convez.ai
SkillDados & AnalyticsAvançado
TAO-QKV

Paper Figures — Gráficos em Nível de Publicação

TAO-QKV/paper-figures

Figuras publication-grade para papers científicos em matplotlib/TikZ

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill Claude para gerar figuras científicas em qualidade de publicação usando matplotlib, seaborn e TikZ. Cria painéis complexos com múltiplos subplots, insets, evidência cruzada e tratamento honesto de incerteza — direto em PDF, PNG e SVG sem revisões de editor.

Instalação

$npx convez-ai TAO-QKV/paper-figures

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
61
Forks
5
Acompanhando
1
Saúde
42%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Gerar painéis de evidência biomédica com diagnósticos integrados
  • Criar figuras hero de framework e métodos em TikZ compilável
  • Produzir gráficos reproduzíveis com validação cruzada e incerteza
  • Automatizar composição de figuras complexas para submissão de journals

Veredito

Ferramenta sólida e madura para pesquisadores que precisam de figuras de alto padrão. Aborda bem o problema de reprodutibilidade e composição visual, mas requer familiaridade com matplotlib/TikZ e domínio do repositório para extrair todo o potencial.

README

A tradução para o português deste README ainda está na fila.

paper-figures — a publication-grade figure system for scientific papers

CI License: MIT Python matplotlib figures

Turn "a dataset + a claim" into a publication-grade figure — the kind a top-journal editor accepts with no revision request. A self-contained Claude skill: reproducible matplotlib/seaborn plots for data figures, and original TikZ for hero/method figures.

Example output

The point is not "many chart thumbnails"; it is complex paper figures with hierarchy, insets, cross-panel evidence, and honest uncertainty.

Complex biomedical evidence panel - a single integrated figure with a hero response manifold, decision field, calibration/ROC diagnostics, volcano + Manhattan omics evidence, imaging plate, survival validation, forest plot, covariance heatmap, and mechanism graph (examples/complex_panels.py):

Complex biomedical evidence panel

Complex spatial-physical evidence panel - field + streamline + trajectory + sensors as the hero panel, with response surface, spectrum, time-series uncertainty, QQ residual check, phase portrait, anomaly zoom, and registered imaging readouts:

Complex spatial-physical evidence panel

Complex method/system figure - architecture, workflow, evidence flow, state machine, production timeline, dependency graph, clustering, set overlap, ternary trade-space, and quality-lift slopegraph:

Complex method and system panel

Framework / method hero figure (original, compilable TikZ) - the figure a paper opens with: a two-paradigm framework (mechanistic + data-driven, fused at the contribution), with a real method object embedded in each lane (a degradation curve crossing its failure threshold, an actual MLP, the fusion's prediction lines) instead of text-in-boxes. One pdflatex away, zero compile risk to the main document (examples/hero_tikz/framework_hero.tex; see framework-figures.md):

Two-paradigm framework hero figure in TikZ

Additional composite galleries are reproducible coverage examples, not the visual front door: showcase_gallery.py, coverage_gallery.py, domain_showcases.py.

Lower-level API smoke/demo galleries are kept as regression checks rather than the visual front door: gallery.py, archetypes_gallery.py, paradigms_gallery.py, modern_gallery.py, scientific_gallery.py, advanced_gallery.py, rich_gallery.py, before_after.py.

Quickstart (with Claude)

  1. Put your data in data/processed/your_data.csv.
  2. Open Claude in this folder and say:

    Plot a figure from data/processed/your_data.csv; the claim is "Y declines linearly with X".

  3. The paper-figure-generation skill triggers and produces:
    • a reproducible script scripts/figN_<name>.py
    • the figure in three formats outputs/figures/figN_<name>.{pdf,png,svg}
    • a caption that states the conclusion
    • a row in outputs/figures/figure_manifest.csv

Use it as a library (no Claude needed)

pip install -e .          # or: pip install -e ".[extras]"  for pandas/seaborn/scipy
import numpy as np
from paperfig import paper_style, save, timeseries_ci, scatter_fit, heatmap

paper_style(font='sans', journal='ieee')   # font: 'sans'|'serif'|'cn'; journal: None|'ieee'|'nature'|'pnas'

t = np.linspace(0, 10, 50); y = 1 - np.exp(-0.3 * t)
fig, ax = timeseries_ci(t, y + np.random.normal(0, .05, 50), y, y - .06, y + .06,
                        ylabel='Signal $y$')
ax.set_title('My result')                  # tweak the returned fig/ax freely
save(fig, 'fig1_main_result')              # → outputs/figures/fig1_main_result.{pdf,png,svg}

48 callable chart types, each returning (fig, ax) or (fig, axes):

DomainFunctions
coretimeseries_ci, sorted_bar, grouped_bar, scatter_fit, heatmap, pareto, tornado
distributionsviolin, raincloud, ridgeline, histogram, box, ecdf, qq, jointplot, hexbin_density
ML / diagnosticsroc_curve, pr_curve, calibration, confusion, residual_diag, feature_importance, convergence_curve, embedding_scatter, attention_map
stats / comparisonbar_err (+ significance brackets), bubble, forest, bland_altman, slopegraph, radar
omics / genomics / medicalvolcano, manhattan, survival (Kaplan-Meier)
fields / dynamics / spectracontour_field, streamplot_field, surface3d, phase_portrait, trajectory, spectrum, ternary
structure / set / flowsankey, chord, dendrogram, network_graph, venn2
imaging platesimage_panel (microscopy / medical / remote-sensing image panels with scalebars)
transfer / methodalignment_scatter (P5 motif)

Best-fit route: quantitative and ML panels use the callable matplotlib API; mechanism / architecture / workflow figures use the TikZ composition paradigms; image-heavy evidence uses image_panel plus quantitative subpanels. Pie/donut charts are intentionally not highlighted because the cookbook treats them as low data-ink defaults; prefer sorted bars unless composition is the point.

  • Journal presets — paper_style(journal=...) sets that publisher's single-column figure size + font sizes:

    journal1-col widthjournal1-col widthjournal1-col width
    nature89 mmcell85 mmacs3.33 in
    science57 mmpnas87 mmrsc83 mm
    ieee3.5 inelsevier90 mmaps86 mm

    (From each publisher's public author guidelines — verify against the current guidelines, specs drift.)

  • LaTeX is optional — figures render without a LaTeX install; opt in with paper_style(tex=True) for true LaTeX typography.

Or just a matplotlib style (no API)

import paperfig                 # registers the style name
import matplotlib.pyplot as plt
plt.style.use('paperfig')       # now any plt.* plot is publication-grade

The idea

Publication-grade = the cookbook's §0b four axes, all true at once: Depth (the figure argues a mechanism) × Elegance (one figure, one claim) × Unimpeachable (carries its own uncertainty + reproducible) × Visible gap (reads journal-grade at a glance). Archetypes A1–A13 are the floor, not the ceiling.

Relation to SciencePlots

SciencePlots is the excellent, popular toolkit for matplotlib styles — and paperfig borrows its best ideas: a plt.style.use('paperfig') .mplstyle, cascading journal presets, and citability. The differences:

  • paperfig does not require LaTeX (SciencePlots does); LaTeX is opt-in (tex=True).
  • paperfig adds what a style sheet can't: a quality bar (§0b four axes), hero/method-figure composition paradigms (§I P1–P6), callable archetype functions (timeseries_ci, alignment_scatter, …), and original TikZ for method figures (§K).

Think of it as: SciencePlots-style presets, plus the judgment and building blocks to make the figure itself publication-grade. If you only want journal styles, SciencePlots is great; if you want the figure's content held to a bar, use paperfig.

Layout

PathWhat
paperfig/the installable package: style.py (preset) + archetypes.py (callable A1–A10 + P5)
examples/runnable galleries (complex_panels.py, showcase_gallery.py, coverage_gallery.py, domain_showcases.py, scientific_gallery.py, advanced_gallery.py, archetypes_gallery.py, paradigms_gallery.py) + five complete compilable TikZ framework heroes (hero_tikz/: P1 pipeline, P2 swimlanes, + T4 graphical-model / data-tensor / matrix-fit families)
tests/pytest smoke tests (every archetype + three-format save)
.claude/skills/paper-figure-generation/SKILL.mdthe Claude skill (triggers, hard rules)
.../references/figure-cookbook.mdmain reference: §0b quality bar · §0a contract · §0 style · §A archetypes A1–A13 · §I composition paradigms P1–P6 · §J craft spec · §K original TikZ · §L external template library · §M Origin front-end
.../references/figure-critique.mdthe quality gate: operationalizes the §0b four axes + worked before→after critiques
.../references/framework-figures.mdframework/method-figure deep-dive: five compilable hero examples + borrowable TikZ techniques + curated resources
.../references/caption-and-quality.mdcaption writing + final quality checklist
scripts/critique.pyrunnable critique gate — static-analyses a figure script against the bar, exits non-zero on a hard-rule miss
scripts/_style.pyback-compat shim → paperfig.style
data/processed/ · outputs/figures/input data · output + figure_manifest.csv
CLAUDE.md · pyproject.tomlproject role + <TEMPLATE_LIB> convention · packaging

Hard rules

Reproducible (a script reads data from a file; no inline data > 20 rows; no AI-generated images) · vector three-format (PDF + PNG + SVG; plus TIFF/EPS at journal submission via save(fig, name, formats=('pdf','tiff'), dpi=600)) · style preset called once · colorblind- and grayscale-safe · units / N / uncertainty on the figure · pass the §0b four axes + §F reproducibility checklist before "done".

These are enforced, not just exhorted: python scripts/critique.py scripts/figN_<name>.py static-analyses a figure script against the bar — hard-rule misses (no uncertainty, jet colormap, equal grid, default style, PNG-only) come back as FAIL (non-zero exit, so it gates CI), and the judgment-half (does the figure argue a mechanism? does one hero panel dominate? B&W-legible?) prints as prompts you answer by hand. See references/figure-critique.md.

Requirements

Python 3.9+, matplotlib, numpy, pandas; seaborn for heatmaps; a LaTeX install with TikZ for §K hero figures (optional). See requirements.txt.

License

MIT © 2026 TAO-QKV. Use it, fork it, adapt it for your papers.

Ver o original no GitHub

Bom para

Relacionados