Canva’s Craft image model gets mixed results in Business Insider test

Canva describes Craft as its first foundation model built and trained from scratch by Canva Research.[1] The company says the model is available within Canva and to developers through a public API.[1] Canva says Craft generates images at 2K resolution, supports more than 30 aspect ratios, and offers precise control and editable outputs.[1] Those specifications point to a tool intended for both in-product creation and developer access, but they do not establish how well it will fit a particular team’s workflow.

Canva says Craft uses a two-stage architecture, planning image composition before rendering.[1] In that planning stage, the model identifies image elements, spatial relationships, relative scale, visual hierarchy and text, according to the company.[1] Canva also says Craft can generate text in 55 languages.[1] For teams assessing image tools, those features are relevant to prompts involving multiple objects, layout or in-image copy; the available evidence does not specify how consistently Craft handles those requests in everyday use.

Canva says Craft debuted as DesignArena’s fourth most advanced image-generation model and ranks first for People & Portraits among 87 models.[1] The company also cites a Landscapes & Nature comparison involving 85 models, but the supplied information does not state Craft’s rank in that category.[1] Canva reports that Craft ranks second for Abstract & Patterns and fourth for Products.[1] The company says more than 25,000 image pairs were assessed in head-to-head comparisons with leading models.[1] These are Canva’s reported results, not a guarantee of performance for a specific brief.

Business Insider compared Craft with image-generation models from OpenAI, Google and Meta.[2] Its test found that Craft fell short at making a birthday card, while it was catching up with more powerful models on realistic images.[2] That split is a practical caveat for teams evaluating the model: performance on one kind of image task may not predict performance on another. Canva says its training data includes millions of human-made templates from its content library.[1] It also says model distillation reduced required inference steps to a fraction of the original model’s, and a custom inference engine made Craft an order of magnitude faster than its first internal version.[1] Canva says two upcoming variants are designed to deliver high-quality image generation at a fraction of leading frontier models’ cost.[1]