Canva develops its first image model and breaks into DesignArena’s top 4

Canva introduces Craft, its first AI image generation model, capable of planning compositions, rendering text, and creating in 2K.

Canva Craft: An AI Image Model That Plans the Composition Before Generating the Picture

Generating a realistic photograph does not necessarily produce an image suitable for a poster, an advertising campaign, or a social media publication. Design also requires arranging elements, allocating space for text, establishing visual hierarchy, and maintaining a readable composition. Those are the constraints behind Craft, Canva's first image generation model, developed by its internal Canva Research laboratory. Introduced on October 22, 2026, the model uses a two-stage architecture that separates composition planning from image rendering. At launch, it reached fourth place on DesignArena's overall image leaderboard with an Elo score of 1,340, performing particularly well in portrait and landscape generation. Now available in Dream Lab, Craft also marks a strategic shift for Canva, which is moving beyond integrating third-party models to developing its own image generation technology. Planning the Layout Before Rendering the Pixels

Craft's defining feature is its architecture. Rather than processing a prompt and immediately generating an image, the system introduces an intermediate planning stage. A dedicated model interprets the request to determine which objects should appear, their relative sizes, spatial relationships, visual hierarchy, and any text that needs to be rendered. A second model then transforms that structured plan into an image.

The method addresses a recurring difficulty in AI-assisted design: a visually convincing result may still fail to follow the requested composition. A product can occupy too much space, a title can appear in the wrong place, or objects may be arranged differently from the original brief. The planning stage attempts to represent these constraints more explicitly before image synthesis begins.

This has practical implications for communication materials. Posters require space for typography, product advertisements often need negative space for a headline, and social graphics must remain legible on small screens. By separating structure from rendering, Canva aims to align image generation more closely with conventional graphic design principles.

The architecture does not guarantee perfect adherence to complex instructions. Planning is itself performed by an AI model, which can misunderstand spatial relationships or simplify requested elements. Final accuracy depends both on the quality of the plan and on how faithfully the rendering system follows it. Fourth Place on DesignArena's Image Leaderboard

Canva highlights DesignArena as an important measure of Craft's performance. According to the results announced at launch, the model achieved 1,340 Elo points and fourth place overall. The three models above it were developed by OpenAI, allowing Canva to position its research laboratory second among image model developers represented near the top of the ranking.

Craft performs differently across categories. It ranks first in People & Portraits among 87 models and first in Landscapes & Nature among 85. It also places second in Abstract & Patterns and fourth in Products. These results reflect relative preferences within the evaluated categories rather than a universal measurement of graphic quality.

Leaderboard positions remain dependent on the prompts, collected preferences, and number of comparisons. They can change when more votes are recorded or new systems enter the evaluation.

Canva also conducted internal assessments. The company says professional designers and dozens of researchers evaluated more than 25,000 image pairs in direct comparisons against leading models. When automated assessments disagreed with designers' judgments, Canva prioritized the human evaluation. This reflects its emphasis on creative applications, although the full results and methodology are not independently reproduced in the announcement. Training on Millions of Human-Made Designs

Craft's specialization extends beyond architecture. Canva says its training process places substantial emphasis on graphic design materials, including millions of human-made templates from its content library. Posters, social posts, presentations, infographics, menus, and typography-heavy compositions provide examples of how information and imagery are organized within a design.

This differs from training primarily on finished photographs or flattened illustrations. Structured design templates can communicate spatial relationships, visual hierarchy, and the organization of text and graphic elements. Canva's objective is for Craft to learn some of the principles behind effective composition rather than simply reproducing the appearance of finished images.

Following pretraining, Craft underwent supervised fine-tuning and reinforcement learning informed by feedback from professional designers. The process focused on instruction following, typography, and visual quality. The model can generate text in 55 languages, a capability particularly relevant to advertising and editorial materials that combine imagery with written messages.

The company has not published a complete inventory of its training corpus or detailed the proportions of different content categories. Its reliance on structured, human-made templates is therefore a disclosed aspect of the training approach rather than a fully auditable dataset. 2K Images Across More Than Thirty Aspect Ratios

Craft generates images at resolutions up to 2K, supporting more than thirty aspect ratios and additional custom framing options. This allows creators to produce images suited to different formats without relying on a single standard canvas.

Canva presents a broad range of applications, including editorial photography, portraits, landscapes, illustrations, patterns, product imagery, and marketing assets. Prompts can specify details such as lighting conditions, shadow direction, and the relative positioning of multiple objects.

Typography is another area of emphasis. Posters and