For most of the AI image boom, the recipe has been the same: type a prompt, hope the model gets the text spelled right, and regenerate until it does. On October 22, 2026, Canva announced Craft, its first image model built in-house, with a different recipe. Craft first plans the layout of an image, then renders it, which is how a human designer works. Canva says the split helps with readable text and precise object placement (Source: Digital Trends).
Canva is not a frontier lab in the usual sense. It makes design software used by a huge number of non-designers, and its pitch is that a model trained on design material rather than mostly photographs will be better at menus, posters, infographics and slides. This post lays out what Craft is, which claims come from Canva, and what remains unproven.
Quick summary: Announced October 22, 2026, Canva Craft is the company's first in-house image model. It uses two stages, a planning model that maps elements, layout, hierarchy and text, then a renderer, and outputs up to 2K resolution with 30+ aspect ratios. It is available to Canva users through Dream Lab (subject to account and usage limits) and to developers through Canva's API. Canva's own claim is a top-four debut on the DesignArena leaderboard. Pricing and independent test results were not disclosed in the coverage reviewed. Faster Craft-fast and Craft-ultrafast variants are upcoming.
How Craft works: plan first, render second
Craft splits image generation into two models. A planning model reads the prompt and drafts the composition: which elements appear, how they relate in space, their relative scale, the visual hierarchy and any text the image needs. A dedicated renderer then creates the final image (Source: TestingCatalog). Canva says this approach improves detail accuracy and readable text, and lets users edit specific elements without changing the rest of the image (Source: Digital Trends).
The training data is the other half of the story. Digital Trends reports that Craft was trained on millions of human-created Canva templates, such as menus, posters, infographics and presentations, rather than mainly on photographs (Source: Digital Trends). Those templates encode how layout, hierarchy and typography fit together, which is exactly what general image models tend to fumble.
The specs and access
- Resolution: up to 2K.
- Aspect ratios: Digital Trends says 30 supported formats; TestingCatalog says more than 30 presets plus custom ratios.
- Languages: TestingCatalog reports prompts can be in any language and that generated text supports 55 languages (single-source detail).
- Access: Dream Lab for Canva users worldwide, subject to account and usage limits, plus Canva's public API for developers.
- Editing: results can be converted into layered, editable designs using Canva's Magic Layers (Digital Trends).
- Speed: Canva says Craft is an order of magnitude faster than its first internal version, using distillation and a custom inference engine (TestingCatalog).
| Detail | What is reported | Who says it |
|---|---|---|
| Announcement date | October 22, 2026 | Digital Trends, TestingCatalog |
| Architecture | Planning model, then renderer | Both outlets |
| Output | Up to 2K, 30+ aspect ratios | Both outlets |
| Leaderboard claim | Fourth-ranked model on DesignArena, making Canva Research the second-ranked image lab (Canva-reported) | TestingCatalog; Digital Trends quotes the second-lab claim |
| Category ranks | First in People & Portraits (of 87) and Landscapes & Nature (of 85); second in Abstract & Patterns; fourth in Products (Canva-reported) | TestingCatalog only |
| Upcoming variants | Craft-fast and Craft-ultrafast | Both outlets |
| Pricing | Not stated | Both outlets |
| Independent tests | None cited | Digital Trends |
The leaderboard claim, read carefully
Canva says Craft debuted fourth among image models on DesignArena, which would make Canva Research the platform's second-ranked image-generation lab. Digital Trends quotes Canva's announcement making the lab-level claim, and TestingCatalog gives the model-level and per-category ranks (Source: TestingCatalog). These are vendor-reported numbers about a leaderboard that rewards design-style prompts, which is a good fit for a template-trained model. A fourth-place debut is impressive, but it does not tell you how Craft performs on photorealism, product shots or your own brand style. Wait for independent comparisons and test it on your real prompts.
What Canva says is still missing
Canva itself lists future work: advanced editing and reference-image capabilities are planned for later releases, along with further improvements to quality, detail, control, fine print and technical drawings. Craft is backed by Canva Shield, which includes input and output moderation and privacy and safety controls (Source: TestingCatalog). On integrations, Digital Trends reports planned ChatGPT and Gemini integrations through Canva's MCP, while TestingCatalog says Canva MCP support is coming soon. That would let AI assistants create design-ready images within larger workflows, but it is not live yet (Source: Digital Trends).
Digital Trends also makes a fair point: Canva is entering a crowded field led by OpenAI, Google and others, and Craft's value depends on whether its outputs are reliably usable without much cleanup (Source: Digital Trends). No pricing for Dream Lab or the API was given in either report.
Why plan-then-render matters
Anyone who has asked an image model for a poster with a headline, a date and a logo knows the failure mode: the picture looks great, the words are garbled. Separating layout from pixels is one way to attack that. If the plan says there is a headline at the top third, a date below it and a logo in the corner, the renderer has a blueprint to follow, and in principle you can change one element without regenerating the whole scene. Whether Craft delivers that consistently is the thing to test.
To test any design-focused image model fairly, write three prompts that stress layout: a poster with a headline and date, a menu with four items and prices, and an infographic with three labeled steps. Check spelling and placement, not just looks, and compare against at least one general image model.
FAQ
01What is Canva Craft?
Canva's first in-house image model, announced October 22, 2026. It plans a layout first and then renders the image.
02How good is it?
Canva says it debuted fourth on the DesignArena leaderboard, making Canva Research the second-ranked image lab. These are vendor-reported claims, and neither report cited independent tests.
03What resolution and formats does it support?
Up to 2K resolution, with 30+ aspect ratios according to TestingCatalog (30 formats per Digital Trends).
04Where can I use it?
In Canva's Dream Lab for Canva users, subject to account and usage limits, and through Canva's public API for developers.
05Is it cheaper or free?
Neither report states pricing. Faster Craft-fast and Craft-ultrafast variants are planned, aimed at lower cost.
Image models differ a lot on text, layout and style, so the only way to know which one fits your work is to run the same prompt through several. LumiChats offers 40+ premium AI models side by side in one subscription, which makes that kind of head-to-head comparison quick.
