Canva's first in-house image model turns AI pictures into editable layers

Most AI image generators feel like pulling a lever on a slot machine. You type in a prompt, cross your fingers, and get back a flat pixel blob where text is garbled and tweaking a single element means rolling the dice all over again.

Canva wants to change that. The design giant just introduced Canva Craft, its first fully in-house image generation model built to take on OpenAI, Google, and Meta.

Blueprint first, pixels second

Instead of generating a finished picture in one blind shot, Canva Craft relies on a two-stage process.

In stage one, the model acts like an art director. It breaks down the prompt to map out spatial relationships, element scale, visual hierarchy, and exact typography placement. In stage two, it actually generates the visual at up to 2K resolution across more than 30 aspect ratios.

The secret sauce is in the training data. Canva skipped scraping generic web photos and instead trained Craft on millions of human-made templates from its own library, spanning social posts, presentations, infographics, and restaurant menus.

Because the system understands structure, images generated by Craft can be imported into Canva Magic and split into layered designs. If you need to swap a headline or shift an icon, you adjust that single layer rather than regenerating the entire image.

On the Design Arena benchmark, Canva Craft currently ranks fourth among top AI image models, sitting just behind OpenAI's trio of GPT-Image-2.5 variants.

Cutting the API cord

The model was developed with heavy input from Leonardo.ai, the generative media platform Canva bought back in 2024.

Craft is available right now for all users worldwide under the Dream Lab tab. Canva is also shipping it via API, with lighter, cheaper Craft-fast and Craft-ultrafast variants on the way.

Notably, Canva plans to plug Craft into external chatbots like ChatGPT and Gemini using Anthropic's Model Context Protocol (MCP), letting third-party AI agents trigger Canva's layout-aware generation.

There is also a massive financial motive here. Canva has historically paid millions to route image prompts through external providers. Bringing the core model in-house cuts those cloud API bills while handing Canva direct control over latency and quality.

That balance sheet discipline matters. While Canva was tracking toward $4 billion in revenue last year, it recently lowered its growth projections, causing its valuation to drop from $42 billion to around $35 billion, according to analyst estimates.

Why it matters

For everyday marketers and creators, Canva Craft attacks the most frustrating flaw in modern generative AI: editability.

Hyper-realistic art is fun, but real work requires legible copy, brand layouts, and the ability to nudge a logo two pixels to the left. By tying visual generation directly to layered design files, Canva turns AI output from disposable stock imagery into working design assets. It is well worth testing today inside Dream Lab.

For the wider industry, this marks a turning point where leading vertical apps stop acting as mere wrappers for foundation model providers. Canva's proprietary archive of templates gave it domain-specific leverage that even frontier lab models struggle to match out of the box.

Still, training and hosting high-performance diffusion models is notoriously capital-intensive. With Canva already weathering valuation markdowns, we think the real test will be whether Craft genuinely lowers operating expenses or simply trades third-party API bills for hefty internal compute costs.