Canva Craft image model ranks 4th, beats Google and OpenAI rivals

What changed

Canva released Canva Craft on October 23, 2026, marking the design platform’s entry into proprietary image generation. The company built the model in-house using Canva Research and trained it from scratch using millions of design templates rather than traditional photo datasets. According to the announcement, Canva Craft can compete with comparable models from Google, OpenAI, ByteDance, and Meta.

The model generates images at up to 2K resolution and supports more than 30 aspect ratios. It renders text across 55 languages and uses a composition-first approach, planning proportions and text layout before generating results. Generated images can be broken into editable layers using Canva’s Magic Layers feature. On DesignArena’s leaderboard, the company says Craft ranked fourth overall and first for portraits and landscapes.

Canva Craft is available now through Dream Lab and the public API. The company plans to add Model Context Protocol integration for chatbots including ChatGPT and Gemini later. Two faster variants named Craft-fast and Craft-ultrafast are expected to launch soon, according to Canva. Researchers from Leonardo.ai, which Canva acquired in 2024, contributed to the project development.

Why it matters

The launch consolidates Canva’s position as a design ecosystem that no longer depends on external image generation APIs. Professionals already using Canva for templates and layouts now have an option to generate imagery without switching to OpenAI’s DALL-E, Google’s Imagen, or other standalone tools.

For Canva users, having an in-house model means tighter integration with the design workflow. Images generated through Craft can flow directly into editable layers, avoiding the friction of downloading and reimporting external results. The company controls iteration speed and can tune the model for design-specific tasks like composition and text rendering.

The ranking on DesignArena signals that Canva Craft is competitive. Placing fourth overall and first in categories like portraits and landscapes suggests the model has particular strength in use cases common to design work. This matters because it establishes Canva as a viable alternative to dedicated image generation platforms for users already invested in the broader design suite.

Cost represents another consideration. By developing the model in-house, Canva gains more control over pricing strategies. The planned Craft-fast and Craft-ultrafast variants suggest tiered pricing options, though specific costs remain unannounced. For teams running high volumes of image generation, this approach could potentially reduce spending compared to pay-per-API-call models from competitors.

The training approach using design templates rather than photographs represents a deliberate strategy. This dataset composition may produce results better suited to design-specific applications than models trained on general photo collections, though real-world performance across different use cases remains to be validated.

What to test

Before adopting Canva Craft as your primary image generation tool within the design workflow, verify these claims:

• DesignArena ranking durability. The company cites fourth-place rank overall and first-place finishes in specific categories. Confirm these rankings remain stable as competing models improve and understand how DesignArena scores models, since rankings can shift.

• 2K resolution consistency. Test whether Canva Craft consistently generates clean output at 2K across all 30+ aspect ratios, particularly non-standard ones. Check for degradation in text rendering quality, especially in languages beyond English.

• Composition planning accuracy. The model’s claim that it plans composition, proportions, and text before generation suggests fewer misaligned or broken text results than some alternatives. Run tests with text-heavy prompts and compare error rates to competitors.

• Layer editability. Verify that Magic Layers extraction works reliably and produces layers that are genuinely editable, not just visual masks. Test complex scenes with overlapping objects.

• API cost and speed. When pricing is public, compare cost per image against OpenAI DALL-E, Google Imagen, and others. Measure response times, especially once Craft-fast and Craft-ultrafast launch.

• Text rendering in non-Latin scripts. The model claims 55-language support. Test text output in Hindi, Arabic, Chinese, and other scripts to assess whether quality degrades compared to Latin-based languages.

• Style consistency. Generate variations of the same prompt to assess whether Craft maintains aesthetic coherence across multiple generations or whether results feel inconsistent.

The conclusion

Canva Craft is a credible competitive entry into image generation. The model’s training approach using design templates rather than photos represents a deliberate strategy that may produce better results for design-specific tasks. The DesignArena rankings suggest this approach is yielding results, at least in certain categories.

For existing Canva users, the model reduces friction and dependency risk from third-party APIs. Integration within the design workflow matters substantially for many professionals. However, the source does not provide direct performance comparisons against DALL-E, Imagen, or other widely-used alternatives, so ranking fourth on one leaderboard does not necessarily mean Craft is objectively superior to all competitors.

What to watch next includes pricing announcements, the speed of rollout for Craft-fast and Craft-ultrafast variants, and real-world performance reports from Canva’s user base. Once pricing is public, cost per image will determine whether the model offers economic advantage beyond workflow convenience. If Canva successfully tiers the model while maintaining quality, the in-house approach could establish a precedent for other design platforms seeking independence from external image generation APIs.

Common questions

How does Canva Craft compare to OpenAI DALL-E and Google Imagen?

Canva says Craft can compete with models from Google, OpenAI, ByteDance, and Meta. On DesignArena, the company reports that Craft ranked fourth overall and first in portraits and landscapes, suggesting it performs comparably to leading alternatives in some categories, though the source does not provide head-to-head metric comparisons on other benchmarks.

What languages does Canva Craft support for text rendering?

The model renders text in 55 languages, according to Canva. The source does not specify which 55 languages are included or whether support covers all major writing systems.

When will the cheaper variants Craft-fast and Craft-ultrafast be available?

Canva says the two cost-efficient variants are “expected soon,” but the company has not announced a specific release date in the source material.

Will Canva Craft work with ChatGPT and other AI assistants?

Model Context Protocol integration for chatbots including ChatGPT and Gemini is planned for later, according to Canva, but an exact timeline is not provided.

Related reading

• Google Agent Compute Engine: microVM isolation, zero credentials
• Granite Speech 5.0 2B ranks #2 on Open ASR Leaderboard
• Replit Switches to OpenAI as Default Model Provider