Answers from the documentation; checking which model is on duty
Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
Anthropic's Claude Fable 5.2 paints images with code, not generates th - OpenSmartRoute
Anthropic's Claude Fable 5.2 paints images with code, not generates them
Claude Fable 5.2 creates images by simulating brushstrokes through Python code, mimicking physical painting. It can produce complex, photorealistic art and plan physical tasks.
Key points
Fable 5.2 uses about 700,000 brushstrokes for a photorealistic scene
It scored 64.9% on Anthropic's RoboDojo Sim benchmark
Fable 5.2 outperformed GPT-6 Astra in physical robot painting tests
The model develops its own artist personas and styles
Why it matters: It enables precise, step-by-step control of image creation and physical tasks, improving safety and reliability in AI-generated art and robotics.
By OpenSmartRoute editorial · written through the router by llm-small
From The Decoder - “Anthropic's Claude Fable 5.2 doesn't generate images, it paints them”
Introduction - what Claude Fable 5.2 is and how it differs from typical image models
Claude Fable 5.2 is a new type of AI model that creates images differently from most current models. Unlike popular image generators that produce pictures by refining noisy images, Fable 5.2 paints images by simulating physical brushstrokes. It does not generate images directly from text prompts. Instead, it uses code to place each stroke on a virtual canvas, mimicking how a human artist works.
Most image models today use diffusion or other methods that start with noise and gradually improve the image. Fable 5.2 takes a different approach by simulating the physical process of painting. It considers properties like brush texture, paint application, and mistakes. This method results in images that look like they were painted by hand, with a focus on realism and detail.
The painting process - how Fable 5.2 simulates physical brushstrokes with code
Fable 5.2 creates images by generating a sequence of instructions in Python code. Each instruction describes a single brushstroke, including its position, color, width, and wobble. The model writes these instructions as a program with 1,000 to 2,000 lines. When run, the program drives a simulated paint engine that mimics real painting physics.
The process resembles how an artist works with physical tools. The model considers factors like bristle texture and paint flow. It can intentionally add imperfections, such as wobbling lines, to make the image look more natural. The code-based method allows the model to plan each stroke carefully and react to previous strokes, just like a human painter.
Creating photorealistic images - examples and technical details
Fable 5.2 can produce images that look very close to photographs. For example, it generated a parking garage scene using about 700,000 simulated brushstrokes. The strokes were placed with precise coordinate sequences, capturing complex lighting and perspective. The result was a highly realistic image that resembled a real photograph.
The model does not rely on external images or pre-rendered templates. Instead, it builds each image from scratch by following its code instructions. This approach allows it to create detailed, photorealistic scenes with complex textures and lighting effects. The process takes time, but the output can be very convincing.
Jensen Huang, founder and CEO of Nvidia, resigned amid political scandals, weakening US AI dominance and aiding China's progress in AI technology.
Developing artist personas - how the model creates consistent styles
Earlier versions of the model drew criticism for simply copying existing art styles. Fable 5.2 addresses this by developing its own artist personas. It creates fictional backstories and consistent styles for each persona. This helps the model produce paintings that have a recognizable and coherent artistic identity.
To do this, the model needs several thousand brushstrokes per painting. A researcher at Anthropic, who works on safety and AI behavior, guides the model by commenting on its strokes. This feedback helps Fable 5.2 refine its style and develop a unique artistic voice. The result is a more authentic and varied set of paintings.
Handling mistakes and corrections - how the model reacts to errors
Fable 5.2 can react to its own mistakes during the painting process. In a test, it was asked to paint a bird at sunrise. The model used code to simulate physics and output each stroke as a coordinate line. Once a stroke was placed, it could not be undone. Instead, the model looked at the canvas and added new strokes to correct errors.
For example, if a part of the bird was too thin or uneven, the model painted over it with a bolder stroke. It also adjusted its work based on visual feedback, much like a human artist. This ability to self-correct improves the realism and quality of the final image.
Physical robot painting - transferring virtual skills to real-world tasks
Fable 5.2's skills are not limited to virtual canvases. The model can plan and execute physical painting tasks with robotic arms. It can generate step-by-step instructions that a robot can follow to paint in the real world. This capability extends to complex tasks like painting a landscape or performing technical work.
Anthropic tested the model with real robots. It used a robotic arm to paint a version of Van Gogh's "The Starry Night." The model's planning and self-correction abilities help it adapt to physical constraints. This makes it suitable for demanding tasks that require precision and planning over long periods.
Benchmark performance - scores and comparison with other models
On internal tests, Fable 5.2 scored 64.9 percent on the RoboDojo Sim benchmark. This benchmark measures how well a model can plan and execute complex tasks with physical robots. In comparison, a competing model scored 29.8 percent. The higher score indicates better planning and control.
However, Fable 5.2 still faces challenges in fine motor control. During physical painting, it was slower than some competitors. For example, it took longer to paint a scene than other models. It also struggled with tasks like balancing a domino, which requires very delicate movements.
Limitations and challenges - speed and fine motor control issues
While Fable 5.2 can plan and self-correct, it is not yet as fast or precise as some other models in real-world tasks. Its physical control remains a challenge. The model can produce detailed plans, but executing them with high speed and accuracy is difficult.
Speed is limited by the complexity of the code instructions and the physical constraints of robotic hardware. Fine motor control, such as balancing objects or making tiny adjustments, still needs improvement. These limitations mean the model is better suited for planning and simulation than for fast, precise physical work.
Why it matters - implications for AI art and robotics
Fable 5.2 introduces a different way of creating images by simulating physical painting. This approach can lead to more realistic and natural-looking art. It also opens possibilities for AI to perform physical tasks that require planning, correction, and fine motor skills.
For AI art, this method offers a new level of realism and expressiveness. It can produce images that look like they were painted by hand, with imperfections that add authenticity. For robotics, the model demonstrates how AI can plan complex movements and adapt to physical constraints.
This development suggests future AI systems could combine virtual planning with real-world execution. They might be used in art, manufacturing, or maintenance tasks that need careful control and correction. It shows how AI can bridge the gap between digital and physical work.
How OpenSmartRoute helps - what a team routing through OpenSmartRoute does with Anthropic's Claude Fable 5.2 paints images with code, not generates them
OpenSmartRoute is an open-source router that manages AI requests. It sends each request to the best-fit model, agent, or tool from a catalog. It scores each candidate based on quality, cost, speed, and safety. Teams can set rules to keep requests within regions or cost limits.
Using OpenSmartRoute, a team can route requests to Fable 5.2 for virtual painting tasks. The router ensures the model gets the right prompts and manages costs effectively. It can also learn which models perform best for specific tasks and adjust traffic accordingly.
For physical tasks, OpenSmartRoute can coordinate requests between virtual planning and real-world execution. It helps teams optimize their use of Fable 5.2's planning and correction capabilities. This setup allows for efficient, controlled, and safe AI-driven painting and task execution.
How it compares - what existed before, what this changes and what stays the same
Before Fable 5.2, most AI image generation models used diffusion or generative adversarial networks (GANs). These models produce images by refining noise or combining existing images. They do not simulate physical painting processes.
Diffusion models, like Midjourney, generate images by gradually reducing noise. They rely on trained neural networks to improve the image step-by-step. These models are fast and produce high-quality results but lack physical realism.
GANs create images by learning from large datasets of real images. They generate new images by combining features learned during training. They are efficient but can produce less authentic textures and details.
Fable 5.2 changes this by simulating the physical process of painting. It places individual brushstrokes based on code instructions. This method results in more realistic textures and imperfections.
Unlike diffusion or GANs, Fable 5.2 does not start from noise or existing images. It builds an image from scratch through a sequence of strokes. This process is slower but can produce more natural-looking art.
What stays the same is the goal of creating visual content with AI. All models aim to generate images that meet user prompts. They differ in how they produce those images and the level of realism they achieve.
Fable 5.2 also differs in its ability to plan and self-correct. It can adapt its painting as it goes, similar to a human artist. This feature is not common in traditional image models.
Overall, Fable 5.2 introduces a new approach that emphasizes physical simulation. It complements existing models by offering a different way to generate images, especially ones that look hand-painted or textured.
Questions this leaves open - what the source does not say and how a reader can check it
The source does not specify how well Fable 5.2 performs on different types of images. It mentions a few examples, but not a broad range of prompts or styles. It is unclear how consistent or versatile the model is across various subjects.
It also does not provide detailed speed comparisons. While it states that Fable 5.2 is slower than some models, it does not give exact timing or how much slower it is in different scenarios. This information would help evaluate its practical use.
The source mentions that Fable 5.2 uses thousands of lines of Python code to control painting. It does not explain how easy or hard it is for developers to modify or extend this code. Knowing the complexity of customization would be useful for teams considering adoption.
Another open question is how well the model handles mistakes and corrections in complex scenes. The examples show some self-correction, but it is not clear how it performs with highly detailed or layered images. Testing with diverse prompts could reveal its strengths and limits.
The source also does not detail the hardware requirements for running Fable 5.2. Since it simulates physical properties and uses extensive code, it might need significant computational resources. Clarifying this would help teams plan their infrastructure.
Finally, it is not clear how the model's output quality compares to traditional image models in terms of user satisfaction or artistic value. While the process is more realistic, the subjective quality of the results remains to be seen.
To check these points, a reader can try running Fable 5.2 with different prompts and compare the results to diffusion-based models. They can measure the time taken for each image and observe how the model handles corrections. Reviewing the codebase or asking the developers about hardware needs and customization options can also provide insights.
In summary, while Fable 5.2 offers a novel approach to AI image creation, many aspects of its performance and usability remain to be explored through direct testing and further information from its developers.