Finding the perfect visual elements in canva to match your design is harder than it should be.
Users had to search and endlessly scroll, hoping to spot something that matched their design. There was no easy way to find similar graphics, apply consistent styles, or keep things cohesive without manual effort.
AI-powered content reimgained how people find content, making discovery faster, more intuitive, and in tune with their design.
3 months
Time
3+
User Tests
CWT stage reveal
Status
my role
- craft lead across all ai powered initiatives within the edtior
- helped frame the vision and early prototypes for magic styles, and design awareness
- led and coordinated research planning and testing
- collaborated with ml engineers on feasibility and edge cases
- built fully working prototype for early team testing and feedback
how it works
magic styles
apply the visual style of one element to others — aesthetic, color, texture, subject — instantly.
design aware filtering
tap a style to see relevant elements across categories. speeds up discovery and visual cohesion.
magic recommendations
context-aware suggestions that adapt based on your layout and selected elements.
smart collections
ai-curated collections designed to match your intent and design context.
on demand generation
generate visual options directly in the editor — no switching panels.
Design awareness everywhere
cursor affordances across different stye transformations
inline generation based on design context
cursor updates with style attributes
generative fallback + eval explorations
what I learned
implicit assumptions
zero-shot generation often misses the mark. being transparent about model assumptions helps manage users expectations and help guide better results.
control comes before delight
ai generated content only works if it feels predictable and in the user's control.
style is subjective, systems aren't
clear defintions and ways of extratcing intent were needed to help show ai how to interpret and apply style.
placement shapes perception
where and when ai appears matters as much as what it produces.
speed beats polish
rapid iteration with real users uncovered insights we didn't consider in the beginning.