Canva 发布自研图像模型，DesignArena 排名仅次于 OpenAI
做设计工具的 Canva 也自己训模型了，图像生成榜排名第二，先把排版和文字规划好再渲染，改图只动改的那块。
Canva 推出首个自研图像生成模型 Canva Craft，在 DesignArena 图像生成榜单上位列第二，仅次于 OpenAI。该模型采用两阶段架构：先由规划模型读取提示词，写出画面中各元素的位置、大小、强调点和需显示的文字；再由渲染模型按计划生成最高 2K 分辨率、支持 30 多种宽高比的图像。文字拼写和排版在渲染前就已确定，修改某个元素时只有该元素变化，不会影响旁边的其他内容。
Canva has entered the frontier AI model race, and its first model is already near the top.
It's first in-house image model made it the second-best (on DesignArena) lab in image generation, with only OpenAI ahead of it.
Canva Craft has a planner that reads the prompt and writes out the composition before any pixels exist: which elements appear, where each one sits, how large it is, what carries the emphasis, and the exact words to display.
A second model, the renderer, paints that plan at up to 2K across more than 30 aspect ratios.
So layout is no more something that the model has to get right inside a single pass. It becomes an explicit intermediate object.
Spelling gets settled before rendering starts. The renderer’s job is to draw words it was already given, in a spot it was already assigned.
The plan pays off a second time during revision. Change one element and only that element changes, so fixing a subhead leaves the product shot next to it alone.
