How Canva Designs AI for Suggestion Over Replacement
TRIGGER
Image similarity model produced strong results for photos but weak results for graphics and text-heavy images—waiting for perfect accuracy would delay value, but fully automated replacement would damage template quality.
APPROACH
Canva's Content Enrichment Team built a reverse image search system using DINOv2 embeddings for their 150M+ image library, stored in an external vector database with metadata filtering (aspect ratio, IP-safety). Input: image flagged for replacement in Template Assistant. Output: top 8 visually similar images ranked by subject, color, tone, and positioning. Designers select a suggestion, browse alternatives, or bypass to manual search; selections go through human review before republishing. DINOv2 performed strongly on photos but weakly on graphics, text-heavy images, and cartoons due to training data composition. Result: 4.5x speed increase for image replacement compared to manual search, capturing value from photos while designers handle edge cases without workflow disruption.
PATTERN
“When model accuracy varies by content type, ship suggestions to domain experts instead of waiting for perfection. You capture value from high-accuracy cases today while humans handle edge cases without workflow disruption.”
✓ WORKS WHEN
- Model accuracy varies predictably by content type or category
- Domain experts are already in the workflow (designers, reviewers, moderators)
- Speed improvement on high-accuracy cases justifies integration effort
- Bypass path exists for low-accuracy cases (fallback to manual search)
- Quality bar requires human approval before publishing
✗ FAILS WHEN
- Fully automated processing required (no human reviewers available)
- Model accuracy is uniformly low across all content types
- Suggestion latency disrupts workflow more than manual search
- Edge cases are the majority of volume, not the minority
- Users lack expertise to judge suggestion quality