synthclaw

What This Skill Does

The synthclaw skill acts as a powerful bridge between the OpenClaw AI agent and Blender, enabling the programmatic generation of synthetic imagery. By interfacing directly with Blender's backend, this skill allows users to manipulate procedural Value Nodes within .blend files and trigger renders using either the high-fidelity Cycles engine or the rapid, real-time EEVEE engine. It is specifically designed for high-throughput synthetic data generation where precise control over scene variables—such as scale, texture, displacement, or lighting intensity—is required for machine learning model training. The skill can optionally compute validation metrics like LPIPS (Learned Perceptual Image Patch Similarity) and naturalness scores, making it a critical tool for automated dataset creation and iteration.

Installation

To begin using the synthclaw skill, ensure that you have Blender 4.0 or higher installed on your host system and accessible via the terminal path. The skill requires Python 3.10+ for optimal execution. Installation is handled via the OpenClaw CLI:

clawhub install openclaw/skills/skills/ayakimovich/synthclaw

Once installed, verify that blender --version returns successfully in your console to ensure the agent has the necessary permissions to execute render commands.

Use Cases

• Synthetic Data Generation: Ideal for training computer vision models where you need thousands of variations of a single object or scene by sweeping through specific parameter values.
• Automated Parameter Optimization: Use the agent to iteratively adjust procedural materials or lighting setups until the render achieves a target visual quality, verified by the built-in metric computation.
• Procedural Dataset Creation: Generate consistent datasets with ground-truth metadata, such as depth maps or object masks, by systematically iterating through scene configurations.

Example Prompts

1. "Using /home/user/assets/character.blend, please render 5 variations of the 'TextureRoughness' parameter ranging from 0.1 to 0.5 using the EEVEE engine for testing."
2. "Perform a production-grade render of /home/user/scene.blend with 'LightIntensity' set to 0.8 and 'GrainScale' at 2.2. Save it to /output/final.png using the Cycles engine."
3. "Render the current procedural scene using Cycles with 512 samples. Also, calculate the LPIPS metrics comparing the result against /reference/ground_truth.png."

Tips & Limitations

• Efficiency: Use EEVEE for iterative testing, as it provides near-instant feedback for material or light node adjustments. Reserve Cycles for the final production passes due to the higher computational cost.
• Environment: This skill is strictly for offline batch processing. Do not attempt to use it for real-time visual previews or interactive scene manipulation.
• Asset Complexity: Ensure your .blend files have clearly labeled Value Nodes. The skill relies on direct node manipulation, so complex dependencies outside the Value Node system may not be affected by the agent’s parameter adjustments.