All articles

AI5 min readBy JH Akash

Canva Craft Explained: Canva's First AI Image Model Plans Before It Renders

Canva launched Craft on Oct 22, 2026, its first in-house AI image model. It plans the design first, then renders at up to 2K in 30 aspect ratios. Full breakdown of how it works and how it compares.

Canva Craft explained infographic: 2K resolution, 30 aspect ratios, 2-stage plan-first pipeline

What happened

On October 22, 2026, Canva released its first in-house image generation model: Canva Craft. It is built to compete with image models from Google, OpenAI, ByteDance and Meta, and it works differently from all of them. Instead of generating pixels straight from a prompt, Craft runs a two-stage pipeline: it plans the design first, then renders the image around that plan.

Stage one interprets your prompt and builds a design plan covering spatial relationships between elements, relative scale, visual hierarchy, and information for text generation. Stage two then generates the image at up to 2K resolution in one of 30 aspect ratios. Canva says researchers from Leonardo.ai, the media generation platform it acquired in 2024, contributed to the new model.

Why the plan-first approach matters

Most consumer image tools use a one-shot approach: prompt in, pixels out. That is great for novelty images but unreliable for marketing work, where you need legible text, exact object placement and consistent branding. Craft was trained on millions of human-designed templates from Canva's own library (posters, social posts, presentations, infographics, menus and typography-rich layouts) rather than photographs, so the model is optimized for editable, repeatable design output.

Canva says the plan-first design also allows precise edits without changing other parts of the image, and generated images can be converted into layered designs using Canva Magic, the company's image-to-design transformer. Beyond quality, there is a business motive: Canva has been paying for third-party image models, and owning the model gives it cost savings plus better control over quality and speed.

![Flow diagram showing Canva Craft's two-stage pipeline: Step 1 Prompt (user describes the design in plain text), Step 2 Plan (spatial relationships, relative scale, visual hierarchy, text layout info), Step 3 Render (image up to 2K resolution, 30 aspect ratios). Caption notes one-shot diffusion skips the plan step.](https://pnmsivdmxysronpzmciy.supabase.co/storage/v1/object/public/blog-media/body/canva-craft-explained-canva-first-ai-image-model-62d0bcc7.webp)

Canva Craft: confirmed specs at launch

SpecCanva Craft
Launch dateOctober 22, 2026
Pipeline2 stages: plan, then render
Plan step coversSpatial relationships, relative scale, visual hierarchy, text generation info
Max resolutionUp to 2K (vendor-reported)
Aspect ratios30 (vendor-reported)
Training dataMillions of human-designed templates (posters, social posts, presentations, infographics, menus), not photographs
Text renderingDesigned for legible text in designs
EditingPrecise edits without touching other parts of the image
Layered outputConverts to layered designs via Canva Magic
AccessDream Lab tab for all Canva users; Canva API; MCP support in third-party chatbots like ChatGPT and Gemini planned
PricingNot disclosed at launch
TeamCanva research team, with contributors from Leonardo.ai (acquired 2024)

![Bar chart of Canva Craft key numbers from the October 22 2026 launch announcement: 30 supported aspect ratios, 2 pipeline stages (plan, then render), and 2 fast versions on the way named Craftfast and Craft-ultrafast.](https://pnmsivdmxysronpzmciy.supabase.co/storage/v1/object/public/blog-media/body/canva-craft-explained-canva-first-ai-image-model-214e09ed.webp)

Craft vs one-shot image generators: quick comparison

Canva CraftTypical one-shot diffusion tools
Generation approachPlan first, then renderPixels generated directly from prompt
Trained onDesign templatesMostly photographs and general web images
Legible textA core design goalStill inconsistent in many tools
Output editingPrecise per-part edits, converts to layersUsually regenerates the whole image
Built-in distributionInside Canva's editor, 30 aspect ratios for social, print and presentationsStandalone apps or API

No independent head-to-head benchmarks exist yet, so treat every capability claim as vendor-reported for now. The real test will be whether plan-first generation actually produces fewer broken text blocks and warped layouts in real marketing work.

Which should you pick

  • You make marketing content inside Canva: Try Craft in the Dream Lab tab. It is purpose-built for your workflow, and the layered output means you can fix anything it gets wrong.
  • You need photorealistic images of people or products: Craft is trained on designs, not photographs. A general-purpose image model will likely fit better for now.
  • You build products on image APIs: Watch the Canva API launch. Two cost-efficient versions, Craftfast and Craft-ultrafast, are coming soon, which suggests Canva is serious about API usage and speed tiers.
  • You run an agency producing client designs at scale: The combination of legible text, 30 aspect ratios and layered output is aimed squarely at you. For teams building custom AI content workflows, our AI agent development services cover exactly this kind of automation.

The bigger picture

Canva building its own model instead of licensing one is a signal: the platforms closest to end users, the ones that own the distribution and the templates, are turning from customers into competitors. For context, TechCrunch reported Canva was on track for $4 billion in revenue last year, though it cut its 2026 growth forecast and external analysts reportedly pegged its valuation down from $42 billion to $35 billion. Owning the image model stack is one way to protect margins while pushing deeper into AI.

Related reading: our breakdowns of GPT-6 Astra, the model that changed everything and Claude Opus 5.5. If you want help putting AI image and content generation into a real business workflow, talk to us: hire an AI automation agency.

FAQ

What is Canva Craft?

Canva Craft is Canva's first in-house AI image generation model, launched October 22, 2026. It generates images for designs using a two-stage pipeline that plans the layout before rendering pixels, competing with image models from Google, OpenAI, ByteDance and Meta.

How does Canva Craft's two-stage pipeline work?

Stage one interprets your prompt and creates a design plan with details like spatial relationships of elements, relative scale, visual hierarchy and information for text generation. Stage two renders the image at up to 2K resolution in one of 30 aspect ratios.

Is Canva Craft better than Midjourney or OpenAI's image models?

They are built for different jobs. Craft is optimized for editable design output with legible text inside Canva's editor. Artistic tools like Midjourney optimize for aesthetics. No independent benchmarks compare them yet, so any "winner" claim would be premature.

How do I access Canva Craft?

It is available to all Canva users from the Dream Lab tab. Canva also plans to offer it through the Canva API and through MCP connections in third-party chatbots like ChatGPT and Gemini.

What are Craftfast and Craft-ultrafast?

Two cost-efficient versions of the Craft model that Canva says are coming soon. Details on speed, pricing and availability have not been disclosed yet.

Is Canva Craft free?

Canva has not disclosed pricing. The model is available on the company's platform to all users via the Dream Lab tab, which suggests it ships within Canva's existing plans, but API pricing is still unknown.

  • canva craft
  • ai image generation
  • canva
  • leonardo.ai
  • ai design tools
Mistral Voxtral 3 vs ElevenLabs Realtime infographic: 50+ languages, $0.0035 per minute transcription, sub-500ms latency, voice agent SDK

AI

Mistral Voxtral 3 Explained: Voice Agent Stack vs ElevenLabs Realtime

Mistral shipped Voxtral 3 on Oct 20, 2026: 50+ languages of transcription, a real-time voice agent SDK on LiveKit, and TTS from $13 per million characters. The same day, ElevenLabs released Eleven Realtime, two speech-to-speech models with 0.57s response time. Full breakdown inside.