Canva launches Craft to compete in AI image generation

A new entry in the AI race

Canva has officially entered the competitive landscape of AI image generation with the launch of Canva Craft. This new model serves as the company's first dedicated tool for creating visuals from text prompts, positioning it alongside major players like Google, Meta, and OpenAI.

The tool is accessible to the public through a feature called Dream Lab. While Canva provides this service to its broad user base, including those on free accounts, the platform notes that there are specific usage limits for the AI features.

By launching Craft, Canva is attempting to integrate generative AI directly into its existing design ecosystem. The goal is to provide users with a way to create custom assets that can be easily incorporated into the platform's familiar workspace.

Training on design, not photos

What distinguishes Canva Craft from many other models is the specific data used to train it. Rather than relying primarily on a massive collection of photographs, Canva trained the model on human-designed templates. These include common design formats such as posters, menus, and presentations.

This training approach is intended to give the AI a better understanding of graphic design principles. Canva stated that the tool plans a design's layout, proportions, and text before it actually generates the final image. This sequence is meant to help the model produce more accurate details and legible lettering.

By focusing on templates, Canva hopes to bridge the gap between raw image generation and functional graphic design. The company wants the AI to understand how elements should be arranged on a page to be effective for a human viewer.

Performance in real-world tests

To see how the new tool handles practical requests, Business Insider conducted a series of tests with escalating difficulty. The first task involved creating a cartoon birthday card with specific details, including a character's hair color, eye color, and clothing patterns. In this instance, Canva struggled with the font and character proportions, and it failed to include the requested paisley pattern.

The second test focused on image editing, specifically moving a necklace from a person's hand to a jewelry showroom display. While Canva successfully identified the pendant and created an atmospheric background, it failed to account for physics. The resulting image showed the jewelry floating in the air or propped against a pole in an unrealistic way.

The final test asked for a photorealistic image of a cat in a Halloween costume on a gothic castle. While Canva produced a complete image with no malformed components, the fur appeared high-contrast and strandy. In this comparison, the model was noted to be less impressive than the output from ChatGPT or Muse.

Comparison with industry rivals

When compared to other leading models, Canva Craft currently shows room for growth in several areas. For example, ChatGPT and Muse were able to better interpret nuanced instructions, such as weaving in seasonal elements or maintaining realistic lighting and white balance in complex scenes.

While Canva's tool performed well in creating a general atmosphere for a showroom, it did not always match the spatial logic of its competitors. The testing suggested that while the tool is capable of generating high-quality visuals, it sometimes misses the specific details requested by the user.

The results indicate that while Canva is making strides in design-centric AI, it is not yet the top performer for high-end photorealism. Users looking for the most accurate physics or specific artistic nuances may still find other models more reliable for the time being.

The road ahead for Canva

The launch of Craft marks a significant step in Canva's evolution from a design tool to an AI-powered creative suite. By leveraging its massive library of human-made templates, the company is trying to create a unique niche that prioritizes layout and design logic over simple image generation.

For the average user, this means more integrated ways to create content quickly within the Canva interface. However, the current test results suggest that for professional-grade work, manual adjustments will likely remain a necessary part of the workflow to correct AI errors.

As the model continues to develop, the focus will likely be on refining how it handles physics and specific prompt adherence. Canva's success will depend on whether it can turn its template-based training into a distinct advantage that outperforms the general-purpose models of its rivals.

How to try it

You can access Canva Craft by signing into your Canva account and navigating to the Dream Lab section to start generating images from text.

Watch

How Canva updates the model to improve physics and spatial logic in edited images.

Whether the template-based training leads to better text rendering than OpenAI's models.