Abstract
Internet Blender tutorials contain rich, real-world creation knowledge, but that knowledge is difficult for an agent to use directly. Important instructions may appear in narration, on-screen captions, changing interface states, or brief node-graph operations, while the final result alone does not reveal how the asset was constructed. We present BlenderLore, an agent pipeline that converts tutorial videos into timestamped multimodal evidence, reconstructs the demonstrated workflow, and generates executable Blender Python from a canonical scene.
The agent combines visual keyframes, OCR, available transcript evidence, Blender-version cues, and retrieved workflow knowledge to build a bounded reconstruction specification. Generated programs are executed inside Blender and evaluated through fresh renders, multi-view observations, and turntable or animation evidence when motion is supported by the tutorial. The primary output is an editable .blend asset rather than a flattened image.
Generalization is central to BlenderLore. Successful reconstructions are retained as candidate procedural knowledge. When facing an unfamiliar generation or editing task, the agent decomposes the target into reusable construction patterns, retrieves relevant procedural knowledge and recombines it into task-specific Blender code.