If you're creating 3D animations with AI agents, there's one file format you should seriously consider: GLB. And the interesting part is that your agent might not need Blender at all. Just a 3D model, some JavaScript, and a browser.
Let me explain. GLB is the binary container format for glTF 2.0, an open standard developed by the Khronos Group for delivering 3D scenes and assets. Think of it as a convenient package containing the information needed to reconstruct an entire 3D scene.
Inside, you can have:
Scene hierarchy: Objects, components, and their relationships.
Geometry: Vertices, triangles, and other mesh data.
Materials and textures: How everything should look.
Animations: Keyframes, transforms, and movement.
Named objects: Individual components that can be accessed and manipulated through code.
And the entire scene can be packaged into a single .glb file.
Here's why this matters for AI agents.
Three.js has excellent support for glTF 2.0 through its GLTFLoader.
An agent can load a GLB model, inspect its scene hierarchy, find specific objects by name, and manipulate them directly with JavaScript.
Imagine a model containing a human skeleton.
Instead of moving thousands of individual vertices, the agent can identify objects such as: Skull Spine Rib cage Pelvis Individual vertebrae Left and right arms Then it can animate, rotate, highlight, or hide those components independently. It's almost like working with a digital construction kit.
Of course, this requires the GLB to contain properly separated and named objects. Not every downloaded model is organized that way. But when it is, the workflow becomes remarkably straightforward. And here's my favorite part.
The agent doesn't necessarily need to open a traditional 3D application.
It can:
1. Load the GLB model.
2. Write a Three.js scene.
3. Position the camera and lights.
4. Animate individual objects through code.
5. Render a frame in the browser.
6. Inspect the result and make corrections.
With browser automation and image inspection tools, the agent can even iterate visually on its own output.
Write code. Render. Look. Adjust. Repeat.
Hugging Face is becoming a useful source of ready-to-use 3D assets. There are datasets containing furniture, household objects, mechanical components, scanned artifacts, architectural elements, and much more. And some of them are surprisingly unusual.
For example, this entire animation was built using real anatomical data.
I used the CT Anatomy GLB dataset by CodePawl, available on Hugging Face.
The underlying medical data originates from TotalSegmentator, a research dataset developed at University Hospital Basel in Switzerland.
Researchers annotated 117 anatomical structures across 1,228 real clinical CT scans.
We're talking about actual medical imaging data, not anatomically questionable models invented by a text-to-3D generator.
The GLB collection converts a small selection of those segmented scans into structured 3D assets.
The published subset contains:
7 reconstructed anatomical scans
20 additional organ and body-part assets
Individually named anatomical structures
Optimized GLB geometry suitable for browser visualization
You can inspect individual organs, isolate anatomical systems, animate components, and build educational visualizations. And that's exactly what makes this format so useful.
The AI agent doesn't need to invent the anatomy. The geometry already exists.
Its job is to understand the scene, write the animation, control the camera, and produce the final visual. That separation of responsibilities is important.
Generative AI is excellent at helping write code and construct workflows. But it doesn't always need to generate every asset from scratch.
Sometimes the most reliable approach is to combine existing structured 3D data with an agent that knows how to manipulate it. And that's why I think formats like GLB deserve much more attention in AI-assisted creative workflows.
Explore GLB datasets on Hugging Face: https://huggingface.co/datasets?sort=likes&search=glb
CT Anatomy GLB dataset used in this video: https://huggingface.co/datasets/codepawl/ct-anatomy-glb
Original TotalSegmentator medical dataset: https://zenodo.org/records/10047292
Three.js GLTFLoader documentation: https://threejs.org/docs/pages/GLTFLoader.html
Official glTF 2.0 specification: https://registry.khronos.org/glTF/specs/2.0/glTF-2.0.html
#GLB #ThreeJS #AIAgents #3DAnimation #GLTF #CreativeCoding #AIAnimation #OpenSource3D
YouTube / ABVX