Anthropic's Claude Fable 5.2 doesn't generate images, it paints them
First reported by The Decoder ·
AI-generated art now has the potential to mimic human artistic process, including error correction and style development.
Anthropic's new language model, Claude Fable 5.2, can generate complex and photorealistic paintings by simulating the physical process of painting with code. Unlike diffusion models like Midjourney, Fable 5.2 uses Python code to place hundreds of thousands of individual brushstrokes on a virtual canvas, accounting for factors like bristle texture, paint application, and even painting over mistakes. Demonstrations showed the model creating a parking garage scene with approximately 700,000 strokes and a bird painting from 111 coordinate sequences. Fable 5.2 operates without external image models or pre-rendered images, generating each piece as a Python program driving a simulated paint engine. While it scored higher than OpenAI's GPT-6 Astra on Anthropic's internal RoboDojo Sim benchmark, it lagged in speed and fine motor control during a real-world robot arm test.
This demonstrates a significant departure from current generative AI image models, which largely rely on diffusion techniques. Fable 5.2's approach simulates the physical act of painting, suggesting a more agentic and potentially creative path for AI art generation. The model's ability to learn from its "mistakes" and incorporate them into the final piece, or even to "paint over" them, introduces a layer of sophistication previously unseen in AI art.
The comparison with GPT-6 Astra highlights a trade-off between simulation fidelity and real-world performance, particularly in speed and fine motor control. This suggests that while AI can now simulate complex creative processes, bridging the gap to practical, high-speed physical execution remains a challenge for the industry. The development also points towards a future where AI models might not just generate, but actively develop unique artistic styles and personas.
AI-written summary. May contain errors.