Canva has released Craft, the first image generation model built by its own research team. The company says it is designed for design work rather than photography. Announcing it on 22 October, Canva said Craft is available to all Canva users and to developers through its API. It claims the model makes Canva “the world’s second most advanced AI lab for image generation, as measured by rankings from DesignArena” (Canva). That ranking is Canva’s claim, and TSN could not check it against DesignArena’s leaderboard.
Canva’s claims for Craft
Canva says Craft “debuted as the fourth most advanced image generation model on DesignArena”, which it calls “the largest crowdsourced assessor of AI design”. By Canva’s account, Craft ranks first for “People & Portraits (out of 87 models)” and “Landscapes & Nature (out of 85 models)”, second for “Abstract & Patterns” and fourth for “Products”. Canva gave no scores and no names of the models ranked above it (Canva).
Craft generates images “at 2K resolution”. In the Canva editor, it offers “more than 30 aspect ratios” or a custom size, Canva says. Prompts can be in any language, and Craft “can create text in 55 languages”. Users reach it through Dream Lab, in the bottom-left panel of the editor. Canva says images can then be split into editable layers with Magic Layers, its existing tool for turning a flat image into separate design elements.
TechCrunch reports that the model was built to compete with models from Google, OpenAI, ByteDance and Meta. It also reports Canva as saying that researchers from Leonardo.ai, the image platform Canva bought in 2024, worked on it (TechCrunch).
How Craft plans a layout before drawing it
Most image models generate pixels straight from a prompt. Craft, Canva says, uses “a two-stage architecture that separates compositional planning from image rendering”. A planning model first writes out the composition: which elements appear, where they sit, their relative scale and any text. A second model then renders that plan as an image. Canva says this gives “legible text, precise object placement” and lets users change one element without altering the rest (Canva).
The training data is the other difference Canva points to. It says Craft learned from “millions of human-made templates from Canva’s content library”, including posters, social posts, presentations, infographics and menus. That was followed by supervised fine-tuning and reinforcement learning using feedback from Canva’s designers. For its internal tests, designers and researchers compared “more than 25,000 pairs of images” against leading models. “When automated evaluations and human judgments differed, we sided with the designer,” Canva says. It names fine print and technical drawings as areas still to improve.
What developers get today
Canva’s API reference lists Craft as craft_v1_2k, “CRAFT-1 2K, Canva’s first-party image generation model”. It sits alongside two other models, Lucid Origin and Z Image Turbo. The endpoint is a preview: “This API is currently provided as a preview,” the reference says, warning of “unannounced breaking changes” and that “Public apps that use preview APIs will not pass the review process”. Requests are limited to 20 per minute for each user of an app, each job “consumes the user’s AI credit allowance”, and download links expire after 24 hours.
Canva has also announced two faster, cheaper variants, Craft-fast and Craft-ultrafast, without dates. A connection to AI assistants through the Canva MCP server is marked “Coming Soon” (Canva). TechCrunch reports that the plan includes ChatGPT and Gemini.
How Canva’s documents differ
Canva’s own documents differ in places. The launch post promises “more than 30 aspect ratios”. The API reference offers six: square, 16:9, 9:16, 3:2, 2:3 and 4:5. The launch post highlights “accurate text rendering”. The same API reference, whose model notes are “based on Canva’s evaluations”, tells developers to pick Z Image Turbo “if the image includes text” and recommends Craft for posters, infographics, portraits, landscapes and products. So for now, developers get a smaller, preview version of what Canva users see in the editor.
The training approach is the more distinctive part. Other recent releases TSN has covered have focused on speed and price, such as Krea 3’s sub-8-second 2K edits, or on open weights, such as Bria’s Fibo 2.0. Craft’s plan-then-render approach is closer to the layout controls in Black Forest Labs’ FLUX.3 Image Klein, with one difference: Canva trained Craft on its own template library. That is Canva’s argument for why Craft understands design. The only outside evidence offered so far is a crowdsourced ranking that Canva reports itself.
Owning the model also changes Canva’s costs. Canva already offered other companies’ image models. TechCrunch suggests an in-house model is “likely” to cut Canva’s API spending, which is TechCrunch’s inference. Canva itself cites “better quality and speed control”, TechCrunch reports. Meanwhile, Adobe has been speeding up Photoshop and Premiere with NVIDIA.
What to watch next
Watch whether DesignArena’s public board matches Canva’s figures, and when the API leaves preview. Watch too whether Craft-fast and Craft-ultrafast come with published prices, and when the MCP connection goes live.
Sources
- Canva Newsroom, “Introducing Canva Craft”, 22 October 2026 (primary; source of all capability, ranking and training claims, all Canva’s own). https://www.canva.com/newsroom/news/canva-craft/
- Canva Developers, “Create image generation job”, Canva REST API reference, read 22 October 2026 (primary; preview status, model list, rate limit, aspect ratios, credits). https://www.canva.dev/docs/apps/rest-apis/reference/generations/create-image-generation-job/
- Ivan Mehta, “Canva releases its first image generation model”, TechCrunch, 22 October 2026 (reported). https://techcrunch.com/2026/10/22/canva-releases-its-first-image-generation-model/
Related stories
- Adobe and NVIDIA Expand Collaboration to Speed Photoshop, Premiere, Frame.io and Brand Concierge
- Krea Launches Krea 3 Image Editing Model for Sub-8-Second 2K Edits
- Bria’s Fibo 2.0 Leads Open-Weights Image Editing, Under a Non-Commercial Licence
- Black Forest Labs’ FLUX.3 Image Klein Lands in Hugging Face Diffusers, With Layout Control Built In
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
