Composite figure prompts
Multiple parts in your content? Write composite paper-figure prompts for the premium look.
The premium look of real paper figures does not come from infographic posters (title bar + numbered modules + conclusion strip) — it comes from composite figures, i.e. Figure 2 in a proceedings paper: a main panel (a) plus dashed-border sub-panels (b)(c) on the right, no title bar, no conclusion strip, and each sub-panel showing a different type of real chart.
When to go composite
Count the parts of your content: multiple method stages, several experiments, architecture + results, mechanism + diagnostics + trends — if you can list two or more, go composite. Each part becomes a panel; do not squeeze them into numbered modules of one poster.
How to describe it (copy-paste template)
A composite research paper figure (like Figure 2 in a CVPR proceedings paper),
landscape 16:9, white background, NO title bar, NO banner, NO conclusion strip.
Layout: left 62% = main panel (a); right 38% = two stacked dashed-border
sub-panels (b) and (c), labels inside each panel's top-left corner.
(a) <method/pipeline topic>: <input> → <stage 1, real parameters> → <stage 2> →
<output>; add a zoom-in detail box connected by thin lines, embed the
formula <spell out every symbol>.
(b) <experiment comparison>: line chart with legend and axis labels, mark the
key values <real numbers>.
(c) <final comparison vs baselines>: bar chart with values and significance;
below it a one-row mini table: <params / FLOPs / training time>.
Use a different chart type per panel; every number must be real.
Use a semantic multi-hue palette: one hue per functional role (task/phase tags = coral red, model/processing modules = pastel green, evaluation/training = amber, metric chips = sky blue, raw data = pale pink); keep the same role the same colour — never one single blue-grey hue.A single sentence — "a composite research paper figure with (a)… (b)… (c)…" — still produces the composite layout, but testing shows two costs: the pipeline adds a title bar on its own (which the template forbids) and may invent values or attach them to the wrong model. For a figure you can put in a paper, spell out every panel's chart type and real numbers as in the template above.
Content part → panel type
| Your content | Suggested panel |
|---|---|
| Method / pipeline | Main panel (a): dashed groups + box chain + zoom-in + embedded formula |
| Training / convergence curves | Line chart (legend + axis labels + key-point values) |
| Final comparison vs baselines | Bar chart (values + crown/star marks + significance brackets) |
| Ablations / group statistics | Grouped bars or a table row (state p / HR / CI) |
| Category shares / filtering funnel | Donut / funnel chart |
| Params / FLOPs / cost | One-row mini table |
| Loss / metric formulas | Standalone formula block (spell out every symbol) |
Three key details
- One chart type per panel — never fill panels with duplicate chart types.
- Only use numbers that actually appear in the paper (e.g. 3.57% top-5 error, BLEU 28.4); invented values are instantly recognizable.
- Write every module title, formula and label verbatim — the image model will not invent structure you did not spell out.
- Never one flat hue: in flow/architecture figures assign a hue per role (tags red, modules green, evaluation amber, indicators blue) — a single blue-grey palette is the number-one tell of AI-generated figures.
Live examples: the residual network, translation model and UGC face recognition figures in the homepage gallery all use the composite layout — use them as style targets for your prompts.