana Pro              │
│   └──────┬──────┘                                                        │
│          │                                                               │
│          ▼                                                               │
│   ┌─────────────┐                                                        │
│   │   Claude    │ ◄─── Step 5: STRICT visual review + SCORE (1-10)      │
│   │  (Reviewer) │      - Verify EVERY arrow direction                    │
│   │   STRICT!   │      - Verify EVERY block content                      │
│   └──────┬──────┘      - Verify aesthetics & visual appeal               │
│          │                                                               │
│          ▼                                                               │
│   Score ≥ 9? ──YES──► Accept & Output                                    │
│          │                                                               │
│          NO                                                              │
│          │                                                               │
│          ▼                                                               │
│   Generate SPECIFIC improvement feedback ──► Loop back to Step 2        │
│                                                                          │
└──────────────────────────────────────────────────────────────────────────┘

Constants

• IMAGE_BACKEND = ${IMAGE_BACKEND:-gemini} — pluggable image backend: gemini | openai | openai-compatible（见下方后端选择）
• IMAGE_MODEL = ${IMAGE_MODEL:-gemini-3-pro-image-preview} — image rendering model; override with IMAGE_MODEL（openai 后端默认 gpt-image-1）.
• REASONING_MODEL = ${REASONING_MODEL:-gemini-3-pro-preview} — default for layout optimization and style checking; override with REASONING_MODEL. Layout/style reasoning 不依赖出图后端：无 Gemini key 时直接由当前主 agent（Claude/GPT/…）在对话内完成同样的 layout 优化与 style 核验步骤即可，只有出图渲染必须走外部图像 API。
• MAX_ITERATIONS = 5 — Maximum refinement rounds
• TARGET_SCORE = 9 — Minimum acceptable score (1-10) — RAISED FOR QUALITY
• OUTPUT_DIR = figures/ai_generated/ — Output directory

出图后端选择（按可用凭据自动降级）

优先级	后端	条件	默认模型
1	gemini	GEMINI_API_KEY 已设	gemini-3-pro-image-preview
2	openai	OPENAI_API_KEY 已设	gpt-image-1（Images API /v1/images/generations）
3	openai-compatible	IMAGE_API_BASE + IMAGE_API_KEY 已设（任何 OpenAI Images 兼容端点：Azure OpenAI、自建网关、其他厂商兼容层）	由 IMAGE_MODEL 指定
4	fallback	以上均无	mermaid-diagram / data-visualization 生成示意图/矢量占位图，标注 illustration generated via fallback

用户可用 IMAGE_BACKEND 显式指定后端跳过自动探测。所有后端共用同一 prompt 流水线（Step 1-3 的 plan → layout → style spec 一字不改），只替换 Step 4 的渲染调用。

Optional: Style reference (— style-ref: <source>, opt-in)

Lets the user steer structural figure conventions (caption length, panel-count distribution, figure-to-table ratio in the parent paper) toward a reference paper. Default OFF — when the user does not pass — style-ref, do nothing differently from before.

Only when — style-ref: <source> appears in $ARGUMENTS, run the helper FIRST, before generating prompts:

# Resolve $STYLE_HELPER via the canonical strict-safe chain (see
# shared-references/integration-contract.md §2). Policy A — gate:
# unresolved helper means --style-ref cannot be satisfied, so abort.
cd "$(git rev-parse --show-toplevel 2>/dev/null || pwd)" || exit 1
if [ -z "${ARIS_REPO:-}" ] && [ -f .aris/installed-skills.txt ]; then
    ARIS_REPO=$(awk -F'\t' '$1=="repo_root"{print $2; exit}' .aris/installed-skills.txt 2>/dev/null) || true
fi
if [ -z "${ARIS_REPO:-}" ] && [ -f "$HOME/.aris/repo" ]; then
    ARIS_REPO=$(cat "$HOME/.aris/repo" 2>/dev/null) || true
fi
STYLE_HELPER=".aris/tools/extract_paper_style.py"
[ -f "$STYLE_HELPER" ] || STYLE_HELPER="tools/extract_paper_style.py"
[ -f "$STYLE_HELPER" ] || { [ -n "${ARIS_REPO:-}" ] && STYLE_HELPER="$ARIS_REPO/tools/extract_paper_style.py"; }
[ -f "$STYLE_HELPER" ] || {
  echo "ERROR: extract_paper_style.py not resolved at .aris/tools/, tools/, \$ARIS_REPO/tools/, or via ~/.aris/repo." >&2
  echo "       Fix: run 'bash install.sh' at the pack root (writes ~/.aris/repo), export ARIS_REPO=<pack-root>, or copy the helper to tools/." >&2
  echo "       --style-ref cannot be satisfied; aborting." >&2
  exit 1
}
STYLE_STATUS=0
CACHE=$(python3 "$STYLE_HELPER" --source "<source>") || STYLE_STATUS=$?
case "$STYLE_STATUS" in
  0) ;;                                       # use $CACHE/style_profile.md as structural guidance
  2) echo "warning: style-ref skipped (missing optional dep)" >&2 ;;
  3) echo "error: --style-ref source failed; aborting illustration" >&2 ; exit 1 ;;
  *) echo "error: helper failed unexpectedly; aborting illustration" >&2 ; exit 1 ;;
esac

Sources accepted: local TeX dir / file, local PDF, arXiv id, http(s) URL. Overleaf URLs/IDs are rejected — clone via /overleaf-sync setup <id> first and pass the local clone path.

Strict rules (full contract in tools/extract_paper_style.py docstring):

• Use style_profile.md to align caption length and figure density with the reference paper. The CVPR/ICLR/NeurIPS visual standards above still take precedence — --style-ref only refines length-and-density tendencies, never image content.
• Never copy figure content, color palettes, or specific design elements from anything reachable through the cache. The visual design comes from the user's prompt, not the reference.
• Never pass — style-ref (or the cache contents) to the Claude vision-checker / Gemini reasoning-checker sub-agents when they score the generated image — the image must be judged on its own merits.

CVPR/ICLR/NeurIPS Top-Tier Conference Style Guide

What "CVPR Style" Actually Means:

Visual Standards

• Clean white background — No decorative patterns or gradients (unless subtle)
• Sans-serif fonts — Arial, Helvetica, or Computer Modern; minimum 14pt
• Subtle color palette — Not rainbow colors; use 3-5 coordinated colors
• Print-friendly — Must be readable in grayscale (many reviewers print papers)
• Professional borders — Thin (2-3px), solid colors, not flashy

Layout Standards

• Horizontal flow — Left-to-right is the standard for pipelines
• Clear grouping — Use subtle background boxes to group related modules
• Consistent sizing — Similar components should have similar sizes
• Balanced whitespace — Not cramped, not sparse

Arrow Standards (MOST CRITICAL)

• Thick strokes — 4-6px minimum (thin arrows disappear when printed)
• Clear arrowheads — Large, filled triangular heads
• Dark colors — Black or dark gray (#333333); avoid colored arrows
• Labeled — Every arrow should indicate what data flows through it
• No crossings — Reorganize layout to avoid arrow crossings
• CORRECT DIRECTION — Arrows must point to the RIGHT target!

Visual Appeal (科研风格 - Professional Academic Style)

目标：既不保守也不花哨，找到平衡点

✅ 应该有的视觉元素：

• Subtle gradient fills — 淡雅的渐变填充（同色系从浅到深），不是炫彩
• Rounded corners — 圆角矩形（6-10px radius），现代感但不夸张
• Clear visual hierarchy — 通过大小、颜色深浅区分层次
• Consistent color coding — 统一的配色方案（3-4种主色）
• Internal structure — 大模块内部显示子组件（如Encoder内部的layer结构）
• Professional typography — 清晰的标签，适当的字号层次

✅ 配色建议（学术专业）：

• Inputs: 柔和的绿色系 (#10B981 / #34D399)
• Encoders: 专业的蓝色系 (#2563EB / #3B82F6)
• Fusion: 优雅的紫色系 (#7C3AED / #8B5CF6)
• Outputs: 温暖的橙色系 (#EA580C / #F97316)
• Arrows: 黑色或深灰 (#333333 / #1F2937)
• Background: 纯白 (#FFFFFF)，不要花纹

❌ 要避免的过度装饰：

• ❌ Rainbow color schemes (彩虹配色)
• ❌ Heavy drop shadows (重阴影效果)
• ❌ 3D effects / perspective (3D透视)
• ❌ Excessive gradients (夸张的多色渐变)
• ❌ Clip art / cartoon icons (卡通图标)
• ❌ Decorative patterns in background (背景花纹)
• ❌ Glowing effects (发光效果)
• ❌ Too many small icons (过多小图标)

✓ 理想的视觉效果：

• 一眼看上去专业、清晰
• 有适度的视觉吸引力，但不抢眼
• 符合CVPR/NeurIPS论文的审美标准
• 打印友好（灰度模式下也能清晰辨认）
• 像精心设计的学术图表，而不是PPT模板

What to AVOID (CRITICAL)

• ❌ Rainbow color schemes (too many colors)
• ❌ Thin, hairline arrows (arrows must be THICK)
• ❌ Unlabeled connections
• ❌ Plain boring rectangles (add some visual interest)
• ❌ Over-decorated with shadows/glows/icons (too flashy)
• ❌ Small text that's unreadable when printed
• ❌ WRONG arrow directions — This is UNACCEPTABLE!

Scope

Figure Type	Quality	Examples
Architecture diagrams	Excellent	Model architecture, pipeline, encoder-decoder
Method illustrations	Excellent	Conceptual diagrams, algorithm flowcharts
Conceptual figures	Good	Comparison diagrams, taxonomy trees

Not for: Statistical plots (use /paper-figure), photo-realistic images

Workflow: MUST EXECUTE ALL STEPS

Step 0: Pre-flight Check

# Resolve image backend (see Constants → 出图后端选择)
IMAGE_BACKEND="${IMAGE_BACKEND:-}"
if [ -z "$IMAGE_BACKEND" ]; then
    if   [ -n "$GEMINI_API_KEY" ]; then IMAGE_BACKEND=gemini
    elif [ -n "$OPENAI_API_KEY" ]; then IMAGE_BACKEND=openai
    elif [ -n "$IMAGE_API_BASE" ] && [ -n "$IMAGE_API_KEY" ]; then IMAGE_BACKEND=openai-compatible
    else
        echo "WARN: no image API credentials (GEMINI_API_KEY / OPENAI_API_KEY / IMAGE_API_BASE+IMAGE_API_KEY)."
        echo "      Falling back to mermaid-diagram / data-visualization (vector schematic)."
        IMAGE_BACKEND=fallback
    fi
fi
echo "Image backend: $IMAGE_BACKEND"

# Create output directory
mkdir -p figures/ai_generated

Step 1: Claude Plans the Figure (YOU ARE HERE)

CRITICAL: Claude must first analyze the user's request and create a detailed prompt.

Parse the input: $ARGUMENTS

Claude's task:

1. Understand what figure the user wants
2. Identify all components, connections, data flow
3. Create a detailed, structured prompt for Gemini
4. Include style requirements AND visual appeal requirements

Prompt Template for Claude to generate:

Create a PROFESSIONAL, VISUALLY APPEALING publication-quality academic diagram following CVPR/ICLR/NeurIPS standards.

## Visual Style: 科研风格 (Academic Professional Style)
### 目标：平衡 — 既不保守也不花哨

#### DO (应该有):
- **Subtle gradients** — 同色系淡雅渐变（如 #2563EB → #3B82F6），不是多色炫彩
- **Rounded corners** — 圆角矩形（6-10px），现代感
- **Clear visual hierarchy** — 通过大小、深浅区分层次
- **Internal structure** — 大模块内显示子组件结构
- **Consistent color coding** — 统一的3-4色方案
- **Professional polish** — 精致但不夸张

#### DON'T (不要有):
- ❌ Rainbow/multi-color gradients (彩虹渐变)
- ❌ Heavy drop shadows (重阴影)
- ❌ 3D effects / perspective (3D效果)
- ❌ Glowing effects (发光效果)
- ❌ Excessive decorative icons (过多装饰图标)
- ❌ Plain boring rectangles (完全平淡的方块)

#### 理想效果：
像顶会论文中精心设计的架构图 — 专业、清晰、有适度的视觉吸引力

## Figure Type
[Architecture Diagram / Pipeline / Comparison / etc.]

## Components to Include (BE SPECIFIC ABOUT CONTENT)
1. [Component 1]:
   - Label: "[exact text]"
   - Sub-label: "[smaller text below]"
   - Position: [left/center/right, top/middle/bottom]
   - Style: [border color, fill, internal structure]
2. [Component 2]: ...

## Layout
- Direction: [left-to-right / top-to-bottom]
- Spacing: [tight / normal / loose]
- Grouping: [how components should be grouped]

## Connections (BE EXPLICIT ABOUT DIRECTION)
EXACT arrow specifications:
1. [Component A] → [Component B]: Arrow goes FROM A TO B, label it "[data type]"
2. [Component C] → [Component D]: Arrow goes FROM C TO D, label it "[data type]"
...
VERIFY: Each arrow must point to the CORRECT target!

## Style Requirements (CVPR/ICLR/NeurIPS Standard)

### Visual Style
- Color palette: Professional academic colors
  - Inputs: Green (#10B981)
  - Encoders: Blue (#2563EB)
  - Fusion modules: Purple (#7C3AED)
  - Outputs: Orange (#EA580C)
- Font: Sans-serif (Arial/Helvetica), minimum 14pt, bold for labels
- Background: Clean white, no patterns
- Blocks: Rounded rectangles (8-12px radius), subtle gradient fill, colored border (2-3px)
- Subtle shadows for depth effect
- Print-friendly (must work in grayscale)

### CRITICAL: Arrow & Data Flow Requirements
1. **ALL arrows must be VERY THICK** - minimum 5-6px stroke width
2. **ALL arrows must have CLEAR arrowheads** - large, visible triangular heads
3. **ALL arrows must be BLACK or DARK GRAY** - not colored
4. **Label EVERY arrow** with what data flows through it
5. **VERIFY arrow direction** - each arrow MUST point to the correct target
6. **No ambiguous connections** - every arrow should have a clear source and destination

### Logic Clarity Requirements
1. **Data flow must be immediately obvious** - viewer should understand the pipeline in 5 seconds
2. **No crossing arrows** - reorganize layout to avoid arrow crossings
3. **Consistent direction** - maintain left-to-right or top-to-bottom flow throughout
4. **Group related components** - use subtle background boxes or spacing to group modules
5. **Clear hierarchy** - main components larger, sub-components smaller

## Additional Requirements
[Any specific requirements from user]

Step 2: Layout Optimization (REASONING_MODEL)

Claude sends the initial prompt to the configured REASONING_MODEL for layout optimization.

#!/bin/bash
# Step 2: Optimize layout using the configured reasoning model
# This step refines component positioning and spacing

set -e

OUTPUT_DIR="figures/ai_generated"
mkdir -p "$OUTPUT_DIR"

API_KEY="${GEMINI_API_KEY}"
URL="https://generativelanguage.googleapis.com/v1beta/models/${REASONING_MODEL:-gemini-3-pro-preview}:generateContent?key=$API_KEY"

# The initial prompt from Claude
INITIAL_PROMPT='[Claude fills in the detailed prompt here]'

# Layout optimization request
LAYOUT_REQUEST="You are an expert in academic figure layout design for CVPR/NeurIPS papers.

Analyze this figure request and provide an OPTIMIZED LAYOUT DESCRIPTION:

$INITIAL_PROMPT

Provide:
1. **Optimized Component Positions**: Exact positions (left/center/right, top/middle/bottom) for each component
2. **Spacing Recommendations**: Specific spacing between components
3. **Grouping Strategy**: Which components should be visually grouped together
4. **Arrow Routing**: Optimal paths for arrows to avoid crossings
5. **Visual Hierarchy**: Size recommendations for main vs sub-components

Output a DETAILED layout specification that will be used for rendering."

# Build JSON payload
python3 << PYTHON
import json
payload = {
    "contents": [{"parts": [{"text": '''$LAYOUT_REQUEST'''}]}]
}
with open("/tmp/gemini_layout_request.json", "w") as f:
    json.dump(payload, f, indent=2)
print("Layout request created")
PYTHON

# Call Gemini gemini-3-pro-preview for layout optimization (DIRECT connection, no proxy)
RESPONSE=$(curl -s --max-time 90 \
  -X POST "$URL" \
  -H 'Content-Type: application/json' \
  -d @/tmp/gemini_layout_request.json)

# Extract layout description
LAYOUT_DESCRIPTION=$(echo "$RESPONSE" | python3 -c "
import sys, json
data = json.load(sys.stdin)
try:
    print(data['candidates'][0]['content']['parts'][0]['text'])
except:
    print('Error extracting layout')
")

echo "=== Layout Optimization Complete ==="
echo "$LAYOUT_DESCRIPTION"
echo "$LAYOUT_DESCRIPTION" > "$OUTPUT_DIR/layout_description.txt"

Step 3: Style Verification (REASONING_MODEL)

Claude sends the optimized layout to Gemini for CVPR/NeurIPS style verification.

#!/bin/bash
# Step 3: Verify and enhance style compliance using the configured reasoning model

API_KEY="${GEMINI_API_KEY}"
URL="https://generativelanguage.googleapis.com/v1beta/models/${REASONING_MODEL:-gemini-3-pro-preview}:generateContent?key=$API_KEY"

# Read layout from previous step
LAYOUT=$(cat figures/ai_generated/layout_description.txt)

# Style verification request
STYLE_REQUEST="You are a CVPR/NeurIPS paper figure reviewer specializing in visual standards.

Review and ENHANCE this figure specification for top-tier conference compliance:

$LAYOUT

Ensure compliance with:
1. **Color Palette**: Use professional academic colors (green for inputs, blue for encoders, purple for fusion, orange for outputs)
2. **Arrow Standards**: Thick (5-6px), black/dark gray, clear arrowheads, all labeled
3. **Font Standards**: Sans-serif, minimum 14pt, readable in print
4. **Visual Appeal (科研风格)**:
   - ✅ Subtle same-color gradients, rounded corners (6-10px), internal structure visible
   - ❌ NO heavy shadows, NO glowing effects, NO rainbow gradients

Output an ENHANCED figure specification with explicit style instructions for rendering."

# Build JSON payload
python3 << PYTHON
import json
payload = {
    "contents": [{"parts": [{"text": '''$STYLE_REQUEST'''}]}]
}
with open("/tmp/gemini_style_request.json", "w") as f:
    json.dump(payload, f, indent=2)
print("Style request created")
PYTHON

# Call Gemini gemini-3-pro-preview for style verification (DIRECT connection, no proxy)
RESPONSE=$(curl -s --max-time 90 \
  -X POST "$URL" \
  -H 'Content-Type: application/json' \
  -d @/tmp/gemini_style_request.json)

# Extract style-enhanced specification
STYLE_SPEC=$(echo "$RESPONSE" | python3 -c "
import sys, json
data = json.load(sys.stdin)
try:
    print(data['candidates'][0]['content']['parts'][0]['text'])
except:
    print('Error extracting style spec')
")

echo "=== Style Verification Complete ==="
echo "$STYLE_SPEC"
echo "$STYLE_SPEC" > "figures/ai_generated/style_spec.txt"

Step 4: Paperbanana Image Rendering (IMAGE_MODEL)

Claude sends the optimized, style-verified specification to Paperbanana for rendering.

#!/bin/bash
# Step 4: Render image using Paperbanana (`IMAGE_MODEL`)
# Internal codename: Nano Banana Pro
# Use DIRECT connection (no proxy) - proxy causes SSL errors

set -e

OUTPUT_DIR="figures/ai_generated"
mkdir -p "$OUTPUT_DIR"

# ── Backend: gemini（IMAGE_BACKEND=gemini 时执行；openai / openai-compatible 见本步末尾）──
API_KEY="${GEMINI_API_KEY}"
URL="https://generativelanguage.googleapis.com/v1beta/models/${IMAGE_MODEL:-gemini-3-pro-image-preview}:generateContent?key=$API_KEY"

# Read the style-enhanced specification from previous step
STYLE_SPEC=$(cat figures/ai_generated/style_spec.txt)

# Add rendering instructions
RENDER_PROMPT="Render a publication-quality academic diagram based on this specification:

$STYLE_SPEC

RENDERING REQUIREMENTS:
- Output a clean, professional diagram suitable for CVPR/NeurIPS submission
- Use vector-quality rendering with sharp edges and clear text
- Ensure all elements are properly aligned and spaced
- The diagram should be immediately understandable at a glance"

# Build JSON payload using Python for proper escaping
python3 << PYTHON
import json
payload = {
    "contents": [{"parts": [{"text": '''$RENDER_PROMPT'''}]}],
    "generationConfig": {"responseModalities": ["TEXT", "IMAGE"]}
}
with open("/tmp/gemini_request.json", "w") as f:
    json.dump(payload, f, indent=2)
print("JSON payload created")
PYTHON

# Call Paperbanana API WITHOUT proxy (direct connection works better)
RESPONSE=$(curl -s --max-time 180 \
  -X POST "$URL" \
  -H 'Content-Type: application/json' \
  -d @/tmp/gemini_request.json)

# Check for error
if echo "$RESPONSE" | grep -q '"error"'; then
    echo "API Error:"
    echo "$RESPONSE" | python3 -m json.tool 2>/dev/null || echo "$RESPONSE"
    exit 1
fi

# Extract and save image
echo "$RESPONSE" | python3 << 'PYTHON'
import sys, json, base64
from pathlib import Path

output_dir = Path("figures/ai_generated")
data = json.load(sys.stdin)

try:
    parts = data['candidates'][0]['content']['parts']
    iteration = 1  # Claude increments this each iteration

    for part in parts:
        if 'text' in part:
            print(f"\n[Paperbanana]: {part['text'][:200]}...")
        elif 'inlineData' in part:
            img_data = base64.b64decode(part['inlineData']['data'])
            img_path = output_dir / f"figure_v{iteration}.png"
            with open(img_path, "wb") as f:
                f.write(img_data)
            print(f"\n✅ Image saved: {img_path}")
            print(f"   Size: {len(img_data)/1024:.1f} KB")

except Exception as e:
    print(f"Parse error: {e}")
    print(f"Raw response: {str(data)[:500]}")
PYTHON

Backend: openai / openai-compatible（IMAGE_BACKEND=openai 或 openai-compatible 时， 用下面的渲染调用替换上方 Gemini 调用；$RENDER_PROMPT 构造方式完全相同）：

if [ "$IMAGE_BACKEND" = "openai" ]; then
    API_BASE="https://api.openai.com/v1"
    API_KEY="$OPENAI_API_KEY"
    MODEL="${IMAGE_MODEL:-gpt-image-1}"
else  # openai-compatible: Azure OpenAI / 自建网关 / 其他兼容端点
    API_BASE="${IMAGE_API_BASE%/}"
    API_KEY="$IMAGE_API_KEY"
    MODEL="${IMAGE_MODEL:?set IMAGE_MODEL for openai-compatible backend}"
fi

MODEL="$MODEL" python3 - "$RENDER_PROMPT" > /tmp/image_request.json << 'PYTHON'
import json, sys, os
json.dump({"model": os.environ.get("MODEL", "gpt-image-1"),
           "prompt": sys.argv[1], "size": "1536x1024", "n": 1},
          sys.stdout)
PYTHON

RESPONSE=$(curl -s --max-time 300 \
  -X POST "$API_BASE/images/generations" \
  -H "Authorization: Bearer $API_KEY" \
  -H 'Content-Type: application/json' \
  -d @/tmp/image_request.json)

if echo "$RESPONSE" | grep -q '"error"'; then
    echo "API Error:"; echo "$RESPONSE" | python3 -m json.tool 2>/dev/null || echo "$RESPONSE"
    exit 1
fi

# Extract image (b64_json or url)
echo "$RESPONSE" | python3 << 'PYTHON'
import sys, json, base64, urllib.request
from pathlib import Path
data = json.load(sys.stdin)
item = data["data"][0]
img_path = Path("figures/ai_generated") / "figure_v1.png"  # increment per iteration
if item.get("b64_json"):
    img_path.write_bytes(base64.b64decode(item["b64_json"]))
else:
    urllib.request.urlretrieve(item["url"], img_path)
print(f"Image saved: {img_path}")
PYTHON

后续 Step 5 的严格视觉评审、打分与迭代循环对所有后端一致。

Step 5: Claude STRICT Visual Review & Scoring (MANDATORY)

Claude MUST read the generated image and perform a STRICT review:

1. Visual Analysis: What does the image show in detail?
2. Strengths: What's good about it?
3. STRICT Verification: Check EVERY item below
4. Score: Rate 1-10 (10 = perfect) — BE STRICT!

STRICT Review Template:

## Claude's STRICT Review of Figure v{N}

### What I See
[Describe the generated image in DETAIL - every block, every arrow]

### Strengths
- [Strength 1]
- [Strength 2]

### ═══════════════════════════════════════════════════════════════
### STRICT VERIFICATION CHECKLIST (ALL must pass for score ≥ 9)
### ═══════════════════════════════════════════════════════════════

#### A. Arrow Correctness Verification (CRITICAL - any failure = score ≤ 6)
Check EACH arrow:
- [ ] Arrow 1: [Source] → [Target] — Does it point to the CORRECT target?
- [ ] Arrow 2: [Source] → [Target] — Does it point to the CORRECT target?
- [ ] Arrow 3: [Source] → [Target] — Does it point to the CORRECT target?
- [ ] Arrow 4: [Source] → [Target] — Does it point to the CORRECT target?
- [ ] Arrow 5: [Source] → [Target] — Does it point to the CORRECT target?
- [ ] Arrow 6: [Source] → [Target] — Does it point to the CORRECT target?

#### B. Block Content Verification (any failure = score ≤ 7)
Check EACH block:
- [ ] Block 1 "[Name]": Has correct label? Has sub-label? Content correct?
- [ ] Block 2 "[Name]": Has correct label? Has sub-label? Content correct?
- [ ] Block 3 "[Name]": Has correct label? Has sub-label? Content correct?
- [ ] Block 4 "[Name]": Has correct label? Has sub-label? Content correct?
- [ ] Block 5 "[Name]": Has correct label? Has sub-label? Content correct?
- [ ] Block 6 "[Name]": Has correct label? Has sub-label? Content correct?
- [ ] Block 7 "[Name]": Has correct label? Has sub-label? Content correct?

#### C. Arrow Visibility (any failure = score ≤ 7)
- [ ] ALL arrows are THICK (≥5px visible stroke)
- [ ] ALL arrows have CLEAR arrowheads (large triangular heads)
- [ ] ALL arrows are BLACK or DARK GRAY (not light colors)
- [ ] NO arrows are too thin or invisible

#### D. Arrow Labels (any failure = score ≤ 7)
- [ ] EVERY arrow has a text label
- [ ] Labels are readable (not too small)
- [ ] Labels correctly describe the data flowing

#### E. Visual Appeal (科研风格 - Balanced Academic Style) (any failure = score ≤ 8)
- [ ] **有适度视觉吸引力** — 有subtle渐变或圆角，但不夸张
- [ ] **不是平淡方块** — 有一定设计感
- [ ] **不过度装饰** — 没有重阴影、发光效果、彩虹配色
- [ ] **专业学术风格** — 像CVPR论文中的图表，不是PPT模板
- [ ] **Internal structure visible** — 大模块内部显示子组件结构
- [ ] **Color palette: 3-4种协调色** — 不是彩虹，也不是纯黑白

#### E2. Visual Appeal - RED FLAGS (immediate score ≤ 7 if found)
- [ ] **NO heavy drop shadows** (重阴影 = too flashy)
- [ ] **NO glowing effects** (发光效果 = too flashy)
- [ ] **NO rainbow gradients** (彩虹渐变 = unprofessional)
- [ ] **NO excessive decorative icons** (过多装饰图标 = distracting)

#### F. Layout & Flow (any failure = score ≤ 7)
- [ ] Clean horizontal left-to-right flow
- [ ] No arrow crossings
- [ ] Data flow traceable in 5 seconds
- [ ] Balanced spacing (not cramped, not sparse)

#### G. Style Compliance
- [ ] CVPR/NeurIPS professional style
- [ ] Color palette appropriate (not rainbow)
- [ ] Font readable
- [ ] Print-friendly (grayscale test)

### ═══════════════════════════════════════════════════════════════

### Issues Found (BE SPECIFIC)
1. [Issue 1]: [EXACTLY what is wrong] → [How to fix]
2. [Issue 2]: [EXACTLY what is wrong] → [How to fix]
3. [Issue 3]: [EXACTLY what is wrong] → [How to fix]

### Score: X/10

### STRICT Score Breakdown Guide:
- **10**: Perfect. No issues. Publication-ready masterpiece. 视觉风格完美平衡。
- **9**: Excellent. Minor issues that don't affect understanding. 可以直接使用。
- **8**: Good but has noticeable issues. 视觉上太平淡或太花哨都需要改进。
- **7**: Usable but has clear problems. 箭头或内容有问题。
- **6**: Has arrow direction errors (箭头指向错误) OR missing major components.
- **1-5**: Major issues. Unacceptable.

### Visual Style Scoring (视觉风格评分):
- **太花哨 (Too flashy)**: 重阴影、发光效果、彩虹配色 → score ≤ 7
- **太平淡 (Too plain)**: 纯黑白方块、无任何视觉设计 → score ≤ 8
- **恰到好处 (Balanced)**: 适度渐变、圆角、清晰层次 → score 9-10

### Verdict
[ ] ACCEPT (score ≥ 9 AND all critical checks pass)
[ ] REFINE (score < 9 OR any critical check fails)

**If REFINE: List the EXACT issues that must be fixed**

Step 6: Decision Point

IF score >= 9 AND all critical checks pass:
    → Accept figure, generate LaTeX snippet, DONE
ELSE IF iteration < MAX_ITERATIONS:
    → Generate SPECIFIC improvement prompt based on EXACT issues
    → Go to Step 2 (Gemini Layout) with refined prompt
ELSE:
    → Max iterations reached, show best version
    → Ask user if they want to continue or accept

Step 7: Generate Improvement Prompt (for refinement)

Claude generates TARGETED improvement prompt with EXACT issues:

Refine this academic diagram. This is iteration {N}.

## ═══════════════════════════════════════════════════════════════
## CRITICAL: Fix These EXACT Issues (from previous review)
## ═══════════════════════════════════════════════════════════════

### Arrow Direction Errors (MUST FIX):
1. [EXACT issue]: Arrow from [A] to [B] is pointing to wrong target. It should point to [C] instead.
2. [EXACT issue]: ...

### Missing Arrow Labels (MUST FIX):
1. Arrow from [A] to [B] is missing label "[data type]"
2. ...

### Block Content Issues (MUST FIX):
1. Block "[Name]" has wrong label. Should be "[correct label]"
2. ...

### Visual Appeal Issues (SHOULD FIX):
1. Blocks are too plain. Add [gradients/shadows/internal structure]
2. ...

## Keep These Good Elements:
- [What to preserve from previous version]

## Generate the improved figure with ALL issues fixed.

Step 8: Final Output

When figure is accepted (score ≥ 9):

% === AI-Generated Figure ===
\begin{figure*}[t]
    \centering
    \includegraphics[width=0.95\textwidth]{figures/ai_generated/figure_final.png}
    \caption{[Caption based on user's original request].}
    \label{fig:[label]}
\end{figure*}

Key Rules (MUST FOLLOW - STRICT)

1. NEVER skip the review step — Always read and STRICTLY score the image
2. NEVER accept score < 9 — Keep refining until excellence
3. VERIFY EVERY ARROW DIRECTION — Wrong direction = automatic fail (score ≤ 6)
4. VERIFY EVERY BLOCK CONTENT — Wrong content = automatic fail (score ≤ 7)
5. BE SPECIFIC in feedback — "Arrow from A to B points to wrong target C" not "arrow is wrong"
6. SAVE all iterations — Keep version history for comparison
7. Claude is the STRICT boss — Accept only excellence, not "good enough"
8. ARROW CORRECTNESS IS NON-NEGOTIABLE — Any wrong arrow direction = reject
9. VISUAL APPEAL MATTERS — Plain boring figures = score ≤ 8
10. Target score is 9 — Not 8, not "good enough"
11. USE MULTI-STAGE WORKFLOW — Claude → Gemini Layout → Gemini Style → Paperbanana → Claude Review
12. USE CONFIGURED MODELS — REASONING_MODEL for reasoning, IMAGE_MODEL for rendering

Output Structure

figures/ai_generated/
├── layout_description.txt  # Step 2: Gemini layout optimization output
├── style_spec.txt          # Step 3: Gemini style verification output
├── figure_v1.png           # Iteration 1 (Paperbanana render)
├── 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         # All review scores and STRICT feedback

Model Summary

Stage	Model	Purpose
Step 1	Claude	Parse request, create initial prompt
Step 2	REASONING_MODEL	Layout optimization (positioning, spacing, grouping)
Step 3	REASONING_MODEL	CVPR/NeurIPS style verification
Step 4	IMAGE_MODEL (Paperbanana)	High-quality image rendering
Step 5	Claude	STRICT visual review and scoring