The launch of an in-house model focused on design
Canva has introduced Craft, its first image generation model developed by its own research team. This launch marks a strategic shift for the platform, which previously relied on external models from providers such as Google, OpenAI, or Meta for its visual generation features.
The company defines Craft as a frontier model designed specifically for design, rather than a generalist image generator. Its primary objective is not merely to produce an image from text, but to create compositions that respect graphic design rules, such as visual hierarchy, element placement, and text legibility.
Two-stage architecture: planning before generating
The most distinctive technical feature of Craft is its two-stage architecture. In the first phase, the model interprets the user's instruction and builds a detailed visual plan. This schema defines the spatial location of objects, their relative scale, visual hierarchy, and the information necessary to generate legible text.
In the second stage, the system uses that plan to render the final image. This separation between planning and execution aims to solve common problems in image generation, such as object distortion or illegible text, by forcing the model to think through the structure before creating the pixels.
Canva Craft workflow
The model's internal process follows a logical sequence that prioritizes structure over immediate aesthetics.
- 01Instruction interpretation: The model analyzes the prompt to identify required elements.
- 02Visual planning: A schema is generated with the position, scale, and hierarchy of elements, including text placement.
- 03Rendering: The final image of up to 2K resolution is produced based on the established plan.
- 04Integration: The image can be converted into an editable design using Magic Layers.
Training with designs and layer-based editing
To train Craft, Canva relied on a library of millions of human-created templates rather than photographs. This database includes posters, presentations, infographics, and menus, allowing the model to understand the conventions of graphic design and typography.
A practical advantage of this approach is the integration with Magic Layers, Canva's tool that converts flat images into editable layered designs. This means the output of Craft is not a static file, but a starting point that allows users to modify individual elements without having to regenerate the entire image.
Performance on Design Arena and competitive positioning
Canva states that Craft debuted in fourth place on the Design Arena image leaderboard, behind three OpenAI models. The company argues that this makes it the second most advanced image lab in the world, a claim based on the position of its single model on the table.
However, analysts point out that Design Arena rankings are dynamic and that a lab's position is not measured in the same way as an individual model's. Furthermore, Canva has not published independent performance comparisons against the models it previously used, so the real improvement in user experience remains an area to be validated.
Initial position on Design Arena (Preliminary data)
The following table shows the position of Canva Craft against other leading models at the time of launch, according to publicly available data.
| Model | Approximate score | Position |
|---|---|---|
| GPT-Image-2.5 Sunburst | 1.406 | 1st |
| GPT-Image-2.5 Flare | 1.382 | 2nd |
| GPT Image 2 | 1.362 | 3rd |
| Canva Craft | 1.341 | 4th |
Availability, API, and current limitations
Craft is available to all Canva users via Dream Lab, with usage limits depending on the subscription plan. For developers, Canva has launched a public API, but it is in a preview phase, meaning it can change without notice and is not suitable for public-facing client applications.
The company has also announced plans to launch faster and more economical versions, called Craft-fast and Craft-ultrafast, as well as an integration via MCP to connect the model with third-party chatbots like ChatGPT and Gemini. However, no specific dates have been provided for the deployment of these features.
Open questions
- No independent performance comparison between Craft and the external models Canva previously used has been published.
- Canva's claim of being the 'second most advanced image lab' is based on the position of a single model and not on a comprehensive lab evaluation.
- Exact dates for the launch of the Craft-fast and Craft-ultrafast versions, as well as the MCP integration with third-party chatbots, are unknown.
- Canva's API documentation recommends using a different model (Z Image Turbo) for images with heavy text, raising doubts about Craft's actual strength in complex typography.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction