An icon design workflow with AI in it
Icons are the hardest thing to generate well, because consistency across a set matters more than the quality of any single one. Generation helps at the start of that process and gets in the way at the end.
AI Tools Hub Editorial · published · updated
Where it helps
- Exploring metaphors — twenty visual approaches to a concept in minutes
- Breaking a block on a difficult concept where the obvious metaphor is taken
- Producing a rough set to test at size before committing to drawing
Where it hurts
- Consistency across a set — generated icons drift in weight, corner radius and optical size
- Grid and pixel alignment, which generation does not respect
- Anything that needs to read at 16 pixels
A loop that works
| Step | Tool | Output |
|---|---|---|
| Explore metaphors | Midjourney or Ideogram | Twenty rough directions |
| Pick three | You | The ones that survive at small size |
| Generate vectors | Recraft | Editable SVG in a consistent style |
| Normalise | Figma or Illustrator | Grid, weight, optical sizing |
| Extend the set | Recraft, with the finished ones as reference | The remaining icons |
Recraft matters here because it outputs real vectors. Generating raster icons and tracing them produces the inconsistency you are trying to avoid.
The test that matters
Put the whole set at 16, 24 and 32 pixels on one artboard. Inconsistency that is invisible at 200 pixels is obvious there, and that is the size users see.
Can AI produce a complete icon set unaided?
Not one that holds together. The individual icons will be fine and the set will not be consistent, which is the only property that matters.
What about using an existing icon library?
Usually the right answer. Generate only when you need something a library does not have, or when the brand requires a distinct style.
Related
More articles
Opinion
Why AI rollouts fail in Japanese companies — four structural reasons
For business
Enterprise AI rollout checklist — SOC 2, ISO 27001, DPAs and personal data
For business
Japanese-developed LLMs for business — ELYZA, Sakana AI, NTT tsuzumi and PLaMo
Comparison
AI coding tools compared — Cursor, Copilot, Claude Code, Windsurf and Codeium