跳到正文
原文
The Decoder:AI News· Manuel Uth·· 15 小时前AI 评分47

Anthropic 的 Claude Fable 5.2 不生成图像,而是用代码逐笔画出照片级作品

Anthropic's Claude Fable 5.2 doesn't generate images, it paints them

文章要点
  • 1Anthropic 发布 Claude Fable 5.2,不用扩散模型生成图像,而是用代码在虚拟画布模拟物理笔触逐笔作画
  • 2员工 Jake Eaton 展示约 700,000 次模拟笔触的停车场照片级画作,称无外部图像模型或临摹源图
  • 3每笔输出坐标且不可撤销,画错后模型自行覆盖修正,Graczyk 的实验用 111 笔完成日出鸟画作
  • 4Anthropic 报告 RoboDojo Sim 基准 64.9%,GPT-6 Astra 为 29.8%;研究员真机对比显示 Astra 更快
决策要点

绘画只是展示窗口,实质是语言模型在物理约束下逐笔规划、自我纠错的长时程能力,并已延伸到机械臂测试。

  • 影响:若笔触级可控生成成熟,内容与广告制作可获得代码级的画面控制;同类能力指向机器人等物理自动化;官方基准与第三方真机结果存在落差,两类数字宜分开看待。
  • 参考:可考虑让设计与内容团队评估笔触级代码生成是否适合需要精确控图的场景;评估机器人或具身方案时,可要求厂商同时提供官方基准与真机实测数据。
  • 依据:演示与实验来自 Anthropic 员工 Jake Eaton 和 Michael Graczyk 的公开分享;RoboDojo Sim 的 64.9% 对 29.8% 为 Anthropic 官方口径,真机对比出自研究员 Aris Zhu。
正文

Anthropic's new language model Claude Fable 5.2 can create complex and even photorealistic paintings by using code to place hundreds of thousands of individual brushstrokes on a virtual canvas.

Current image models like Midjourney typically use diffusion, which gradually refines a noisy image into a clear result. Fable 5.2 takes a different approach by simulating the physical painting process, including bristle texture, how the paint goes on, and mistakes it then has to paint over.

Developer Jake Eaton showed on X what the approach can do. Eaton works at Anthropic, where he regularly contributes to research on language model values, safety mechanisms, and AI agent behavior.

When he asked Fable 5.2 to create a painting as hard to distinguish from a photograph as possible, the model generated an image of a shopping cart in a parking garage using roughly 700,000 simulated brushstrokes.

A photorealistic painting of a shopping cart in a parking garage, with bright sunlight casting shadows on the wall.
About 700,000 simulated brushstrokes went into this photorealistic parking garage scene by Claude Fable 5.2. | Image: Jake Eaton

According to Eaton, Claude works without an external image model and without pre-rendered or traced source images. Each piece is a Python program of 1,000 to 2,000 lines that drives a simulated paint engine. The approach resembles agentic image editing, where language models pick tools and evaluate results.

Earlier attempts with the sibling model Opus 5.5 drew criticism for merely imitating existing art. Eaton responded by having Fable 5.2 develop its own artist personas with fictional backstories and consistent styles. The model needs several thousand brushstrokes per painting, and Eaton redirects them with comments.

A painting of a yellow tube slide on a playground beneath a blue sky with a white cloud.
For its paintings, Fable 5.2 develops its own artist personas and controls brushwork through Python code. | Image: Jake Eaton

Without an undo button, the model has to commit to every stroke

Anthropic engineer Michael Graczyk shared a chat log and video of an experiment in which he asked Fable 5.2 to paint a bird at sunrise. The model could only use code to simulate the physics of bristles, wet paint, and canvas texture. It had to output every brushstroke manually as its own coordinate line, with no way to undo strokes once placed. That's what sets this method apart from AI-generated videos, where models like Flux 3 generate image sequences and their temporal consistency directly.

The resulting 21-second video uses a counter to track how the painting takes shape from exactly 111 strokes on a blank canvas. For the first stroke, the model output the color code #30346a, a width of 90 pixels, and five slightly wobbling waypoints. Fable 5.2 said it added the roughly seven-pixel wobble on purpose because a human arm doesn't draw a perfectly straight line either.

Between passes, the model looked at the canvas and reacted to its own mistakes like a real painter would. When the bird's back came out too thin and scraggly, the AI laid a bolder stroke over it, so the stray bristle marks along the edges now look like ruffled feathers.

In a second, 280-stroke piece showing two people painting at dusk, a late lighting effect accidentally smeared paint tubes and a cup on the table. Claude repainted those objects after the fact. The model also had both figures paint with their outer hands so their inner shoulders could touch, just as the prompt requested.

An overhead view of a painting showing a long table with a white tablecloth after a meal, with plates, glasses, a red wine spill, and a child’s drawing.
Even complex perspectives and lighting are handled by the language model through thousands of individual code instructions. | Image: Jake Eaton

The same spatial reasoning transfers to physical robots

Fable 5.2's ability to plan step by step and self-correct under physical constraints isn't limited to virtual canvases. According to Anthropic, the model is built for extremely demanding long-horizon technical tasks, as is Mythos 5.2, a version designed for verified security and biology researchers.

In internal tests, Fable 5.2 solved two open problems in black hole physics, designed three working electronic devices with circuit board manufacturer Diode, and used more than 1,000 sub-agents to program its own laptop operating system in eight days.

AI researcher Aris Zhu tested how the model performs with real robotic arms. He pitted Fable 5.2 against OpenAI's competing model GPT-6 Astra, with each using a physical robot arm to paint Van Gogh's "The Starry Night." Anthropic reports a score of 64.9 percent for Fable 5.2 on its internal RoboDojo Sim benchmark, compared with 29.8 percent for GPT-6 Astra.

Zhu's real-world comparison still shows limits in fine motor control. Astra painted faster, and so far only OpenAI's model has managed to stand a domino on end. It did so by slowly loosening its gripper to use gravity as a pivot point.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.

来源:The Decoder:AI News · the-decoder.com