On October 22, Canva released a new image generation model, Canva Craft. The foundation model was trained from scratch by its in-house research group, Canva Research. Rather than starting to draw as soon as it receives a prompt, it plans the composition first and then generates images at up to 2K resolution.
The design aims to make images easier to fix after they are created. Craft generates the image, and the existing Magic Layers feature breaks that image into editable pieces such as text and background. How does Canva intend to bridge the gap between getting one good-looking image and being able to change an ad's date or headline later?
Decide the layout first, then draw
Craft contains two models inside: one that plans the composition and one that renders the image. The first model reads the prompt and decides what to place on the canvas, how the elements are arranged, and how large each should be. The plan also covers the priority of information and any text to appear in the image, and the second model turns that plan into the final picture.
Take an ad that carries a product photo and a headline. The result depends not only on the texture of the product but also on where the text is placed so it can be read and in what order things are shown. Canva explains that handling composition and text placement before drawing makes it easier to produce complex layouts and legible text. It also says that when you change a targeted element, the surrounding areas need not be altered unnecessarily. That is not a guarantee, however, that everything outside the specified area will always be preserved perfectly.
The training data draws on what Canva has built up from running a design tool. It includes millions of human-made templates from Canva's library: posters, social media posts, presentations, as well as diagrams and menus, all material for learning how layout and text fit together. The aim is to teach ways of arranging information that photographs alone capture poorly. This does not mean no photographs were used.
After pretraining, Canva applied supervised fine-tuning and reinforcement learning, bringing in feedback from professional designers to tune how faithfully the model follows instructions and how accurate its text is. In internal evaluation, designers and researchers compared more than 25,000 pairs of images head to head. Where automated metrics and human judgment disagreed, Canva says it gave priority to the designers' judgment.
The specifications cover more than 30 aspect ratios, and arbitrary ratios can also be entered. Text in images can reportedly be generated in 55 languages, and prompts can be written in any language. Still, the number of supported languages says little about how precise fine Japanese characters will be, and Canva itself lists small text and technical drawings as areas to improve.
One step makes the image, another turns it back into parts
Magic Layers is an existing feature that analyzes an image and separates it into editable elements such as text, objects, and background. Canva Research says it trained a learned decomposition model that understands the elements and layers within a composition. Combined with a font library, the goal is to return text to an editable form instead of leaving it as a cluster of pixels.
Craft plans the composition and draws the image, Magic Layers splits the image into parts, and the user then edits text and layout after the split. A generated image and an editable design are products of different steps.
| Step | Handled by | What to check when using it |
|---|---|---|
| Decide the composition and create the image | Canva Craft | Whether the requested text and objects are placed appropriately |
| Split the image into editable parts | Magic Layers | Whether text and background separate as the intended elements |
| Edit the parts and finish | Edits by the user | Whether layout and text legibility hold up after changes |
The table arranges Canva's explanation of Craft and the research group's explanation of Magic Layers in the order of production. It lets you check image quality at generation separately from how easy the result is to fix after decomposition. It is not a measured comparison of decomposition accuracy, and it does not mean that Craft's first output is already a set of editable parts.
Consider a case where you only want to change the date of an event ad. Regenerating the whole image means you have to watch that the background and product you wanted to keep don't change. If the text can be returned to an independent part, the job becomes simply editing the date. On the other hand, if the text decomposition or font reproduction breaks down, fixing it takes extra effort. The claim that Craft has a structure suited to Magic Layers points to Canva's aim of linking generation to revision within its own system, but how many minutes it actually saves needs separate verification.
Canva says it will expand more advanced editing and reference-image features in future releases. It is too early to judge the generation and decomposition tools available now by including features that have yet to arrive.
In-house models matter most when you revise repeatedly
More than one billion designs are created on Canva every month. Canva's research group explains that at this scale, a difference of a few cents per generation adds up significantly. However, the number of designs created is not the same as the number of AI image generations, so Craft's cost savings cannot be calculated from this figure.
In production work that repeatedly generates and revises images, waiting time until completion and the compute cost of each round accumulate. If a model can be tuned to your own use case, it becomes easier to design not just image quality but also which processes to speed up and where to spend compute. Behind building Craft in-house, one can read an intent to control responsiveness and cost to fit production workflows on Canva.
As part of that effort, the company used model distillation to reduce the number of inference steps needed for image generation, and also developed its own engine for the planning stage. It plans to offer faster versions, Craft-fast and Craft-ultrafast, in the future. The announcement gives no specific generation times or user-facing pricing for the released version, so it is not possible to say how many times faster or how much cheaper it is than competitors.
Having its own model is also different from replacing all external models at once. Canva Research says it needs the flexibility to develop and operate its own models alongside partners' models. It fits Canva's explanation better to see Craft as one of a group of models used according to purpose.
Craft is available to Canva users worldwide and can be used from Dream Lab. A public API for developers is also available, while use through MCP, which lets AI assistants call Canva's features, is planned for the future. The announcement of availability to all users cannot be read as free, unlimited generation, so plan-specific usage limits and Craft pricing need to be confirmed at the time of use.
If you try it in a production setting, the criteria will include not only how the finished image looks but also how far you can get with revisions such as changing a date or fixing a headline in Magic Layers. If Japanese text is generated accurately, can still be edited after being split into parts, and the waiting time and cost remain reasonable through repeated rounds, Craft could become useful from the stage of proposing image ideas through to finishing a design for publication.