Expanding our AI 3D Toolkit: Exploring Sequence-Based Mesh Generation with MeshGPT

A while ago, we shared our Image-to-3D demo with the Peerlist community. We received great feedback on that pipeline, and we appreciate everyone who took the time to test it. As we continue building and exploring the AI 3D generation space, we wanted to experiment with a fundamentally different underlying architecture.

Today, we are opening up access to a new experimental demo based on MeshGPT. Instead of relying on implicit representations, this tool treats 3D mesh generation purely as a sequence prediction task.

What to Expect in the Demo

This tool is designed to focus strictly on structural topology and clean geometry. When you run a generation prompt, here is what the output delivers:

• Native Triangulation: The model does not generate point clouds or rely on surface extraction algorithms like Marching Cubes. It directly outputs pure meshes containing explicit vertex coordinates and connected faces.

• Compact, Low-Poly Outputs: The generated assets maintain a minimal polygon count with sharp, defined edges. We highly recommend toggling the wireframe view in the demo to inspect the routing structure—it is highly organized and avoids the messy clusters typical of older generation methods.

• Engine-Ready Geometry: Because the output avoids over-tessellation, the exported .obj or .gltf files bypass the need for intensive retopology. They can be imported directly into WebGL environments, Three.js projects, or standard game engines with minimal performance overhead.

The Underlying Mechanics

For the builders curious about the architecture, MeshGPT applies the logic of Large Language Models to spatial and geometric data. It operates on two main principles:

• Geometric Vocabulary: The system utilizes graph convolutions to encode local 3D shapes and topologies into discrete tokens. It essentially builds a finite "dictionary" of geometric components.

• Autoregressive Generation: Just as an LLM predicts the next word in a sentence, this Transformer-based model autoregressively predicts the next triangular face (including its positional and relational data) until the entire object is fully assembled.

Try It Out

We built this demo so developers and creators can get hands-on with this specific type of generative pipeline. It is particularly effective at handling hard-surface objects, such as furniture, basic architectural elements, and low-poly props.

[MeshGPT]

We invite you to break it, test its limits, and see how the geometry holds up in your environment. Drop your feedback, note your generation times, or share any interesting topological bugs you encounter in the comments below.