ML paper figure stack: from architecture to flow to sketch

An ML or systems paper has roughly four kinds of figures, each with a different formality register. Trying to do them all in one tool produces visual fatigue. This stack assigns one skill per kind.

Step 1: Architecture diagram (formal, ML-aware)

Use drawio-diagram by xstongxue. Output is .drawio XML — opens in app.diagrams.net, editable forever. The skill is ML-aware: knows how to depict attention heads, conv layers, residual connections, and normalization in their conventional shapes.

If you want the visual style to match a specific reference paper, use style migration: give it a screenshot of the reference figure, then your content. Output matches the reference visual style.

Step 2: Algorithmic flow / pseudocode-adjacent

Use mermaid-tools. Write the flow in Mermaid syntax inside your Markdown draft, render to PNG when you compile to PDF. The output is tidy, technical, and reproducible — if a reviewer asks you to re-arrange a step, you re-render in seconds.

Step 3: Conceptual sketch (introduction)

The “what’s wrong with the existing approach” illustration in the introduction has a different job than the architecture diagram. It should signal thinking out loud, not finished system. Use excalidraw-diagram — its hand-drawn vibe makes the figure read as conceptual rather than definitive.

Step 4: Hero / cover figure

For papers that go on arXiv with a hero image (increasingly common for ML papers wanting Twitter circulation), use canvas-design. One editorial composition with your paper’s title, your method’s signature visual, and a single accent color. This is what people share when they tweet your paper.

Three different vocabularies, one paper

The point is that each figure signals what tier of formality applies to that section. Architecture (formal) and conceptual sketch (exploratory) need different tools because they communicate different levels of confidence.

Variations

Systems paper — replace step 1 (drawio for ML) with drawio’s general architecture mode (still works, just less ML-specific shapes).
Theoretical paper — drop step 4 (no hero image), heavier on step 2 (more flow / pseudocode).
Slides version of the paper — pipe outputs into pptgen-drawio which natively imports .drawio files.

What this pipeline doesn’t do

3D model visualization — none of these handle 3D. You’ll need custom Three.js or Manim.
Interactive figures — for online supplements, write a separate interactive HTML using web-artifacts-builder.

This stack covers the standard ML/CS paper visual vocabulary in 2026.