What problem does it solve? Creating architecture diagrams and method illustrations that meet CVPR/NeurIPS visual standards is time-consuming, and AI-generated figures often contain wrong arrow directions, missing labels, or unprofessional styling that requires many manual correction rounds. ## Core Features & Use Cases - Multi-Stage Generation Pipeline: Claude plans the figure, Gemini (gemini-3-pro) optimizes layout and verifies style compliance, then Paperbanana (gemini-3-pro-image-preview) renders the final image. - Strict Iterative Review Loop: Claude visually inspects every generated figure against a detailed checklist (arrow directions, block content, labels, visual appeal), scores it 1-10, and refines until the score reaches 9 or the 5-iteration limit. - Optional Style Reference: Pass a reference paper (TeX, PDF, arXiv ID, or URL) via --style-ref to align caption length and figure density conventions without copying visual content. - Use Case: A researcher needs a model architecture diagram for a CVPR submission; the skill generates a color-coded, print-friendly figure with thick labeled arrows and outputs a ready-to-use LaTeX snippet. ## Quick Start Ask the assistant to generate an architecture diagram of your model pipeline for a CVPR paper, describing the components and data flow you want shown.
