Paper Figure Craft
Create dense, crisp, publication-grade research figures in code (exact-point SVG for schematics and classic four-spine Matplotlib for quantitative evaluation charts) that match top systems, architecture, EDA, and AI venues without generic AI-generated flowchart tells. The Chinese equivalent is SKILL.zh-CN.md.
Core Workflow
-
Combine Toolchain Skills & Split Figure Roles (
1 Figure = 1 Job, Heterogeneous Containers)- Synergize with
figgenie-paper-diagramwhen installed: Referencefiggenie-paper-diagramvenue exemplars for information density, emit semantic attributes (data-kind,data-role,data-symbol="tex"), and run the 5-step pipeline:tex_fill.py --check(MathJax inline vector<path>) $\to$lint.py(containment/overlap checks) $\to$4× PNGvisual check $\to$svg_to_pdf.py/export_pdf.py(zero-rescale vector PDF) $\to$svg2pptx.py(editable slides). Whenfiggenie-paper-diagramis absent, use templates/svg_diagram_scaffold.py + templates/svg_to_pdf.py directly. - Motivation (
Fig. 1, single-column244 × 152–185 pt, Page 1 top-right): Pick the archetype fitting your thesis:- Archetype A (Paradigm Contrast): Top
(a)contrasts 2–3 prior paradigm cards (mini trend curves ++ strength / - limitation) + interlocking challenge chevron barC1–C3; bottom(b)previews your 3-step solution with a1-to-1column-alignedchevron_bar. - Archetype B (Motivating Case Trajectory & Convergence):
(a)Concrete benchmark/hardware setup,(b)single-line step trajectoryE1..Ekhighlighting rapid overall convergence, and(c)convergence curve paired with explicit colored square swatches (swatch_legend_row:DF: WS / OS,GLB: 2M / 8M / 32M).
- Archetype A (Paradigm Contrast): Top
- Never put heterogeneous content in identical boxes: Differentiate static setups, sequential step rows, directed edges, verified fact ledgers, and agent hypotheses visually, keeping copy to crisp single-line phrases.
- System Overview (
Fig. 2, double-column504 × 180–192 pt): Closed-loopSteps 1–6with numbered black circle badges, fast-bypass vs. on-demand routing, and standalone output artifact. - Worked Mechanism / Tool Loop (
Fig. 3,244 ptor504 × 146–160 pt): Concrete numerical walkthroughs (with embedded candidate leaderboard mini-tables highlightingPickand mini scatter plots) or binder-ring verified ledger $\to$ agent hypothesis speech bubble $\to$ sampling tool strip. - Quantitative Evaluation Charts (
Figs. 4–6, Matplotlib vector PDF/SVG): Dual-row search trajectories, grouped hatched bars, stacked runtime breakdowns, and Pareto scatter plots.
- Synergize with
-
Enforce Layout Micro-Grammar, 4-Layer Palette Hierarchy & 5-Category Micro-Iconography Read references/schematic-grammar.en.md (or schematic-grammar.zh-CN.md) and build on templates/svg_diagram_scaffold.py + templates/plot_style.py:
- Exact-point interlocking grid & typography: Fix
W = 504 pt(\textwidth) or244 pt(\columnwidth) with1 pt = 1 SVG unit;3–4 ptouter margins,4.5–5.5 ptinter-stage gaps bridged by outlinedblock_arrow_h,3.5–4 ptinner padding. Use centered white knockout banners (stage_container) or light dashedoutline_zone+2.8 ptleft_stripe_card. Reserve Italics for[Bracketed Role Tags], branch conditions, descriptor fields, and 1-line physical mechanisms, keeping font weight uniform across wrapped list lines. - 4-layer mineral-ink palette (
plot_style.pyshared across SVG & Matplotlib): Layer 1 structural ink#18181B+#3F4752/#525E6Bgreys +#C9D1DBblock arrows; Layer 2 muted architectural stage grounds (#F4ECDC,#DAE8F4,#DDE2E8,#F3DFCE,#DCEAD7); Layer 3 crisp#FFFFFFforeground cards +#E2E7EEtabs (#E6EFEAmint for winningPickrows); Layer 4 high-contrast accents (#1B4965Navy,#A61C24Crimson,#C46D1FOchre,#2C6E58Sage) capped under<10%area—never flooding large container backgrounds. - 5-category vector micro-iconography: (1)
13.5×13.5 pthardware/workload thumbnails (icon_systolic,icon_noc,icon_sparse); (2) dataset drums (icon_drum) & left-edge notebookbinder_ringssignaling verified facts; (3) role-badged agents (icon_agent: magnifying glass for diagnosis vs. sampling grid for experiment) & pointedspeech_bubblepolygons for testable hypotheses; (4)icon_wrench&icon_stacked_docsfor standalone artifacts; (5)icon_refresh_loop&swatch_legend_row. - Distinguish nodes vs. directed edges & state explicit units: Draw interfaces/filters as pill-labeled arrows (
directed_edge_with_pill) and write0 LLM Calls,0 Sim Calls,9.2s (3.6× faster).
- Exact-point interlocking grid & typography: Fix
-
Enforce Classic IEEE/ACM Quantitative Chart Conventions Read references/eval-charts-ieee.en.md and start from templates/ieee_eval_plot_scaffold.py:
- Four-spine box + inward ticks +
Arimo + STIXType-42 fonts: Enable all four spines (0.8 pt), inward ticks, andpdf.fonttype = 42. - Grayscale-safe hatches & dual step-hold trajectories: Hatch baseline bars (
'//','\\\\','..'), render Ours in solid primary accent with direct numeric labels, and plot dual wall-clock / evaluation-count trajectories viaax.step(..., where='post'). Plot only universal baselines in grouped cross-domain bar charts.
- Four-spine box + inward ticks +
-
Render, Lint, and Visually Inspect Every Figure via Screenshot
- Confirm
>= 3 ptinline math clearance, baseline<tspan dy>resets, and 1:1 terminology sync with manuscript prose. - Visually inspect high-DPI PNG/PDF renders to eliminate reference-line label collisions, cramped text, or legend overlaps.
- Confirm