Canva claims the 2nd most advanced image lab as Craft slips to 5th

Canva’s launch post for Canva Craft says the model “makes Canva the world’s second most advanced AI lab for image generation, as measured by rankings from DesignArena”. That label rests on arithmetic: Craft debuted fourth, and the three models above it were all OpenAI’s, leaving Canva Research second among labs. On DesignArena’s live image board, pulled from its own API on 24 October, Craft sits fifth. The model now ahead of it is Riverflow 2.5 Pro, from Sourceful, so by Canva’s own measure the company is third.

Canva released Craft on Thursday, its first image generation model built in-house to compete with models from Google, OpenAI, ByteDance and Meta. It renders at up to 2K resolution across more than 30 aspect ratios, and it takes prompts in any language while generating text in 55 of them. Access runs through the Dream Lab tab for all users globally, plus the public API.

The mechanism behind that claim is a two-stage split. A planning model fixes spatial relationships, relative scale, visual hierarchy and text placement first, and a rendering model then turns that plan into the image, which is what Canva credits for legible lettering. The company says it trained Craft on millions of human-made templates from its own library rather than on photographs, a choice TechCrunch reported covers posters, presentations, infographics and menus.

It debuted as the fourth most advanced image generation model on DesignArena.

Canva, Introducing Canva Craft

What the live board shows

DesignArena ranks 86 models in its image arena, and 78 Elo points separate first place from fifth. Here’s the top five as of 24 October, with the head-to-head count each rating is built on.

Rank	Model	Lab	Elo	Battles
1	GPT-Image-2.5 Sunburst	OpenAI	1,408	11,790
2	GPT-Image-2.5 Flare	OpenAI	1,383	12,218
3	GPT Image 2	OpenAI	1,363	61,582
4	Riverflow 2.5 Pro	Sourceful	1,354	16,722
5	Canva Craft	Canva	1,330	2,086

That last column is the caveat nobody printed. Craft’s 1,330 rating comes from 2,086 matchups, the smallest sample in the top ten, against 61,582 for OpenAI’s GPT Image 2 and 16,722 for the model that just passed it. DesignArena rates on a Bradley Terry system it labels Elo, scored from crowdsourced head-to-head votes, and it publishes no error bar on any row in the arena.

That matters because a rating built on roughly a thirtieth of the leader’s votes, by our calculation of the two battle counts, can move a long way. We’ve seen the pattern before, when Modulate raised $25M on a transcription lead it had already lost.

One published hands-on test came out cooler than the ranking, though. Business Insider ran three prompts of escalating difficulty past Craft, ChatGPT, Gemini and Muse. Craft’s birthday card came back with an awkward font and an outsize cartoon head, and like Gemini it placed a pendant in a way that defied physics. Its photorealistic cat, the reviewer wrote, “pretty much matches the quality of Muse and Gemini”, but the overall verdict was that Canva “still has a lot of room to grow compared to other models”.

The case for building it anyway is cost. Canva had been using a range of other labs’ image models on its platform, and TechCrunch read the switch as likely to save it API fees, with Canva itself citing better quality and speed control. Researchers from Leonardo.ai, which Canva bought in 2024, contributed to the model, and The Verge reported that cheaper Craft-fast and Craft-ultrafast versions are still to come, along with agent access over the Model Context Protocol.

Money gives that urgency a shape. Canva was on track for $4 billion in revenue last year, but it cut its 2026 growth forecast, and external analysts put its valuation down from $42 billion to $35 billion, TechCrunch reported. Craft swaps a bought-in input for one the company owns.

The thing worth watching is whether the 1,330 holds as the vote count grows. OpenAI’s three entries are the benchmark to beat, and it shipped ChatGPT Images 2.5 in September. If Craft’s rating drifts down as its battle count climbs toward the tens of thousands the leaders carry, the “second most advanced lab” line won’t survive its own source.