设计平台Canva发布了首款由其研究团队自主研发的图像生成模型Canva Craft，旨在与谷歌、OpenAI和Meta等公司的同类模型展开竞争。Canva表示，该模型分两个阶段运行。第一阶段，模型解读用户输入的提示词，并制定生成计划，其中包括各元素之间的空间关系、相对比例、视觉层级以及文本生成所需的信息。第二阶段，模型根据计划生成图像，最高支持2K分辨率，并提供多达30种宽高比，以适配不同平台。
Canva表示，这种方法能够使模型准确呈现物体细节，并生成清晰可读的文字。此外，用户还可以对图像进行精确编辑，而不影响图像的其他部分。
该公司指出，与使用照片训练的图像生成模型不同，Canva Craft主要使用设计作品进行训练。训练数据来自Canva素材库中由人类设计师制作的模板，包括海报、社交媒体帖子、演示文稿、信息图表、菜单以及包含大量文字排版的设计作品。
这意味着，Canva Craft生成的图像可以通过该公司的图像转设计工具Canva Magic，转换为具有独立图层的可编辑设计作品。去年，Canva推出了自主研发的设计理解模型，能够识别设计结构和图层。
Canva Craft现已向Canva平台的所有用户开放，用户可以通过Dream Lab（梦幻实验室）选项卡使用该模型。该模型还将通过Canva API提供服务。
此外，Canva计划通过其模型上下文协议（MCP）集成，为ChatGPT和Gemini等第三方聊天机器人提供图像生成功能。
此前，Canva一直在平台内使用多种外部图像生成模型，因此，新模型有望帮助公司节省API调用成本。
Canva表示，除了降低成本之外，自研模型还使平台能够更好地控制图像生成质量和速度。
2024年，Canva收购了媒体生成平台Leonardo.ai。该公司表示，来自Leonardo的研究人员参与了新图像模型的研发工作。
Canva还表示，正在开发两款成本更低、效率更高的模型版本，分别名为Craftfast和Craft-ultrafast，预计很快发布。
Canva去年有望实现40亿美元营收。不过，该公司下调了2026年的增长预期。据外部分析师估计，此举导致其估值从420亿美元大幅下降至350亿美元。
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。
The first image generation model developed by its research team was designed to compete with similar models from Google, Open AI and Meta. According to Canva, the model operates in two phases. In the first phase, the model interprets the prompt entered by the user and develops the production plan, which includes spatial relationships between the elements, relative proportions, visual layers and the information required for text generation. Phase II, the model produces images according to the plan, supports up to 2Kresolution and provides as many as 30 wide and high ratios to fit different platforms.
According to Canva, this method enables the model to present object details accurately and to generate clear readable text. In addition, the user can edit the image accurately without affecting the rest of the image.
It states that, unlike the image generation model using photo training, Canva Craft used mainly design works for training. The training data are from templates produced by human designers in the Canvaasset library, including posters, social media posts, presentations, information charts, menus and designs with a large number of text layouts.
This means that the images generated by Canva Craft can be converted through the company ' s image-to-design tool, Canva Magic, into an editable design with a stand-alone layer. Last year, Canva launched an autonomous design understanding model that identifies design structures and layers.
Canva Craft is now open to all users of the Canva platform, who can use the model through the Dream Lab tab. The model will also provide services through Canva API.
In addition, Canva plans to integrate through its model context protocol (MCP) to provide image generation for third-party chat robots such as ChatGPT and Gemini.
Previously, Canva had been using multiple external image generation models within the platform, and it was therefore expected that the new model would help companies save API access costs.
In addition to reducing costs, Canva states that self-study models enable the platform to better control the quality and speed of image generation.
In 2024, Canva acquired Leonardo.ai, the media generation platform. It indicated that researchers from Leonardo were involved in the development of new image models.
It also indicated that two more cost-effective versions of the model, called Craftfast and Craft-ultrafast, were under development and were expected to be soon publish.
Canva is expected to reach $4 billion last year. However, the company reduced its growth expectations for 2026. According to estimates by external analysts, this led to a significant decline in its valuation from $42 billion to $35 billion.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.