Canva Launches Craft, Its First In-House Image Model, Ranked 4th on Design Arena

TL;DR

Canva released Canva Craft on October 22, its first image generation model built in-house. Canva trained it on millions of human-made templates (posters, social posts, slides, infographics, menus) instead of a pile of photographs. Craft writes a layout plan before it renders anything and outputs up to 2K. Every Canva user can reach it in Dream Lab, and developers can call it as craft_v1_2k through a preview REST API. On Design Arena's image leaderboard it sits 4th at 1,341, behind three OpenAI models and ahead of everything from Google, Meta, Microsoft, xAI and Black Forest Labs.

What Canva shipped

According to TechCrunch, Craft is the first image model to come out of Canva's research team, and it goes up against models from Google, OpenAI, ByteDance and Meta. Unite.AI reports it is the first foundation model Canva Research trained from scratch. Researchers from Leonardo.ai, which Canva bought in 2024, worked on it.

The spec sheet, as reported by TechCrunch, Unite.AI and Wilson's Media:

• Output up to 2K resolution, in about 30 aspect ratios inside the Canva editor.
• Training data: millions of human-made templates from Canva's own content library, including posters, social posts, presentations, infographics, menus and typography-heavy layouts.
• Available now to all users worldwide through Dream Lab, and to developers through Canva's API.
• Output can be turned into an editable layered design with Magic Layers, Canva's image-to-design model.
• Two cheaper variants, Craft-fast and Craft-ultrafast, are announced but have no date.
• Access from ChatGPT and Gemini through Canva's MCP server is planned but not live. Unite.AI lists it as coming soon.

Canva has spent more than a decade selling templates to people who would rather not open Illustrator. Now the templates have trained a model of their own.

How it works: plan first, then paint

Most diffusion-style image models go straight from prompt to pixels. Craft splits the job. A planning model reads the prompt and writes a composition plan covering spatial relationships, relative scale, visual hierarchy and what text goes where. A second rendering model then turns that plan into the image. Canva says the split gives it "greater precision, consistency and control."

Think of it as the difference between a sign painter who freehands the shop window and one who chalks the layout on the glass first. The chalk lines are why the lettering ends up centered.

The training data matters as much as the architecture. A photo-heavy dataset teaches a model what the world looks like. A template library teaches it where the headline goes, how big the logo should be, and how much empty space a menu needs. Those are the decisions Canva's users are actually trying to make. General-purpose labs can scrape finished posters off the web. They can't scrape Canva's template library.

The leaderboard: fourth, and first outside OpenAI

Design Arena ranks image models by Elo from blind, crowdsourced head-to-head votes on identical prompts. When we checked the image board at 01:05 UTC on October 23, Craft was 4th overall. The three models above it are all from OpenAI.

Those numbers support Canva's claim, repeated to Capital Brief and Unite.AI, that it is now the second-highest-ranked AI lab for image generation, behind OpenAI. Reve 2.1 is 20 points back, Microsoft's MAI-Image-2.6 is 25 back, Google's best entry is 50 back, and Meta's Muse Image and Black Forest Labs' FLUX 3 Image are around 80 back. Co-founder and COO Cliff Obrecht told Capital Brief: "This is a huge moment for us, and hopefully it shows that you don't need to be in Silicon Valley to build world-class AI."

Canva also says Craft ranks first in Design Arena's people and portraits and landscapes and nature categories. We didn't check those tabs. The Abstract tab we did check had Craft 2nd at 1,370, eight points behind GPT-Image-2.5 Sunburst at 1,378.

Read the ranking with care

• Arena Elo moves as votes arrive. A model that launched yesterday has fewer votes than the incumbents, so its rating can swing more. Capital Brief calls the placement an early indication.
• Design Arena's image categories (marketing, product, typography, portraits, architecture and so on) sit close to the work Craft was trained for. A high score there says less about photorealism or physics.
• The overall ranking ignores price and latency, and we found no published per-image price for Craft.

What developers get

The Canva developer changelog added preview image generation endpoints on October 1. Craft now appears in them as craft_v1_2k, described as "CRAFT-1 2K, Canva's first-party image generation model." From the API reference:

• POST https://api.canva.com/rest/v1/image-generations starts an asynchronous job. You poll the get endpoint for the result, and the download URL expires after 24 hours.
• It needs a user access token with the asset:write scope, and the user needs the image_generation capability. Each job spends that user's Canva AI credit allowance. An empty allowance returns a 429 with credit_quota_exceeded.
• The rate limit is 20 requests per minute per user. Prompts run up to 10,000 characters, and an idempotency_key is required.
• The API exposes only six aspect ratios: square, 16:9, 9:16, 3:2, 2:3 and 4:5. You get the 30-ish ratios only inside the editor.
• It is a preview API. Breaking changes may land without a version bump, and public apps that use it will not pass Canva's app review.

The model picker is the useful part. Canva recommends leaving model unset so it can route each prompt and fall back if one model is down. If you pin one and it can't serve the request, the job fails instead of falling back. The docs pitch Craft for infographics, posters, slides, portraits, landscapes, products and marketing material. They point to lucid_origin (Leonardo's model) for architecture, logos and fantasy scenes, and to z_image_turbo "if the image includes text."

That last line is odd for a model sold on its handling of typography-heavy layouts. Canva's launch materials promise legible text, but its own API reference still sends text-heavy prompts somewhere else. If your use case is posters with real copy on them, test both before you pin anything.

Why Canva built its own

TechCrunch's reading is cost. Canva uses several image models, and TechCrunch expects an in-house one to cut its API bills while giving Canva more say over quality and speed. The promised fast and ultrafast variants point the same way. The timing fits too: TechCrunch notes that Canva was on track for $4 billion in revenue last year, has since cut its 2026 growth forecast, and is valued by outside analysts at about $35 billion, down from $42 billion.

If you build creative software, look at the dataset. Canva got past Google on this board with a library of finished designs nobody else has, then shipped the model inside a product that already has an audience. If your product sits on a pile of finished, human-made work in one narrow format, you may be holding a training set the general-purpose labs can't buy.

Key Takeaways

• Canva Craft, released October 22, is Canva's first in-house image model. Canva trained it on millions of its own human-made templates rather than photographs.
• A planning model lays out composition, scale, hierarchy and text before a render model produces the image at up to 2K, and Magic Layers can convert the output into an editable design.
• On Design Arena's image board Craft sits 4th at 1,341 Elo, behind three OpenAI models and ahead of Reve, Microsoft, Google, Meta and FLUX. It is early and the votes are still coming in.
• It is live for all users in Dream Lab and in a preview REST API as craft_v1_2k, billed against the user's AI credits, limited to 20 requests per minute and six aspect ratios.
• Canva's own API docs still recommend Z Image Turbo for text-heavy images, so test Craft on your real layouts before you pin it.
• Craft-fast, Craft-ultrafast and MCP access from ChatGPT and Gemini are announced but undated.

Sources: TechCrunch, Design Arena image leaderboard, Canva API reference: Create image generation job, Canva developer changelog, Capital Brief, Unite.AI, Wilson's Media