mlatlas: Building Declarative Machine Learning Architecture Diagrams with Typst

mlatlas is a Typst library designed for creating declarative machine learning architecture diagrams. Developed by j-vaught and hosted on GitHub (released on 2026-05-31), it addresses key pain points in ML visualization by allowing users to describe architectures via code, which the compiler then turns into charts. Core benefits include solving manual adjustment hassle, enabling easy version control, ensuring style consistency, and integrating deeply with technical documents.

Original source: mlatlas GitHub repo

In ML, clear architecture diagrams are crucial for communication, but traditional tools have several issues:

• Manual adjustment: Nodes and lines require tedious drag-and-drop alignment.
• Version control: Binary image files are hard to track changes for.
• Style inconsistency: Different people’s diagrams vary, reducing document professionalism.
• Doc separation: Charts are separate from text, making updates cumbersome.

With Typst’s rise as a modern typesetting system, there’s a need for native code-generated chart solutions—mlatlas fills this gap.

mlatlas offers these core features:

1. Declarative syntax: Use concise Typst code to define layers, connections, and styles.
2. High contrast rectangle style: Optimized for technical docs, clear in print and on screen.
3. Built-in ML components: Pre-made elements like fully connected layers, conv layers, attention mechanisms.
4. Full customization: Adjust node shapes, color themes, layout parameters.
5. Deep doc integration: Charts are part of Typst code, supporting version control, automatic typesetting, cross-references.

mlatlas uses Typst’s drawing primitives (box, line, place) to build charts. The workflow is:

1. Import the mlatlas module in your Typst doc.
2. Define layers (input, hidden, output) using preset functions.
3. Declare connections between layers—layout is auto-calculated.
4. Apply preset themes or custom styles (colors, fonts, spacing).
5. Embed the chart directly into the doc for seamless integration.

This turns charts into code assets, simplifying maintenance.

mlatlas is useful in: Academic papers/reports: Neural net architectures, data flow processes, model comparisons, attention mechanism visualizations. Technical docs/tutorials: API architecture, system module relationships, training flows, deployment diagrams. Demos & teaching: Lecture notes, workshop slides, online course materials.

mlatlas leverages Typst’s strengths and outperforms traditional tools:

Feature	TikZ (LaTeX)	draw.io	mlatlas (Typst)
Learning curve	Steep	Gentle	Medium
Version control	Friendly	Difficult	Friendly
Compile integration	Native	Need export	Native
Modern syntax	No	N/A	Yes
ML-specific components	Need extensions	Manual draw	Built-in

It fills the ML visualization gap in the Typst ecosystem.

mlatlas represents a shift toward 'docs as code'—integrating visual content into code management. For ML practitioners, this means focusing more on content than tooling.

As Typst’s ecosystem matures, more specialized libraries like mlatlas will emerge, lowering the barrier to creating high-quality technical docs.