figure-quality-standards

Overview

Figures are the first thing reviewers look at. A sloppy figure signals sloppy science. This skill defines mandatory quality standards for every figure and table produced during Phase 4 (experiment execution) and Phase 5 (results integration), and enforced again during Phase 6 (paper writing).

Invoke this skill before generating any figure intended for a paper.

Universal Figure Standards

Resolution and Format

| Property | Requirement | |----------|------------| | Format | Vector (PDF or SVG) for plots; PNG at ≥300 DPI only for raster images (photos, heatmaps) | | Minimum DPI | 300 for raster, vector preferred for all line/bar/scatter plots | | File format for LaTeX | PDF (first choice) or EPS; avoid PNG/JPG for plots | | Size | Match column width of target venue (typically 3.25" single column, 6.875" double column for IEEE/ACM) |

Typography

| Property | Requirement | |----------|------------| | Font family | Match venue profile (see Venue-Specific Styles below) — sans-serif for CNS, serif for CS/IEEE | | Axis label size | ≥ 8pt after scaling to final print size | | Tick label size | ≥ 7pt after scaling | | Legend text size | ≥ 7pt after scaling | | Figure title | OMIT — do NOT put a title on the figure. The LaTeX \caption{} serves as the title. | | Panel labels | Bold lowercase for CNS-style (a, b, c); uppercase for CS/IEEE (A, B, C) — see venue profile |

Test: After generating a figure, mentally scale it to its final column width. If any text becomes unreadable at that size, increase the font.

Color

| Property | Requirement | |----------|------------| | Color palette | Match venue profile — Nature palette for CNS, Tableau 10 for CS, see below | | Consistency | ALL figures in the same paper must use the SAME color → method mapping | | Grayscale fallback | Figures must be distinguishable in grayscale (some venues print in B&W). Use markers/hatching in addition to color | | Maximum colors | ≤ 8 distinct colors per figure; beyond that, use subplots |

Layout and Readability

| Property | Requirement | |----------|------------| | Axis labels | Present on every axis; include units (e.g., "Accuracy (%)", "Time (s)") | | Grid lines | Depends on venue profile: subtle for CS, often absent for CNS | | Legend placement | Inside the plot area if space allows, otherwise outside. Never overlap data points. | | White space | Tight layout (bbox_inches='tight' in matplotlib); no excessive margins | | Aspect ratio | Standard ratios (4:3, 16:9, 1:1). Never stretched or squished. | | Subplot spacing | Consistent spacing; shared axes where appropriate to save space |

What Goes ON the Figure vs. IN LaTeX

<IMPORTANT> Figures and captions are SEPARATE things. The figure is an image file (PDF/SVG). The caption is LaTeX text in `\caption{}`. Do NOT confuse them. </IMPORTANT>

ON the figure (in the image file itself):

• Axis labels with units
• Tick labels
• Legend (if multiple series)
• Panel labels (a, b, c — positioned top-left of each subplot)
• Annotations (arrows, text callouts if needed)
• NO figure title (the caption replaces it)
• NO caption text

IN LaTeX \caption{}:

• What the figure shows (one sentence)
• Key takeaway / main observation
• Per-panel descriptions for multi-panel figures: "(a) Method comparison on Dataset X. (b) Ablation study..."
• Define abbreviations not defined in main text
• Statistical details if relevant ("Error bars indicate ± 1 std over 5 seeds")
• Must be self-contained: reader should understand the figure from caption alone

LaTeX pattern:

\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figures/main_comparison.pdf}
\caption{Comparison of methods on three benchmarks.
(\textbf{a}) Accuracy on Dataset X. Our method (blue) outperforms all baselines.
(\textbf{b}) Training efficiency. Our method converges 2$\times$ faster.
Error bars indicate $\pm$ 1 std over 5 random seeds.}
\label{fig:main}
\end{figure}

---

Venue-Specific Figure Styles

Read target_venue from research-anchor.yaml and select the matching profile. If unsure which profile to use, ask the user.

Profile: CNS (Nature, Science, Cell and their sub-journals)

Nature/Science/Cell have a distinctive, recognizable figure aesthetic. Matching it signals professionalism.

| Property | CNS Standard | |----------|-------------| | Font | Helvetica / Arial (sans-serif). Nature explicitly requires this. | | Font size | 5–7pt for figure text (Nature allows small text because figures are high-resolution) | | Panel labels | Bold lowercase: a, b, c, d — top-left of each panel, outside plot area | | Color palette | Nature palette: ['#E64B35', '#4DBBD5', '#00A087', '#3C5488', '#F39B7F', '#8491B4', '#91D1C2', '#DC0000', '#7E6148', '#B09C85'] | | Background | White. No gray background. | | Grid lines | None or extremely subtle. CNS figures are clean and minimal. | | Spines | Usually left + bottom only. No top/right spines. | | Line width | 0.5–1pt for data lines, 0.25–0.5pt for axes | | Multi-panel | Very common. 4–8 panels per figure. Use plt.subplot_mosaic() for complex layouts. | | Figure width | Single column: 89mm. Double column: 183mm. Full page: 183mm × 247mm. | | Annotations | Clean arrows, minimal text. Let the data speak. | | Bar plots | Thin bars, often with individual data points overlaid (strip/swarm plot on top of bars) | | Statistical markers | Brackets with asterisks: *, **, ***, ns |

CNS_STYLE = {
    'font.family': 'sans-serif',
    'font.sans-serif': ['Helvetica', 'Arial', 'DejaVu Sans'],
    'font.size': 7,
    'axes.titlesize': 8,
    'axes.labelsize': 7,
    'xtick.labelsize': 6,
    'ytick.labelsize': 6,
    'legend.fontsize': 6,
    'axes.linewidth': 0.5,
    'xtick.major.width': 0.5,
    'ytick.major.width': 0.5,
    'lines.linewidth': 1.0,
    'lines.markersize': 4,
    'axes.spines.top': False,
    'axes.spines.right': False,
    'axes.grid': False,
    'figure.dpi': 300,
    'savefig.dpi': 300,
    'savefig.bbox': 'tight',
}

NATURE_COLORS = ['#E64B35', '#4DBBD5', '#00A087', '#3C5488',
                 '#F39B7F', '#8491B4', '#91D1C2', '#DC0000',
                 '#7E6148', '#B09C85']

Profile: CS Conferences (NeurIPS, ICML, ICLR, CVPR, AAAI, ACL, EMNLP)

CS conferences prioritize clarity and information density over aesthetics.

| Property | CS Conference Standard | |----------|----------------------| | Font | Serif (Times, Computer Modern) to match paper body, OR sans-serif if consistent | | Font size | 8–10pt (larger than CNS because columns are wider) | | Panel labels | Uppercase or "(a) (b) (c)" in caption text; less common as on-figure labels | | Color palette | Tableau 10, ColorBrewer, or custom — must be colorblind-safe | | Background | White | | Grid lines | Light gray dashed — acceptable and often helpful for reading values | | Spines | Left + bottom preferred; all four acceptable | | Line width | 1.5–2pt for data lines (thick enough to see in projected slides too) | | Multi-panel | 2–4 panels typical. Subfigures common. | | Figure width | Single column: ~3.25". Double column: ~6.875" (LaTeX \textwidth). | | Error bands | Shaded regions (alpha=0.2) with mean line. Very standard for learning curves. | | Tables > figures | CS values tables highly; main results are often a table, not a figure |

CS_STYLE = {
    'font.family': 'serif',
    'font.serif': ['Times New Roman', 'DejaVu Serif', 'Computer Modern Roman'],
    'font.size': 10,
    'axes.titlesize': 11,
    'axes.labelsize': 10,
    'xtick.labelsize': 9,
    'ytick.labelsize': 9,
    'legend.fontsize': 9,
    'axes.linewidth': 0.8,
    'lines.linewidth': 1.5,
    'lines.markersize': 6,
    'axes.spines.top': False,
    'axes.spines.right': False,
    'axes.grid': True,
    'grid.alpha': 0.3,
    'grid.linestyle': '--',
    'figure.dpi': 300,
    'savefig.dpi': 300,
    'savefig.bbox': 'tight',
}

CS_COLORS = ['#4e79a7', '#f28e2b', '#e15759', '#76b7b2',
             '#59a14f', '#edc948', '#b07aa1', '#ff9da7',
             '#9c755f', '#bab0ac']  # Tableau 10

Profile: IEEE (Transactions, Conference Proceedings)

IEEE has strict formatting requirements documented in their author guidelines.

| Property | IEEE Standard | |----------|-------------| | Font | Times New Roman (mandatory) | | Font size | 8–10pt in figures | | Column width | Single: 3.5". Double: 7.16". | | Color | Allowed but paper may be printed B&W — MUST be readable in grayscale | | Captions | "Fig. 1." format (not "Figure 1") | | Line markers | Essential — distinguish lines by marker shape, not just color | | Grid lines | Optional, light |

IEEE_STYLE = {
    'font.family': 'serif',
    'font.serif': ['Times New Roman'],
    'font.size': 9,
    'axes.titlesize': 10,
    'axes.labelsize': 9,
    'xtick.labelsize': 8,
    'ytick.labelsize': 8,
    'legend.fontsize': 8,
    'axes.linewidth': 0.6,
    'lines.linewidth': 1.2,
    'lines.markersize': 5,
    'axes.spines.top': False,
    'axes.spines.right': False,
    'axes.grid': True,
    'grid.alpha': 0.2,
    'grid.linestyle': ':',
    'figure.dpi': 300,
    'savefig.dpi': 300,
    'savefig.bbox': 'tight',
}

Profile: Bioinformatics / Life Sciences (Genome Research, Bioinformatics, PNAS, eLife)

Life science journals generally follow CNS aesthetics with domain-specific plot types.

| Property | Life Science Standard | |----------|---------------------| | Font | Helvetica / Arial (following Nature/CNS tradition) | | Style | Very close to CNS profile above | | Domain-specific plots | Volcano plots, MA plots, heatmaps with dendrograms, Kaplan-Meier survival curves, Manhattan plots, circos plots | | Heatmap conventions | Row/column clustering dendrograms, diverging colormap (red-white-blue for expression), annotated color bars | | Statistical notation | Brackets with */**/***/ns between groups, Bonferroni-corrected p-values | | Bar plots | Individual data points overlaid (strip/swarm), NOT just bars with error bars |

Use CNS_STYLE and NATURE_COLORS from the CNS profile.

Profile: Physical Sciences (APS/Physical Review, ACS, RSC)

| Property | Physical Sciences Standard | |----------|--------------------------| | Font | Computer Modern or Helvetica depending on journal | | Figure width | APS single column: 3.375". Double: 6.75". | | Conventions | SI units on all axes, scientific notation for large/small numbers, insets common for zoomed regions | | Color | Conservative — fewer colors, more line style variation |

Use CS_STYLE as base, adjust font to Computer Modern.

How to Select a Profile

At project start (Phase 0/1), when target_venue is set in research-anchor.yaml:

1. Read the venue name
2. Map to profile:
   • Nature, Science, Cell, Nature *, Science *, Cell *, PNAS, eLife → CNS profile
   • NeurIPS, ICML, ICLR, CVPR, ECCV, AAAI, ACL, EMNLP, KDD, WWW → CS profile
   • IEEE *, any IEEE transaction or conference → IEEE profile
   • Bioinformatics, Genome Research, Nucleic Acids Research → Life Science profile
   • Physical Review *, ACS *, RSC *, J. Chem. Phys. → Physical Sciences profile
3. If venue is unclear or not listed, ask user: "Which figure style matches your target venue? (1) CNS/Nature style (2) CS conference style (3) IEEE style (4) Other — please describe"
4. Write the selected profile to src/plot_style.py and use it for ALL figures

Figure Type Selection Guide

| Data type | Recommended figure | Avoid | |-----------|-------------------|-------| | Method A vs B vs C on multiple datasets | Grouped bar chart or table | Pie chart | | Performance vs hyperparameter | Line plot with error bands | Scatter without connection | | Ablation (component contribution) | Grouped bar chart or stacked bar | Line plot (components aren't ordered) | | Training dynamics | Line plot (x: epoch, y: metric) with shaded std | Bar chart | | Feature importance / attention | Heatmap with annotated values | 3D plots | | Distribution comparison | Violin plot or box plot | Overlapping histograms | | Embedding visualization | t-SNE/UMAP scatter with class colors | PCA (usually uninformative for high-dim) | | Qualitative examples | Grid of input→output pairs | Random cherry-picked singles | | Architecture diagram | Clean schematic (tikz, draw.io, or programmatic) | Hand-drawn or overly complex | | Confusion matrix | Annotated heatmap with numbers in cells | Plain matrix without annotations |

Style Template

At project start, create src/plot_style.py based on the selected venue profile. This file is imported by every plotting script.

import matplotlib.pyplot as plt
import matplotlib as mpl

# ──────────────────────────────────────────────
# SELECT ONE profile based on target venue.
# See Venue-Specific Figure Styles section above.
# Copy the matching STYLE dict and COLORS list here.
# ──────────────────────────────────────────────

# Example: CS conference profile (NeurIPS, ICML, etc.)
STYLE_CONFIG = CS_STYLE   # Replace with CNS_STYLE, IEEE_STYLE, etc.
COLORS = CS_COLORS         # Replace with NATURE_COLORS, etc.

mpl.rcParams.update(STYLE_CONFIG)

Save as src/plot_style.py and import in every plotting script. This ensures ALL figures have consistent, venue-appropriate style.

Method-Color Mapping

At the start of the project, define a global color mapping and use it everywhere:

METHOD_COLORS = {
    'Ours': COLORS[0],       # Always blue
    'Baseline A': COLORS[1], # Always orange
    'Baseline B': COLORS[2], # Always red
    'Baseline C': COLORS[3], # Always teal
    'Ablation': COLORS[4],   # Always green
}

Store this mapping in src/plot_style.py and update it as methods are added. Never assign colors ad-hoc per figure.

Table Standards

| Property | Requirement | |----------|------------| | Format | booktabs style in LaTeX (\toprule, \midrule, \bottomrule); no vertical lines | | Best result | Bold the best value in each column/metric | | Second best | Underline the second best (if comparing ≥4 methods) | | Uncertainty | Always report mean ± std (or CI); bare numbers without variance are unacceptable | | Alignment | Decimal-aligned numbers; consistent decimal places per column | | Significance | Mark statistically significant improvements (e.g., † or * with p-value in caption) | | Our method highlight | Use light gray row shading or clear label; never bury it in the middle |

LaTeX Table Template

\begin{table}[t]
\centering
\caption{Main results on [datasets]. Best in \textbf{bold}, second best \underline{underlined}.
$\dagger$: statistically significant improvement over best baseline ($p < 0.05$, paired t-test).}
\label{tab:main}
\begin{tabular}{@{}lccc@{}}
\toprule
Method & Dataset A & Dataset B & Dataset C \\
\midrule
Baseline 1 & $83.2 \pm 0.4$ & $76.1 \pm 0.8$ & $91.3 \pm 0.2$ \\
Baseline 2 & $\underline{85.1 \pm 0.3}$ & $77.4 \pm 0.6$ & $\underline{92.0 \pm 0.3}$ \\
Baseline 3 & $84.7 \pm 0.5$ & $\underline{78.2 \pm 0.5}$ & $91.8 \pm 0.4$ \\
\midrule
Ours & $\mathbf{87.3 \pm 0.2}^\dagger$ & $\mathbf{80.1 \pm 0.4}^\dagger$ & $\mathbf{93.5 \pm 0.2}^\dagger$ \\
\bottomrule
\end{tabular}
\end{table}

Per-Figure Quality Checklist

Before including ANY figure in the paper, verify every item.

Figure file (the image itself):

• [ ] Venue profile applied — using the correct style from src/plot_style.py?
• [ ] Readable at print size — scale to final column width; all text ≥ 7pt (CNS: ≥ 5pt)?
• [ ] Axis labels present — with units (e.g., "Accuracy (%)", "Time (s)")?
• [ ] No figure title — title belongs in LaTeX \caption{}, not on the figure
• [ ] Legend present — if multiple series; not overlapping data?
• [ ] Panel labels — if multi-panel: bold a, b, c (CNS) or (A), (B), (C) (CS/IEEE) top-left?
• [ ] Color consistent — same method = same color as all other figures?
• [ ] Colorblind safe — distinguishable without color (markers, line styles, hatching)?
• [ ] Error bars / variance — shown where applicable (shaded region or error bars)?
• [ ] Vector format — PDF/SVG for plots (not PNG/JPG)?
• [ ] No chartjunk — no 3D effects, no excessive decoration, no rainbow gradients?
• [ ] White space optimized — tight layout, no giant margins?

LaTeX side (in the .tex file):

• [ ] \caption{} self-contained — reader understands figure from caption alone, without reading body text?
• [ ] \caption{} describes each panel — for multi-panel: "(a) ... (b) ..."?
• [ ] \caption{} states key takeaway — not just "Results on Dataset X" but what the results show?
• [ ] \caption{} notes statistical details — "Error bars: ± 1 std over 5 seeds" or similar?
• [ ] \label{fig:xxx} present and descriptive?
• [ ] \includegraphics width — matches venue column width (\columnwidth or \textwidth)?
• [ ] Referenced in text — every figure is \ref{}'d in the body text; no orphaned figures?

Anti-Patterns — NEVER Do These

| Anti-pattern | Why it's bad | What to do instead | |-------------|-------------|-------------------| | Default matplotlib style (white bg, thin lines, small fonts) | Unreadable at print size | Apply the style template above | | Rainbow colormap for categorical data | Perceptually nonlinear, colorblind-hostile | Use qualitative palette (Tableau 10, ColorBrewer) | | 3D bar charts or pie charts | Distort proportions, waste ink | 2D grouped bar chart | | Inconsistent colors across figures | Reader can't track methods | Global method-color mapping | | Screenshots of terminal output | Unreadable, unprofessional | Proper table or formatted code block | | Figures without error bars | Results look unreliable | Always show variance (std, CI, min-max) | | Tiny axis labels that need zooming | Will be illegible in print | ≥ 8pt at final size | | Cherry-picked qualitative examples | Misleading | Show representative range (good + average + failure) |

Integration with Workflow

This skill should be invoked:

1. Phase 4 (experiment-execution): When generating any experimental figure — apply style template, use method-color mapping
2. Phase 5 (results-integration): When producing the content outline — verify all planned figures meet standards; run the per-figure checklist
3. Phase 6 (paper-writing): Before including each figure in a .tex file — final quality check; verify LaTeX \includegraphics path and caption

Rationalization Prevention

| Excuse | Reality | |--------|---------| | "I'll fix the figures later" | You won't. Style issues compound. Apply the template from the first plot. | | "Default matplotlib looks fine" | On screen at 100%, maybe. At conference poster or PDF zoom, it's unreadable. | | "Color doesn't matter" | 8% of men are colorblind. Reviewers print in B&W. Color always matters. | | "Error bars clutter the plot" | Error bars ARE the data. Without them, your plot is a lie. | | "One quick plot is fine for now" | Quick plots become final figures 90% of the time. Do it right the first time. |