Stop Drawing Neural Nets by Hand

What it does

NN-SVG is a browser-based tool that produces neural network architecture diagrams as scalable vector graphics. Instead of manually laying out nodes and edges in a drawing program, researchers adjust parameters—layer sizes, colors, layout—to generate figures ready for academic papers or teaching materials. It handles three distinct visual conventions: classic fully-connected networks, LeNet-style convolutional diagrams, and AlexNet-style deep network schematics.

The interesting bit

The project treats neural net diagrams as a formal visual language, codifying the exact styles of seminal papers rather than inventing a generic look. FCNN and CNN figures render via D3, while the AlexNet-style 3D blocks use Three.js, giving each diagram type the specific visual grammar researchers expect.

Key highlights

• Exports to SVG for lossless inclusion in LaTeX documents or web pages
• Parametric controls for size, color, and layout remove the need for manual illustration
• Supports three canonical styles: FCNN, LeNet-style CNN, and AlexNet-style deep networks
• Published in the Journal of Open Source Software (JOSS)
• Runs in the browser via D3 and Three.js

Caveats

• Limited to three specific diagram styles; modern architectures like transformers or graph neural networks are not mentioned in the documentation
• Appears to require manual parameter entry for each layer rather than importing model definitions directly from deep-learning frameworks

Verdict

Machine learning researchers who need clean, consistent architecture figures for papers or lectures should bookmark this; deep-learning practitioners looking to auto-visualize arbitrary model checkpoints will need to look elsewhere.