What problem does it solve? Creating paper-ready architecture diagrams and method illustrations for academic submissions is slow and iterative, and generic image generators often produce slide-style figures with wrong arrows, unreadable labels, or decorative styling unsuitable for CVPR/NeurIPS papers. ## Core Features & Use Cases - Multi-stage iterative workflow: Claude plans the figure, optimizes layout, verifies academic style, renders through the local Codex app-server MCP bridge, then strictly reviews and scores the output (target score 9/10, up to 5 refinement rounds). - Academic style enforcement: Built-in CVPR/ICLR/NeurIPS style guide covering palettes, arrow standards, typography, and grayscale readability. - Deterministic artifact pipeline: A helper script runs preflight checks, finalizes the accepted figure to figure_final.png, emits a LaTeX include snippet and review log, and verifies all artifacts. - Use Case: Given a method section describing an encoder-decoder pipeline, produce a publication-ready architecture figure with correct data-flow arrows and English labels, plus a ready-to-paste LaTeX figure block. ## Quick Start Ask the assistant to generate a paper-ready architecture figure for your method description and save it under figures/ai_generated/ with a LaTeX snippet.
