Claude Fable 5.2 'paints' images stroke by stroke in code, Anthropic demos show

Started by DodgyCoder, Today at 07:05 PM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Topic: Claude Fable 5.2 'paints' images stroke by stroke in code, Anthropic demos show   Views(Read 66 times)
Active members in this topic:
DodgyCoder(1)

DodgyCoder

Anthropic staff have shown an unusual way an AI model can make pictures. The Decoder reports that Claude Fable 5.2 writes Python programs of 1,000 to 2,000 lines that place each brushstroke on a virtual canvas, rather than generating pixels directly

One photorealistic parking garage scene used about 700,000 simulated strokes, according to Anthropic's Jake Eaton. A bird at sunrise used just 111 strokes

On Anthropic's internal RoboDojo Sim benchmark, Fable 5.2 scored 64.9 percent against 29.8 percent for GPT-6 Astra. In a real robot arm test, GPT-6 Astra painted faster

The source is a set of demos posted on X by Anthropic employees rather than a formal launch. The Fable 5.2 release itself was covered earlier

Painting through code is a fascinating test of planning and spatial reasoning, quite different from normal image generators. It could have uses in robotics, where a model must plan physical movements step by step. Internal benchmarks need independent checks, but the demos are striking

Is painting through code a meaningful test of AI ability? And could this approach help robots learn physical tasks?