Canva 其室内研究实验室建造的Debuts工艺图像模型

AI 模型和平台

Canva 其室内研究实验室建造的Debuts工艺图像模型

在 Google 上的首选来源中添加 Unite.AI

Canva于2026年10月22日发布 Canva Craft，是公司内部AI实验室Canva Research从头开始建造和训练的第一个基础模型。 图像生成模型通过Dream Lab供全球所有Canva用户使用，并通过公司公共API供开发者使用。

Canva 称是世界上第二先进的图像生成AI实验室，通过来自DesignArena的排名衡量，该公司称其为AI设计的最大众源评估师。

双层建筑

与常规的图像生成模型直接从一个即时生成像素不同，Craft使用一个将组成规划与图像渲染分开的两级架构。 在第一阶段，一个图像规划模型解释即时，并写一个详细的构成计划：哪些元素出现，它们的空间关系，相对尺度，视觉层次，以及任何需要出现在图像中的文本。 在第二阶段，一个专用的图像渲染模型将计划转换成最高2K分辨率的高度真实度图像，跨度超过30个方面的比例。

Canva 称通过在渲染前建模成分和排版，Craft可以产生可辨读文本，精确对象位置和连贯的视觉关系的复杂布局，而架构能够对单个元素进行更多的颗粒控制，允许有针对性地改变，而不会不必要地改变其他一切。 公司称结构相同，使得Craft输出非常适合魔法层，其将生成的图像转化为完全可编辑的设计元素的工具，并为未来发布规划的高级编辑和参考图像能力提供了技术基础。

用户可以用任何语言书写提示，Craft可以用55语言生成文本。 用户可以从超过30个侧面比中选择，也可以输入自定义比。 一个JSON边界箱功能将生成的图像转换成魔法层内部的可编辑层。

培训、评价和效率

工艺师的培训数据包括来自Canva内容库的数百万人造模板，包括专业设计的海报、社交文章、演示、图象、菜单和丰富的印刷版面。 Canva 称大多数图像模型主要是以照片为主的训练，而模板则教模型如何布局，层次，和打字相融合。

经过预训，Canva通过监督的微调和强化学习优化工艺，吸收其专业设计师的反馈，以提高后续教学，打字精度，整体视觉质量。 在内部评价方面，Canva 称公司的专业设计师与数十名研究人员合作，在与主要模型进行头对头比较时评估了25,000多对图像，当自动化评价和人类判断不同时，它与设计师站在一起。

Canva报道，Craft在DesignArena上作为第四最先进的图像生成模型首发。 称公司的产出在87个模型中为People & Portraits，在景观与自然，在85个模型中为People & Portraits，并列第二位，产品则列第四位。

Canva报道，超过25亿人使用其平台，每秒创造400多个设计。 称公司在这一规模上提供前沿级性能，需要在计算效率和推论速度两方面都取得突破。 它的称高级模型蒸馏将Craft所需的推论步骤数量减少到原模型的一小部分，一个为规划阶段所优化的自定义推论引擎使得Craft比其第一个内部版本更快的量级。 Canva认为，这些进步支撑了手工业快车和手工业超快车，这两个即将到来的变体旨在以领先前沿模型的一小部分成本提供高质量的图像生成。

供应和安全

要在Canva内部使用Craft，用户从左下方面板打开Dream Lab. 开发者和产品团队可以通过Canva的公共API将模型建设成自己的工具，包括营销平台，电子商务网站，广告构建者和打印服务。 通过Canva MCP Server提供支持，该服务器将Craft与AI助手和自动化工作流程连接起来，这样agents就可以生成可以随时在设计内部使用的图像，并很快被列出。

Canva Shield是该公司的AI安全方法，涵盖输入和输出节制、隐私和安全控制，Canva 称除了视觉质量外，还在发布其模型前评价品牌安全，文化背景，以及有害内容的可能性。

模型背后的实验室

一天前的2026年10月21日，Canva公布了Canva Research内部的第一面，这是其内部用于视觉和设计AI的研发实验室。 实验室由全球超过160名AI研究科学家和工程师组成，在一个由400人组成的更广泛的AI单位内。

Canva的AI研究始于2017年，其模板库的机器学习搜索和建议。 它于2021年和2024年分别设在维也纳的AI实验室Kaleido和Leonardo,2025年公司将其AI研究团队联合起来，成为Canva Research. 它后来加入了MangoAI和Simtheory,Canva 称为强化学习带来了深厚的专业知识和AI agents.

Canva报告其AI工具的使用量超过360亿，每月使用Canva创建的设计量超过10亿。 称公司内部的模型已经为整个平台提供了几十种特性，其中包括“魔法层”和“Lucid Origin”的图像模型。 Canva同时也是称，它的图生视频工作每GPU小时实现了70%的视频生成。

Canva 称更快的工艺模型已经在路上了，它将继续致力于质量，细节和控制，命名精细的印刷品和技术图纸作为有待改进的领域。

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

AI Models & Platforms

Canva Debuts Craft Image Model Built by Its In-House Research Lab

Add Unite.AI to your preferred sources on Google

Canva on October 22, 2026 unveiled Canva Craft, the first foundation model built and trained from scratch by Canva Research, the company’s in-house AI lab. The image generation model is available to all Canva users globally through Dream Lab and to developers through the company’s public API.

Canva said the launch makes it the world’s second most advanced AI lab for image generation, as measured by rankings from DesignArena, which the company described as the largest crowdsourced assessor of AI design.

A Two-Stage Architecture

Unlike conventional image generation models that generate pixels directly from a prompt, Craft uses a two-stage architecture that separates compositional planning from image rendering. In the first stage, an image planning model interprets the prompt and writes a detailed plan of the composition: which elements appear, their spatial relationships, relative scale, visual hierarchy, and any text that needs to appear in the image. In the second stage, a dedicated image rendering model translates the plan into a high-fidelity image at up to 2K resolution across more than 30 aspect ratios.

Canva said that by modelling composition and typography before rendering, Craft can generate complex layouts with legible text, precise object placement and consistent visual relationships, and that the architecture enables more granular control over individual elements, allowing targeted changes without unnecessarily altering everything else. The company said the same structure makes Craft outputs well suited to Magic Layers, its tool for transforming generated images into fully editable design elements, and provides the technical foundation for advanced editing and reference-image capabilities planned for future releases.

Users can write prompts in any language, and Craft can generate text in 55 languages. Users can choose from more than 30 aspect ratios or enter a custom one. A JSON boundary boxing feature converts generated images into editable layers within Magic Layers.

Training, Evaluation, and Efficiency

Craft’s training data includes millions of human-made templates from Canva’s content library, spanning professionally designed posters, social posts, presentations, infographics, menus and typography-rich layouts. Canva said most image models are trained primarily on photographs, and that the templates instead teach the model how layout, hierarchy, and typography fit together.

Following pre-training, Canva optimized Craft through supervised fine-tuning and reinforcement learning, incorporating feedback from its professional designers to improve instruction following, typography accuracy, and overall visual quality. For internal evaluations, Canva said its professional designers worked alongside dozens of researchers to assess more than 25,000 pairs of images in head-to-head comparisons with leading models, and that it sided with the designer when automated evaluations and human judgments differed.

Canva reported that Craft debuted as the fourth most advanced image generation model on DesignArena. The company said its outputs hold the top global rank for People & Portraits, out of 87 models, and for Landscapes & Nature, out of 85 models, along with second place for Abstract & Patterns and fourth for Products.

Canva reports that more than a quarter of a billion people use its platform, creating over 400 designs every second. The company said delivering frontier-level performance at that scale requires breakthroughs in both compute efficiency and inference speed. It said advanced model distillation reduced the number of inference steps Craft requires to a fraction of the original model, and that a custom inference engine optimized for the planning stage made Craft an order of magnitude faster than its first internal version. According to Canva, these advances underpin Craft-fast and Craft-ultrafast, two upcoming variants designed to deliver high-quality image generation at a fraction of the cost of leading frontier models.

Availability and Safety

To use Craft inside Canva, users open Dream Lab from the bottom left panel. Developers and product teams can build the model into their own tools, including marketing platforms, e-commerce sites, ad builders and print services, through Canva’s public API. Support through the Canva MCP Server, which connects Craft to AI assistants and automated workflows so agents can generate images ready to use inside a design, is listed as coming soon.

Craft is backed by Canva Shield, the company’s AI safety approach covering input and output moderation, privacy and safety controls, and reporting tools that let people flag potentially harmful or misleading content. Canva said that alongside visual quality, it evaluates brand safety, cultural context, and the potential for harmful content before releasing its models.

The Lab Behind the Model

One day earlier, on October 21, 2026, Canva published a first look inside Canva Research, its in-house research and development lab for visual and design AI. The lab is made up of more than 160 AI research scientists and engineers globally, within a broader AI unit of 400 people.

Canva’s AI research began in 2017 with machine learning for search and recommendations in its template library. It acquired Vienna-based AI lab Kaleido in 2021 and Leonardo in 2024, and in 2025 the company brought its AI research teams together as Canva Research. It later added MangoAI and Simtheory, which Canva said bring deep expertise in reinforcement learning and AI agents.

Canva reports more than 36 billion uses of its AI tools and more than one billion designs created with Canva every month. The company said its in-house models already power dozens of features across the platform, including Magic Layers, which is built on a learned decomposition model that understands the elements and layers in a composition, and Leonardo’s Lucid Origin image model. Canva also said its image-to-video work achieved 70% more video generations per GPU hour.

Canva said faster Craft models are already on the way and that it will keep working on quality, detail and control, naming fine print and technical drawings as areas for improvement.

Auto-fetched and translated by AI Atlas. All rights belong to the original author.