Canva Launches Craft, Its First AI Image Model Trained on Design Templates Not Photographs

Canva Craft Enters Competitive AI Image Generation Market

Canva has launched Canva Craft, its first proprietary AI image generation model, marking a strategic shift from relying on third-party AI tools to developing in-house generative AI capabilities1

1

. The design platform, used by 95% of Fortune 500 companies, is now competing directly with AI image models from OpenAI, Google, Microsoft, and Meta3

3

. Available globally to all Canva users through the platform's Dream Lab, Craft represents a fundamental departure from traditional AI image generation approaches by focusing specifically on design workflows rather than photographic realism2

2

.

Two-Stage Architecture Plans Before Generating

What sets Canva Craft apart is its two-stage architecture that mimics how human designers work1

1

. According to Stefano Corazza, Canva's head of AI research, the model first analyzes prompts and maps out composition, including object placement, scale, visual hierarchy, and typography placement before generating the actual image3

3

. This design-focused AI approach allows the system to "reason" before creating visuals, resulting in more complex and comprehensive designs with accurate details and readable text5

5

. The model generates images at up to 2K resolution across more than 30 aspect ratios, making it versatile for social media posts, marketing content, presentations, infographics, and menus2

2

.

Training on Human-Designed Templates Sets New Standard

Unlike competing AI image generation models trained primarily on photographs, Craft was developed using millions of human-made templates from Canva's content library3

3

. The training data includes professionally designed posters, social posts, presentations, infographics, menus, and typography-rich layouts, teaching the model how layout, composition, and typography work together rather than simply learning what finished images look like3

3

. Corazza explained that the model was trained on Canva's library, licensed and public databases, and data from Canva users who opted into training1

1

. This unique training approach gives Craft a deeper understanding of design principles, making it particularly effective for AI for designers who need marketing graphics and social content where accurate text and layout matter as much as imagery5

5

.

Strong Performance on Design Arena Leaderboard

Early benchmarking results show Canva Craft ranking fourth overall on Design Arena's leaderboard for AI image models, trailing only three models from OpenAI while surpassing offerings from Google's Nano Banana and Microsoft's MAI-Image models1

1

. The model achieved first place among 87 models for people and portraits, first among 85 models in landscapes and nature, and second for abstract designs and patterns4

4

. This performance establishes Canva Research, the company's internal research division that has tripled in size over the past year, as the second-highest performing AI lab for image generation models behind OpenAI4

4

. However, Craft trails in the marketing subcategory, an area where many Canva users focus their efforts1

1

.

AI Output Editing Through Magic Layers Integration

Canva has prioritized AI output editing capabilities, addressing a common frustration where generative AI produces results that are almost but not quite right1

1

. Craft builds on Canva's existing Magic Layers technology, which breaks down AI designs into basic elements for precise adjustments without affecting other parts of the image1

1

5

5

. This technologically complex process solves a significant problem for users who need to refine AI-generated content. The ability to edit specific elements while leaving the rest untouched means designers spend less time fixing mistakes and more time on productive work5

5

.

Strategic Move Toward In-House AI Development

Canva's development of Craft follows its acquisition of Leonardo.Ai and a substantial internal ramp-up of AI research capabilities1

1

. Corazza noted that keeping AI technology in-house helps the company control cloud costs and reduce inference time, addressing the unpredictable economics of relying on third-party generative AI services1

1

4

4

. This represents a significant strategic shift for Canva, which previously integrated third-party AI models and launched its first creative AI tool, Magic Media, in 20231

1

. The company believes developing proprietary models gives it greater control over AI economics and performance.

API Access and Future Model Variants

Canva is making Craft available beyond its own platform through an API, allowing developers to integrate the model into other applications4

4

. The company plans to enable AI agents to connect with Canva Craft through the open Model Context Protocol, with integrations planned for third-party AI assistants including ChatGPT and Gemini2

2

5

5

. Canva is also developing two more efficient variants called Craft-fast and Craft-ultrafast to provide faster, more cost-effective options for different use cases2

2

5

5

. The model includes robust input and output moderation, privacy and safety controls, and reporting tools2

2

.

Ongoing Controversy Around Creative AI Tools

The launch comes amid ongoing controversy surrounding creative AI tools, with photographers, illustrators, and artists raising concerns about AI training on copyrighted work, creation of abusive material, and potential replacement of human creators1

1

. Eric Wittman, CEO of photo editing platform VSCO, responded to Canva CEO Melanie Perkins' announcement by asserting that real photographers, not AI, will always rank first for portraits1

1

. As AI models improve, real creators face increasing pressure to verify their work as human-made, though digital verification methods remain sporadic and inconsistent1

1

. Creative companies like Canva and Adobe maintain that AI serves as a tool to enhance rather than replace human creativity, though this position remains contested within the creative community.