// Yapay Zeka

Canva released its first in-house image model and made it available to all users

Canva has made Craft, an image generation model trained from scratch by its research lab, available through Dream Lab and a public API. The model uses a two-stage architecture that separates composition planning from drawing.

Updated: 23 October 2026

On October 22, 2026, Canva announced Canva Craft, its first foundation model trained from scratch by its own research lab, Canva Research. The image generation model is now available to all Canva users worldwide through Dream Lab; developers can also access the model via the company’s public API. Canva says that, according to DesignArena rankings, this launch has made it the world’s second most advanced AI lab for image generation.

Plan first, draw second

Craft works differently from models that generate pixels directly from prompts. In the first stage, a visual planning model interprets the prompt and writes a detailed plan for the composition: which elements will appear, their positions relative to one another, their relative scales, visual hierarchy, and the text that will appear in the image. In the second stage, a separate rendering model turns this plan into an image at up to 2K resolution and in more than 30 aspect ratios.

According to Canva, modeling composition and typography before drawing makes it possible to produce complex layouts with readable text and provides more precise control over individual elements. The same structure also forms the basis of Magic Layers, a feature that converts the generated image into editable design layers. Users can write prompts in any language, and the model can generate text in 55 languages. The JSON bounding box feature also converts the image into editable layers within Magic Layers.

Training data: templates, not photos

Craft’s training data includes millions of human-made templates from Canva’s content library: professional posters, social media graphics, presentations, infographics, menus, and typography-heavy layouts. Canva says most image models are trained primarily on photos, while templates teach the model how layout, hierarchy, and typography come together. After pretraining, the model went through supervised fine-tuning and reinforcement learning stages using feedback from professional designers.

Where it ranks

Canva says Craft entered DesignArena as the fourth most advanced image generation model. According to the company, the model ranks first in the world in the People and Portraits category among 87 models and in the Landscape and Nature category among 85 models; it is second in the Abstract and Patterns category and fourth in the Products category. In internal evaluations, the company’s designers and dozens of researchers compared more than 25,000 image pairs against leading models. When automated evaluation conflicted with human judgment, the designer’s decision was used.

Speed and cost

Canva notes that more than 250 million users are producing over 400 designs per second on the platform, and delivering frontier model performance at this scale requires advances in computational efficiency. Advanced distillation reduced the model’s inference steps to a tiny fraction of the original model’s; a custom inference engine written for the planning stage made Craft an order of magnitude faster than its first internal version. This work lays the foundation for Craft-fast and Craft-ultrafast versions that will run at lower cost.

Canva Research, with more than 160 AI research scientists and engineers, is part of the company’s 400-person AI unit. The company points to fine print and technical drawing as areas for improvement. The model is supported by Canva Shield’s security approach; it will soon also connect to AI assistants and automated workflows through Canva MCP Server.

Source: Unite.AI

// JOIN THE CONVERSATION

Share your experience

Bir sorunun, farklı bir çözümün veya eklemek istediğin bir notun var mı? Giriş yapmadan da yazabilirsin.

// KEEP READING

One more note.

All posts ↗︎