The agent separates the character, background, text, and outline into transparent layers, then offsets them at different depths with view-dependent coordinates to create parallax. Alpha masks preserve each silhouette, while procedural Voronoi and wave patterns add iridescent highlights, sparkles, and a gold outline. Animated reflections complete a holographic foil card that shifts as it moves.
BlenderLore
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.
Method
How It Works
Start from high-quality internet Blender tutorials and identify the target asset, workflow, and supported motion.
Align keyframes, OCR, narration or audio, interface actions, timestamps, and Blender-version cues.
Convert evidence into ordered operations and retrieve relevant procedural knowledge.
Generate Blender Python, execute it in Blender, render the scene, compare the result, and repair failures in a closed loop.
Package editable assets and visual evidence, then retain validated workflow patterns as candidate procedural knowledge for reuse.
What You Get
Each run recreates a tutorial workflow and delivers an editable Blender project, a reproduction script, renders, and validation results.
asset.blendreproduce.pyThrough collection, repair, and reconstruction, we built a high-quality dataset of 23K procedural 3D assets.
The procedural knowledge library serves as a key mechanism for generalization: when facing an unfamiliar target, the agent decomposes it into reusable construction patterns, retrieves relevant procedural knowledge, and recombines it into task-specific Blender code.
Tutorials become executable, editable assets.
Source image and tutorial cover to editable Blender material.
Blender material nodes
image = bpy.data.images.load(SOURCE_IMAGE, check_existing=True)
voronoi = nodes.new("ShaderNodeTexVoronoi")
voronoi.feature = "DISTANCE_TO_EDGE"
set_input(voronoi, "Scale", 50.0)
bpy.ops.wm.save_as_mainfile(filepath=str(OUTPUT_DIR / "asset.blend"))Tutorial-learned surfaces, shaders, and animated effects.
Editable forms, scientific visualizations, and compact scene studies.
Learned techniques reshape existing assets while preserving what should stay.
Original yellow asset and material tutorial to a generated chromatic result.
Blender material nodes
noise_large = nodes.new("ShaderNodeTexNoise")
layer_weight = nodes.new("ShaderNodeLayerWeight")
links.new(layer_weight.outputs["Facing"], tint_mix.inputs[0])
links.new(tint_mix.outputs["Color"], principled.inputs["Base Color"])
bpy.ops.wm.save_as_mainfile(filepath=str(OUTPUT_DIR / "asset.blend"))
One material language, transferred across forms.
Citation
@misc{blenderlore2026,
title = {BlenderLore: Learning 3D Coding from Internet Tutorial Videos},
author = {FreedomAI Team},
year = {2026},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/FreedomIntelligence/BlenderLore}},
}