VINNT

Visual Illustration of Neural Networks in Typst

user Manual · Gallery · Examples

Layered neural network architectures, drawn as isometric block diagrams — the kind that open a paper's method section. Built on CeTZ.

#import "@preview/vinnt:0.2.0": *

Requires Typst 0.15 or newer. Not yet on Typst Universe so you will have to point the import at a local copy: #import "path/to/src/lib.typ": *.

---

Start here

#draw-network((
  conv(label: "a"),
  pool(),
  conv(label: "b"),
))

Blocks are sized from their tensor shapes

State what each layer produces and the geometry follows. Change a channel count and the figure changes with it — there is no second copy of the pyramid to drift out of step.

#draw-network((
  conv(shape: (64, 128, 128), channels: (64, 128)),
  conv(shape: (128, 64, 64), channels: (128, 64)),
  conv(shape: (256, 32, 32), channels: (256, 32)),
  conv(shape: (512, 16, 16), channels: (512, 16)),
))

Spacing, lane heights and arrival points are all computed too. Nothing above sets an offset.

Skips, sums and group brackets

#draw-network(
  (
    conv(name: "in", label: "in", widths: (0.3,)),
    convres(name: "c1", label: "3x3", widths: (0.5,)),
    convres(name: "c2", label: "3x3", widths: (0.5,)),
    sum(name: "add"),
    conv(name: "out", label: "out", widths: (0.3,)),
  ),
  connections: (
    (from: "in", to: "add", color: rgb("#466A9F"),
     legend: "identity shortcut"),
  ),
  groups: ((from: "c1", to: "add", label: "residual block"),),
  show-legend: true,
)

Encoder–decoder

Routes land on the block itself with touch-layer, and stack into lanes by how far they reach, so a longer skip always arcs over a shorter one.

Genuinely parallel branches

Two-stream fusion, CSP interiors and multi-head detectors are drawn as they are, not collapsed into one block with arrows pointed at it. Branches nest, and may be open at either end.

#draw-network((
  branch(spread: 11, branches: (
    (input(label: "RGB", image: "default"), conv(shape: (32, 160, 160))),
    (input(label: "IR"), conv(shape: (32, 160, 160))),
  )),
  concat(label: "concat"),
  conv(label: "fused", shape: (128, 80, 80)),
))

Fourteen layer types

Plus custom, which is the generic block you extend when you need something the package has no name for:

#let attention(..a) = custom(fill: rgb("#466A9F"), width: 0.35,
                             legend: "Attention", ..a)

A whole classifier

Neuron diagrams

The other canonical figure — a circle per neuron, an edge per weight — is draw-mlp, and zero configuration is publication-ready:

#draw-mlp((4, 6, 6, 3))

Per-layer control is constructors, like everything else in the package:

#draw-mlp((
  mlp-layer(4,  label: "input"),
  mlp-layer(8,  label: "h1", activation: "relu"),
  mlp-layer(8,  label: "h2", activation: "relu"),
  mlp-layer(3,  label: "output", activation: "softmax"),
), bias: true, cutoff: 6)

Wide layers collapse to an ellipsis with the true count badged beneath. Also in the box: weight-driven edge color and thickness with negatives dashed by default (explicit matrices, a function over indices, or seeded random), an activation glyph catalog with the softmax bracket, skip and recurrent edges with ⊕ merge nodes, bias nodes, square and split node shapes, dropout and highlight node states, and a bottom-to-top direction: "up".

---

Typos are errors, not silence

error: unknown layer option "hieght" on layer 3 (type "conv").
Did you mean "height"? Options accepted here: bandfill, channels,
connection-label, depth, fill, height, image, label, ...

Every option is checked against what that layer type actually reads, including connection and group options and any name a connection points at. A key that is quietly ignored produces a figure that is merely wrong, which reads as the package being broken rather than as the typo it is.

Import a model instead of drawing one

uv run tools/import_model.py --torchvision resnet18 -o resnet18.json

#draw-network(from-shapes(json("resnet18.json")))

---

Documentation

The manual — 98 pages, 176 figures. Every option on its own, several figures each, with the failure shown next to the fix. Every figure in it is compiled from the code printed beside it. Build it with doc/build.sh.

Writing VINNT with an LLM? Point it at doc/manual.typ and doc/examples/ rather than at the PDF. That is the manual's actual source: plain text throughout, and every example a standalone file that compiles on its own.

Gallery — AlexNet, LeNet-5, VGG16/19, ResNet18, U-Net, FCN-8, SynthMorph, YOLO26-n and five RGB-IR fusion variants, plus six MLP figures in gallery/mlp/. Sources in examples/.

• YOLO26-n
• YOLO26-n — early fusion
• YOLO26-n — mid fusion
• YOLO26-n — late fusion
• YOLO26-n — gated fusion
• YOLO26-n — multiscale fusion
• SynthMorph

Licence

MIT-0.

---

All due credit to prior work is provided in ACKNOWLEDGEMENTS.md.