figure-pipeline

name

figure-pipeline

description

Five-step figure generation pipeline inspired by PaperVizAgent (Google Research, 2026). Orchestrates Retriever → Planner → Stylist → Visualizer → Critic stages for publication-quality scientific figures. Retrieves reference figures from literature, plans layout and composition, applies journal-specific styling, generates the figure, then critiques and refines. Use when the user needs high-quality figures for papers/presentations and wants a more deliberate, reference-driven approach than direct code generation. Especially useful for multi-panel figures and complex data visualizations.

Five-Step Figure Generation Pipeline

Generate publication-quality scientific figures through a structured pipeline with reference retrieval, planning, styling, generation, and self-critique. Inspired by PaperVizAgent's five-agent architecture.

When to Use

• Multi-panel composite figures (e.g., "Figure 1: A) boxplot, B) KM curve, C) heatmap, D) network")
• Figures that need to match a specific journal's style
• Complex visualizations where direct code generation often produces suboptimal layouts
• User explicitly asks for "publication-quality" or "journal-ready" figures
• Graphical abstracts

When NOT to use (overkill):

• Single simple plot → just write matplotlib/ggplot2 code directly
• Mechanism diagrams → use svg-scientific-figures skill
• Quick exploratory plots → direct code

---

Pipeline Steps

Step 1: RETRIEVE — Find Reference Figures

Search for published figures on the same topic to establish visual expectations:

bash: echo "=== PubMed figure captions ===" && \
curl -s "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi?db=pubmed&retmode=json&retmax=10&sort=relevance&term=TOPIC+AND+(figure+OR+visualization)" && \
echo -e "\n=== bioRxiv recent figures ===" && \
curl -s "https://api.biorxiv.org/details/biorxiv/2025-01-01/2026-03-18?jou=biorxiv" 2>/dev/null | head -c 2000

Also search for figure captions in full-text papers via Asta or Europe PMC:

curl -s "https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=TOPIC+AND+FIG_TYPE:figure&format=json&pageSize=5"

From reference figures, extract:

• Figure types used (violin plot vs boxplot, clustered vs simple heatmap)
• Panel arrangement (2x2 grid, horizontal strip, vertical stack)
• Color schemes (which journal palette)
• Annotation patterns (significance brackets, gene labels, axis formatting)

Step 2: PLAN — Design Figure Composition

Create a structured figure plan:

{
  "figure_id": "Figure 1",
  "title": "THBS2 Expression and Survival Analysis Across Cancer Types",
  "layout": {
    "type": "grid",
    "rows": 2,
    "cols": 2,
    "width_cm": 17.5,
    "height_cm": 15
  },
  "panels": [
    {
      "id": "A",
      "type": "boxplot",
      "data": "TCGA pan-cancer expression",
      "x": "cancer_type",
      "y": "THBS2_expression_TPM",
      "notes" 
    
    
       
       
       
       
       
    
    
       
       
       
       
    
    
       
       
       
       
    
  
   
     
     
     
     
  

Step 3: STYLE — Apply Journal-Specific Formatting

Select and apply styling rules based on target journal:

Journal Category	Width	DPI	Font	Palette	Panel Labels
Nature/Science/Cell	8.9cm (single) / 18.3cm (double)	300	Arial/Helvetica	NPG	Bold uppercase A, B, C
Lancet/NEJM/JAMA	8.5cm / 17.5cm	300	Arial	Lancet/NEJM	Bold a, b, c
Cancer Research	8.5cm / 17.5cm	300	Arial	AACR	Bold A, B, C
Default (no journal specified)	17.5cm (double)	300	Arial	NPG	Bold A, B, C

Styling rules:

• Axis labels: 8-9pt, sentence case
• Axis tick labels: 7-8pt
• Panel labels: 10-12pt, bold, top-left corner outside plot area
• Legend: 7-8pt, outside plot or in open space
• Significance annotations: * p<0.05, ** p<0.01, *** p<0.001, ns not significant
• Color: use colorblind-friendly palettes; never use red-green only
• Grid lines: minimal or none; let data speak

Step 4: VISUALIZE — Generate the Figure

Write Python or R code implementing the planned figure. Use the styling rules from Step 3.

Python approach (for multi-panel):

import matplotlib.pyplot as plt
import matplotlib.gridspec as gridspec

fig = plt.figure(figsize=(17.5/2.54, 15/2.54), dpi=300)
gs = gridspec.GridSpec(2, 2, hspace=0.35, wspace=0.35)

ax_a = fig.add_subplot(gs[0, 0])
# Panel A code...
ax_a.text(-0.15, 1.05, 'A', transform=ax_a.transAxes, fontsize=12, fontweight='bold')

ax_b = fig.add_subplot(gs[0, 1])
# Panel B code...

# ... etc

plt.savefig(out_path, dpi=300, bbox_inches='tight', facecolor='white')

R approach (using patchwork):

library(ggplot2); library(patchwork)
p_a <- ggplot(...) + labs(tag = "A")
p_b <- ggplot(...) + labs(tag = "B")
combined <- (p_a | p_b) / (p_c | p_d)
ggsave(out_path, combined, width=17.5, height=15, units="cm", dpi=300)

Step 5: CRITIQUE — Review and Refine

After generating the figure, perform a quality review:

Checklist:

1. All panel labels (A, B, C, D) present and correctly positioned?
2. All axis labels present, readable, and correctly spelled?
3. Legend present where needed? Not obscuring data?
4. Font sizes consistent across all panels?
5. Color palette consistent across all panels?
6. Significance annotations correct and complete?
7. White space balanced? No cramped or empty areas?
8. Data accurately represented? (spot-check values against source data)
9. Figure would be understandable from caption alone?
10. Resolution sufficient for print (300 DPI at target size)?

If any check fails, go back to Step 4 and fix. Maximum 2 refinement rounds.

After passing critique:

✅ Figure review passed (10/10 checks)
Saved: figures/figure1_thbs2_landscape.png (17.5 x 15 cm, 300 DPI)

---

Output

File	Description
figures/figure1_DESCRIPTION.png	Final PNG at 300 DPI
figures/figure1_DESCRIPTION.pdf	Vector PDF (when using R/matplotlib PDF backend)
figures/figure1_DESCRIPTION.svg	SVG if generated via the SVG skill

---

Integration with Research Recipes

When a Recipe generates multiple figure outputs, offer to combine them into a composite figure using this pipeline:

"已生成 4 张独立图表。需要用 Figure Pipeline 组装为一张多面板组合图吗？"