{"slug": "canva-built-its-own-ai-image-model-and-says-only-one-lab-now-makes-better", "title": "Canva built its own AI image model, and says only one lab now makes better pictures", "summary": "Canva launched Canva Craft on October 22, 2026, the first image generation model it built and trained from scratch, and said Craft ranks second only to one AI lab worldwide for image generation. Craft, developed by Canva Research with researchers from acquired Sydney startup Leonardo.ai, went live inside Canva's Dream Lab for all users and through Canva's public API; it uses a planning model that writes a composition plan before a separate rendering model produces images at up to 2K resolution in more than 30 aspect ratios, writes text in 55 languages, and was fine-tuned on millions of Canva templates plus more than 25,000 designer-compared image pairs. Canva said Craft debuted fourth on DesignArena and first of 87 models for people and portraits, though Craft did not yet appear on DesignArena's public image leaderboard on Thursday evening, where OpenAI's GPT-Image-2.5 models held the top two places; CEO Melanie Perkins had told shareholders the company was \"relying too heavily on frontier models\" after Canva cut its 2026 growth forecast from 30% to 20% in August.", "body_md": "Canva’s launch artwork for Craft, a collage of sample images, from its announcement on October 22, 2026. Image: Canva\n\nCanva, the design app used by more than a quarter of a billion people, has [launched Canva Craft](https://www.canva.com/newsroom/news/canva-craft/), the first image generation model it has built and trained from scratch, and says it puts the company second only to one AI lab in the world for making images. The model went live on Thursday, October 22, inside Canva’s Dream Lab for every user, and through Canva’s public API for developers.\n\nCraft comes from Canva Research, the company’s in-house AI lab, and is pitched squarely at image models from Google, OpenAI, ByteDance and Meta. Until now Canva has leaned on other companies’ models: Dream Lab launched in October 2024 [running on Phoenix](https://www.businesswire.com/news/home/20241022640737/en/Canva-Boosts-AI-and-Workplace-Tools-with-the-Introduction-of-Dream-Lab-and-Significant-Visual-Suite-Upgrades), the model from Leonardo.ai, the Sydney start-up Canva bought that year. Researchers from Leonardo worked on Craft, Canva [told TechCrunch](https://techcrunch.com/2026/10/22/canva-releases-its-first-image-generation-model/).\n\n## It plans the picture before it paints it\n\nMost image models turn a prompt straight into pixels. Craft splits the job in two. A planning model first reads the prompt and writes out a detailed plan of the composition: which elements appear, where they sit in relation to each other, their relative size, what matters most, and any text that has to appear in the image. A separate rendering model then turns that plan into a picture at up to 2K resolution, in more than 30 aspect ratios or a custom size.\n\nCanva says planning first gives legible text, objects placed where you asked and edits to one element without redrawing everything else. Craft can write text into images in 55 languages. The layered structure also feeds Magic Layers, Canva’s tool that turns a flat image into separate, editable design elements.\n\n## Trained on Canva templates, not just photos\n\nCanva says most image models learn mainly from photographs, while Craft’s training leaned heavily on millions of human-made templates from its own library: posters, social posts, presentations, infographics, menus and typography-heavy layouts. The model was then fine-tuned with feedback from Canva’s professional designers, who compared more than 25,000 pairs of images against leading rival models. “When automated evaluations and human judgments differed, we sided with the designer,” the company wrote.\n\nThe rankings Canva quotes come from DesignArena, a crowdsourced site where people vote between AI designs. Canva says Craft debuted as the fourth best image model there and ranks first of 87 models for people and portraits and first of 85 for landscapes and nature. Those placings are Canva’s own claims: Craft did not yet appear on DesignArena’s public image leaderboard on Thursday evening, where OpenAI’s GPT-Image-2.5 models held the top two places.\n\n## “Relying too heavily on frontier models”\n\nBuilding its own models is also about money. In August, Canva cut its 2026 growth forecast from 30% to 20%, and chief executive Melanie Perkins told shareholders the average cost of serving an AI task had been too high, [Startup Daily reported](https://www.startupdaily.net/topic/business/canva-cuts-revenue-forecast-by-a-third-as-it-tackles-high-ai-costs/). She wrote that the company had been relying too heavily on frontier models, and added:\n\nSeveral of our first-party models were not yet ready for release, and our pricing, consumption model and usage controls had not caught up with the outsized demand we were seeing.\n\nMelanie Perkins, Canva CEO, in her update to shareholders\n\nCraft is the first of those homemade models to reach every user. Canva says model distillation and a custom inference engine make it an order of magnitude faster than its first internal version. Two cheaper variants, Craft-fast and Craft-ultrafast, are on the way, which Canva says will cost a fraction of what leading frontier models charge.\n\n## Who can use it and what it costs\n\nCraft is available now to all Canva users worldwide, free and paid, through Dream Lab in the bottom left panel of the editor or at canva.com/dream-lab. Canva hasn’t said how many Craft images each plan includes, or what the API costs per image. Developers can build it into their own apps through Canva’s public API.\n\nSupport through the Canva MCP server is “coming soon”, so AI assistants and agents can make a ready-to-use graphic as one step in a bigger task. Canva already has a plug-in inside ChatGPT, shown off at [OpenAI’s DevDay](https://madrobot.blog/2026/09/29/openai-devday-2026-everything-announced-dots-gpt-6-1-sol-pro-500/) last month.\n\n## Why it matters\n\nCanva says people make more than 400 designs on it every second, so every AI image it can make itself is one it no longer pays OpenAI, Google or anyone else for. Rivals are cutting prices fast, with Google [halving the cost of an image with Nano Banana 2.1](https://madrobot.blog/2026/10/08/google-nano-banana-2-1-image-model-price-gemini-api/) this month, so a cheap, design-trained model of its own could matter as much to Canva’s margins as to its users.\n\n*Sources: [Canva Newsroom](https://www.canva.com/newsroom/news/canva-craft/), [TechCrunch](https://techcrunch.com/2026/10/22/canva-releases-its-first-image-generation-model/), [Startup Daily](https://www.startupdaily.net/topic/business/canva-cuts-revenue-forecast-by-a-third-as-it-tackles-high-ai-costs/), [Canva (Business Wire)](https://www.businesswire.com/news/home/20241022640737/en/Canva-Boosts-AI-and-Workplace-Tools-with-the-Introduction-of-Dream-Lab-and-Significant-Visual-Suite-Upgrades)*", "url": "https://wpnews.pro/news/canva-built-its-own-ai-image-model-and-says-only-one-lab-now-makes-better", "canonical_source": "https://madrobot.blog/2026/10/22/canva-craft-ai-image-model-dream-lab-canva-research/", "published_at": "2026-10-22 19:33:00+00:00", "updated_at": "2026-10-22 19:39:53.737534+00:00", "lang": "en", "topics": ["generative-ai", "ai-products", "artificial-intelligence", "ai-research"], "entities": ["Canva", "Canva Craft", "Canva Research", "Leonardo.ai", "Melanie Perkins", "DesignArena", "OpenAI", "GPT-Image-2.5"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/canva-built-its-own-ai-image-model-and-says-only-one-lab-now-makes-better", "markdown": "https://wpnews.pro/news/canva-built-its-own-ai-image-model-and-says-only-one-lab-now-makes-better.md", "text": "https://wpnews.pro/news/canva-built-its-own-ai-image-model-and-says-only-one-lab-now-makes-better.txt", "jsonld": "https://wpnews.pro/news/canva-built-its-own-ai-image-model-and-says-only-one-lab-now-makes-better.jsonld"}}