Neural Architecture Designer: A Python Generator and Renderer for Publication-Ready Neural Network Architecture Diagrams

Abstract

Neural-network architecture diagrams are widely used to communicate model topology, layer roles, dimensional transitions, information pathways, and methodological design. However, publication figures are still often assembled manually or produced by model-inspection systems whose visual objectives differ from those of a scientific schematic. This paper presents Neural Architecture Designer, a Python-based generator and geometry-aware renderer for publication-ready neural network architecture diagrams. The rendering kernel treats an architecture as a declarative layered graph and separates topology from graphical style. It supports heterogeneous node glyphs, mathematical labels, explicit activation blocks, input/output item boxes, dense, sampled, and adjacent-only connectivity, pair-specific connection overrides, straight and cubic Bézier links, line and arrow modes, shape-aware endpoint trimming, surface-referenced gaps, signed outside-junction offsets, deterministic palette assignment, seeded line-width variation, annotations, ordinal path highlighting, and cascade fan-out highlighting. Endpoint geometry is resolved in display space only after the final axes transform is fixed, which preserves visible node-boundary offsets across figure dimensions and export targets. The article derives the principal geometry and connection equations, presents the corresponding Python code, and places representative vector outputs beside the mechanisms they illustrate. It additionally serves as a practical application guide: six interface screenshots document the workspace, layer editor, group-header editor, annotation editor, connection controls, terminal arrows, input/output boxes, project workflow, preview, auto-fit, and export operations. A standalone reconstruction kernel and a complete rendering-only reference implementation are included so that the figure-generation method can be reproduced independently of the desktop control layer. Version 1.0 is publicly released under the MIT License, with source code hosted on GitHub and the archived software release preserved on Zenodo under DOI 10.5281/zenodo.22135298.