| name | figure-generation |
| description | Generate all publication-quality PNG figures for the paper from figure prompts using Nano Banana Pro (Gemini 3 Pro Image). Use when user says "生成图", "generate figures", "画图", or after all upstream skills/agents have produced figure prompts. |
| argument-hint | ["all / specific figure_id"] |
| allowed-tools | Bash(*), Read, Write, Edit, Grep, Glob, WebSearch, WebFetch |
Skill-7: Academic Figure Generation (Nano Banana Pro)
Generate publication-quality PNG figures for: $ARGUMENTS
Overview
This skill reads figure prompt files from output/figure_prompts/, polishes each prompt for academic quality, matches it with a reference example image, then calls Nano Banana Pro (Gemini 3 Pro Image, gemini-3-pro-image-preview) to directly generate publication-quality PNG figures.
Key features:
- Prompt polishing — Claude reviews and refines each prompt before sending to Gemini (serif font, academic color palette, clear arrows, no legend)
- Reference image matching — each figure type maps to an example PNG for style guidance, passed to Gemini as inline image input
- Direct PNG output — Gemini generates high-resolution PNG natively (1K/2K/4K)
- Academic style enforcement — consistent palette, serif typography, professional flowchart/architecture style
- REST API (urllib) — uses raw REST calls, NOT the Python SDK (SDK has known issues with proxy URLs)
Environment Setup
export GEMINI_BASE_URL="https://api.openai-proxy.org/google"
export GOOGLE_API_KEY="your-api-key-here"
Or configure in scripts/config.yaml:
api:
base_url: "https://api.openai-proxy.org/google"
key: "your-api-key"
model: "gemini-3-pro-image-preview"
Inputs
| Input | Path | Purpose |
|---|
| Figure prompts | output/figure_prompts/{review,methodology,results,paper}/*.md | Content descriptions for each figure |
| Reference examples | references/examples/{review,methodology,results,paper}/*.png | Style reference PNGs sent to Gemini as inline image |
| Style guide | references/academic_figure_style.md | Color palette, typography, layout rules |
| Config | scripts/config.yaml | API settings |
Outputs
output/figures/
├── review/
│ ├── fig_prisma_flowchart.png # Publication-quality PNG (2K default)
│ ├── fig_research_taxonomy.png
│ └── ...
├── methodology/
│ ├── fig_system_architecture.png
│ ├── fig_kg_pipeline.png
│ └── ...
├── results/
│ ├── fig_retrieval_comparison.png
│ ├── fig_llm_comparison.png
│ └── ...
├── paper/
│ ├── fig_introduction_overview.png
│ └── ...
└── figure_generation_report.json # Per-figure success/failure log
Format: PNG only. No SVG/XML (Gemini outputs raster PNG natively; for publication this is sufficient at 2K/4K resolution).
References & Examples
Reference Example PNGs (User Provides)
Place PNG reference images in these folders. Each image is sent to Gemini as an inline image reference — Gemini matches the visual style but uses the content from the figure prompt.
references/examples/
├── review/
│ ├── prisma_flowchart_example.png
│ ├── research_taxonomy_example.png
│ ├── concept_trend_example.png
│ └── literature_positioning_example.png
├── methodology/
│ ├── system_architecture_example.png
│ ├── kg_pipeline_example.png
│ ├── retrieval_architecture_example.png
│ ├── ontology_schema_example.png
│ └── chatbot_interface_example.png
├── results/
│ ├── kg_visualization_example.png
│ ├── bar_chart_comparison_example.png
│ ├── radar_chart_example.png
│ └── table_figure_example.png
└── paper/
└── graphical_abstract_example.png
Style Guide
| Reference | Path | Purpose |
|---|
| Academic Figure Style | references/academic_figure_style.md | Palette (hex), serif typography, layout rules |
Workflow
Phase 1: Pre-check
- Verify
output/figure_prompts/ exists and has .md files
- Verify
scripts/config.yaml has API key configured (or GOOGLE_API_KEY env var set)
- If API key missing, ask user to configure
- List all figure prompts and their matching reference images
Report to user:
Found [N] figure prompts:
review/fig_prisma_flowchart.md → ref: prisma_flowchart_example.png ✓
review/fig_research_taxonomy.md → ref: research_taxonomy_example.png ✓
methodology/fig_system_architecture.md → ref: system_architecture_example.png ✓
methodology/fig_kg_pipeline.md → ref: (no reference) ⚠
...
Phase 2: Prompt Polishing (Claude Does This)
Before sending each prompt to Gemini, Claude MUST review and polish it. This is where prompt quality translates to figure quality.
For each figure prompt file:
- Read the raw prompt from
output/figure_prompts/.../*.md
- Read the style guide from
references/academic_figure_style.md
- Polish the prompt to ensure these requirements (from the Nano Banana Pro template):
- Language: All text labels in English
- Font: Serif font (Times New Roman style)
- Overall style: Professional academic architecture/flowchart
- Colors: Primary blue and gray; accent colors (warm orange/teal/green) for innovation highlights
- Data flow: Clear arrows indicating information flow. No legend.
- Schematics: Include representative illustrative elements for each stage/component
- Begin the prompt with "Create a..." — per Nano Banana Pro template
- No meta-text: Prompt must not ask the image to include layout annotations or font specifications
- Save the polished prompt back (overwrite or save as
_polished.md)
Prompt polishing template (inspired by the Nano Banana Pro template):
Create a [figure type] for "[topic]".
Requirements:
- Language: All text labels in English
- Font: Serif font (Times New Roman style)
- Overall aesthetic: Professional academic architecture/flowchart
- Primary colors: clear blue and gray
- Accent colors for key innovations: warm orange / teal / green
- Data/information flow: clear directional arrows (NO legend)
- Schematics: AI generates representative illustrations for each stage
- Layout: [horizontal / vertical / grid], balanced composition
- Size: single-column (1000px) or double-column (2000px)
Content:
[detailed content description from the original figure prompt]
Reference context:
[any relevant data source or paper reference]
Polishing checklist per prompt:
Phase 3: Generate Figures
Run the generation script:
python scripts/generate_figure.py --all --output-dir output/figures --verbose
python scripts/generate_figure.py --prompt output/figure_prompts/review/fig_prisma_flowchart.md -o output/figures -v
What the script does for each figure:
- Load polished prompt + frontmatter (aspect_ratio, image_size)
- Find matching reference example PNG
- Build API request: polished prompt + style guide + inline reference image (if exists)
- Call Gemini via REST API (urllib) — not SDK (proxy compatibility)
- Extract PNG from response (
inlineData field, base64-decoded)
- Save as
.png to output/figures/{section}/{figure_id}.png
Phase 4: Quality Review
After generation, review each figure by opening the PNG:
- Content matches prompt? All requested elements present?
- Labels readable? All text rendered correctly in English?
- Colors match academic palette? Blue/gray primary, warm accents?
- Professional appearance? No artifacts, proper spacing?
- Serif font used? Check text rendering.
If a figure is unsatisfactory:
- Refine the prompt with more specific instructions (add examples, clarify colors, specify layout)
- Regenerate that single figure (not the whole batch)
- Do NOT regenerate all figures — cost and time.
Phase 5: Generate Report
The script automatically creates output/figures/figure_generation_report.json.
Additionally, print a summary for the user:
## Figure Generation Summary
| Figure | Section | Aspect | Size | Reference Used | Status |
|--------|---------|--------|------|----------------|--------|
| fig_prisma_flowchart | Review | 16:9 | 2K | prisma_example.png | ✓ |
| fig_system_architecture | Methodology | 16:9 | 2K | system_arch_example.png | ✓ |
| ... | ... | ... | ... | ... | ... |
Total: [N] figures generated, saved as PNG to output/figures/
Phase 6: Update plan04.md
If figure generation is tracked in plan04.md, update task status per Auto-Update Protocol.
Figure Type → Reference Mapping
| Figure ID Pattern | Reference Example | Category |
|---|
fig_prisma* | review/prisma_flowchart_example.png | Review |
fig_research_taxonomy* | review/research_taxonomy_example.png | Review |
fig_concept* | review/concept_trend_example.png | Review |
fig_*positioning* | review/literature_positioning_example.png | Review |
fig_system_architecture* | methodology/system_architecture_example.png | Methodology |
fig_kg_pipeline* | methodology/kg_pipeline_example.png | Methodology |
fig_retrieval* | methodology/retrieval_architecture_example.png | Methodology |
fig_ontology* | methodology/ontology_schema_example.png | Methodology |
fig_chatbot* | methodology/chatbot_interface_example.png | Methodology |
fig_kg_vis* | results/kg_visualization_example.png | Results |
fig_*comparison* | results/bar_chart_comparison_example.png | Results |
fig_*quality* / fig_*radar* | results/radar_chart_example.png | Results |
fig_introduction* | paper/graphical_abstract_example.png | Paper |
Nano Banana Pro API Details
- Model:
gemini-3-pro-image-preview
- Endpoint:
{BASE_URL}/v1beta/models/gemini-3-pro-image-preview:generateContent
- Method: REST (urllib), not SDK
- Response modalities:
["TEXT", "IMAGE"]
- Image config:
aspectRatio: 1:1, 3:2, 16:9 (default), 9:16, 21:9imageSize: 1K (fast), 2K (balanced, default), 4K (highest quality)
- Timeout: 180s default (increase for 4K)
- Response format:
candidates[0].content.parts[*].inlineData.data (base64 PNG)
Common Pitfalls (from Nano Banana Pro Reference)
| Problem | Cause | Fix |
|---|
Extra inputs are not permitted | Using Python SDK with proxy URL | Use REST API (urllib) — the bundled script does this |
| Empty/blocked response | Prompt triggered safety filter | Rephrase prompt, avoid restricted content |
| Timeout | Large image (4K) or slow network | Increase --timeout, try 2K first |
| Text labels in wrong language | Prompt didn't specify English | Add "Language: All text labels in English" to prompt |
| No serif font | Prompt didn't specify font | Explicitly state "Serif font (Times New Roman style)" |
| Legend appears uninvited | Prompt didn't explicitly forbid it | Add "No legend" to the prompt |
Key Rules
- Polish before generating: Claude MUST review every prompt before it goes to Gemini. Quality of input = quality of output.
- PNG output only: Nano Banana Pro generates raster PNG natively. No SVG/XML — 2K/4K PNG is sufficient for publication.
- REST API only: Use urllib (the bundled script), NEVER the
google-genai SDK with proxy URLs.
- Reference images guide style, not content: Gemini matches visual style but uses content from the prompt.
- Academic palette enforced: All figures use the palette from
academic_figure_style.md (blue + gray primary, warm accents).
- Serif font mandatory: All text labels in serif font (Times New Roman style).
- No legend by default: Unless specifically needed, prompt states "No legend" to keep figures clean.
- Config-driven: API base_url, key, model in
scripts/config.yaml. No hardcoded credentials.
- One figure at a time if quality issues: Don't regenerate all — fix and retry the specific one.
- Report missing references: If no reference PNG exists for a figure, warn the user before proceeding.
- All output to
output/figures/: Organized by category (review, methodology, results, paper).
Composing with Other Skills
skill-5 /review-writing → output/figure_prompts/review/
@kg-graphrag-builder-agent → output/figure_prompts/methodology/ + results/
@evaluation-agent → output/figure_prompts/results/
skill-6 /paper-writing → output/figure_prompts/paper/
skill-7 /figure-generation → output/figures/ (PNG) ← YOU ARE HERE
skill-8 /paper-doc → embeds PNG figures into final_paper.docx
