Thinking-Mode Infographics — Four Dense Visuals, One System

Four information-rich infographics — an LLM training explainer, a cost-vs-quality frontier chart, an academic poster, and a system-diagram of an AI agent runtime — all produced by Iris from single prompts of ~600–1200 words. The technique follows OpenAI's Yuguang Yang demonstration that GPT Image 2 at quality=high faithfully renders precise text, numbers, equations, legends, and layout constraints when given long, disciplined briefs.

Technique

Long-prompt "Thinking"-style briefs

Palette

Ink navy · Cyan · Magenta · Amber · Paper

Grammar

12-col grid · numbered sections · monospaced data

Deliverables

4 infographics · 4 formats

01 — How Large Language Models Learn

Portrait 9:16 educational explainer. Three stages (pre-training → SFT → RLHF) with real equations and stat tiles.

Infographic: How LLMs learn — pre-training, SFT, RLHF

How LLMs Learn 9:16 · 1152×2048

Portrait explainer with a headline block, three numbered panels (Pre-training, SFT, RLHF), precise stat tiles (tokens seen, FLOPs, loss curve, reward-model equation), legend and source footer. ~600 words of structured constraints.

GPT Image 2 · thinking WebP · 9:16 · 1152×2048 prompt ≈ 3,480 chars ~$0.19

02 — The Cost–Quality Frontier, 2026

Landscape 16:9 data visual. Log-log scatter plot with eight labelled model points and a drawn efficient-frontier line.

Infographic: frontier LLMs plotted on cost vs. quality, log-log scale

Cost × Quality Frontier 16:9 · 2048×1152

Scatter plot on log-log axes with exact dot placements and labels for eight frontier models (GPT-5, Claude 4.7 Opus/Sonnet, Gemini 2.6 Pro/Flash, Llama 4 405B, Mistral Large 3), a dashed efficient-frontier overlay, three stacked takeaway cards, legend, and data-source footer.

GPT Image 2 · thinking WebP · 16:9 · 2048×1152 prompt ≈ 3,265 chars ~$0.19

03 — Academic Poster & 04 — Agent Runtime Diagram

A portrait conference poster (scaling-laws paper) next to a landscape 9-step process diagram of an AI agent's request cycle.

Infographic: academic conference poster about scaling laws

Scaling-Laws Poster 9:16 · 1152×2048

A0-style conference poster: banner with authors, two-column body with Abstract, Setup, Loss-scaling figure, Data-vs-parameters isocurve, stat tiles (best val-loss, compute-optimal D/N, inference cost), Discussion, References, and a QR-style square in the footer.

GPT Image 2 · thinking WebP · 9:16 · 1152×2048 prompt ≈ 3,172 chars ~$0.19

Infographic: 9-step AI agent runtime diagram

Agent Runtime in 9 Steps 16:9 · 2048×1152

Horizontal flow with 9 numbered nodes (prompt → context → tokenizer → forward pass → plan → tool pick → tool exec → observation → response), a return-loop arrow, a timing strip with real per-step latencies summing to ~2.2s, legend, and footer.

GPT Image 2 · thinking WebP · 16:9 · 2048×1152 prompt ≈ 3,033 chars ~$0.19

05 — Mind the Gap · London Underground Poster

Tribute to Harry Beck's diagrammatic grammar — 45°/90° geometry, canonical TfL line colours, line legend with opening years and track length, and a platform-edge "Mind the Gap" callout. A different visual system (Johnston-style sans, cream paper) to prove Iris isn't locked to one house-style.

Infographic: London Underground Harry Beck-style diagrammatic map with line legend, network stats, and Mind the Gap callout

Mind the Gap — London Underground 16:9 · 2048×1152

Beck-style diagrammatic tube map on cream paper with exact TfL line colours (Bakerloo brown, Central red, Circle yellow, Piccadilly dark blue, Victoria light blue, Elizabeth purple, …), 10 named interchanges, Zone 1 ring. Right half: 12-row line legend with opening year + length, network-at-a-glance tiles (272 stations · 402 km · 1.35bn journeys), a yellow "MIND THE GAP" platform-edge callout, and an Oyster fare note. Johnston-style humanist sans throughout.

GPT Image 2 · thinking WebP · 16:9 · 2048×1152 prompt ≈ 3,700 chars ~$0.19

Method Notes — Iris

One shared system block. Every prompt starts with the same ~200-word VISUAL SYSTEM — palette hex codes, type grammar, grid rules, icon style, legend placement, source footer. That's how four very different infographics end up looking like siblings.
Write the numbers, don't hint at them. Axis ticks, equation symbols, stat tiles, latency figures — every digit that appears on the canvas is spelled out in the prompt. Image 2 reproduces them crisply at quality=high.
Lay out in words. Each prompt uses numbered panels, percentages of canvas height, and explicit positions ("bottom-right", "top third", "left column"). The model treats the prompt like a brief, not a vibe.
Anchor the equations. For the RLHF panel, the prompt includes the preference-learning loss verbatim. For the scaling-laws poster, the Chinchilla isocurve and α-exponent labels are written out in text. No equations are generated from a vague hint.
Cohesion, then variety. Portrait-explainer, landscape-chart, tall-poster, landscape-diagram — four native aspect ratios so the showcase proves format fluency as well as information density.

Source / inspiration: Yuguang Yang, OpenAI — Slides & Infographics with ChatGPT Images 2.0 (video).

"One of the standout strengths of Image 2 is that it can follow very long and detailed instructions that include precise text and numbers, equations, and technical terms, layout constraints, legends, color and style requirements."

Iris took that claim seriously: four briefs averaging ~3,200 characters each, every figure written out, every layout hand-specified.

Takeaway: Infographics aren't a new model capability — they're a prompt discipline. Write the brief like a real design spec, lock the palette and grid across the set, and Image 2 will render the equations, the ticks, the legend, and the layout exactly as specified. Iris now has this as a repeatable play for papers, investor decks, board reports, and explainer blogs.