Figure

Figure Skill

Publication-quality figures via the hybrid pipeline: let Gemini (Nano Banana) render photorealistic / textured / 3D components, let TikZ own all symbols, arrows, labels, equations, and layout. You get the visual richness of AI rendering AND the precision of code.

When to use which backend

Copy code
┌─────────────────────────────────────────────────────────────────┐
│  What are you drawing?                          │ Backend       │
├─────────────────────────────────────────────────┼───────────────┤
│  Data plot (lines, bars, heatmaps from data)    │ pgfplots      │
│  Quantum circuit (Hadamard, CNOT, measure)      │ quantikz      │
│  Feynman diagram                                │ feynman       │
│  Molecule / Lewis structure                     │ chemfig       │
│  Electronic circuit (op-amp, RLC)               │ circuitikz    │
│  Atomic energy levels + transitions             │ energy_levels │
│  Phase-space trajectory                         │ phase_space   │
│  Pulse sequence (NMR / qubit timing)            │ pulse_sequence│
│  Optical bench cartoon (2D, flat)               │ optical_setup │
│  Concept schematic with textured 3D objects     │ HYBRID        │
│    (Paul trap, cavity, apparatus, animal etc.)  │               │
└─────────────────────────────────────────────────────────────────┘

For HYBRID, pick raster components sparingly: only for textured / 3D / glossy objects where TikZ would take 50+ lines of patches. Everything else → TikZ.

The hybrid pipeline

Copy code
1. decide which elements are raster (3D/textured/glossy) vs vector (everything else)
2. for each raster element:
     python skills/figure/scripts/hybrid_gen.py \
       --name <outName>  --prompt "<object spec>"  --style "<style suffix>" \
       --out assets/<outName>.png
   → generates isolated-object PNG on white bg, then rembg → transparent PNG
3. write figure_X.tex starting from skills/figure/templates/<best>.tex
     - \includegraphics{assets/<outName>.png} for raster slots
     - TikZ \draw \node for all labels, arrows, equations
4. compile:    pdflatex figure_X.tex
   preview:    pdftoppm -r 200 figure_X.pdf preview -png
5. Read preview-1.png (vision) → find issues → edit .tex → recompile (≤3 iters)
6. final output: figures/figure_X.{tex,pdf}

Style consistency

ALWAYS check figures/style_guide.md first (if it exists) for mandatory palette, fonts, line weights. If absent, default to Okabe-Ito (see references/palettes.md).

Before starting, look at existing figures/*.png via Read — match their palette, font family, and panel-label style exactly.

Strict rules (when using this skill as illustrator)

• Never originate content decisions (what data, what physics, what params).
• Only make style/composition/rendering decisions.
• LaTeX symbols (\ket{r}, F_{C_4}, \mu m, etc.) MUST be TikZ-native — never rely on Nano Banana to render text correctly (it can't).
• Raster components must have NO text in the prompt ("no labels, no captions").
• Raster components must have consistent style suffix across a figure for lighting/palette coherence.

Templates

See templates/ for starting points. Each is a complete \documentclass{standalone} ready to compile. Pick the closest match, copy to figures/, edit.

References

• references/decision_tree.md — full decision flow for tool selection
• references/palettes.md — Okabe-Ito, Nature, Tableau colorblind-safe palettes
• references/pitfalls.md — package conflicts (chemfig vs tikz-cd), font fallbacks, common bugs

Scripts

• scripts/hybrid_gen.py — Gemini image generation + rembg background removal. CLI: --name --prompt --style --out [--model gemini-2.5-flash-image] [--no-rembg]
• scripts/requirements.txt — Python deps

First-time setup

bashCopy code
cd skills/figure
pip install -r scripts/requirements.txt
export GEMINI_API_KEY=...   # if not already set