🚀 3DCodeVerse lets AI coding agents create entire 3D worlds using code instead of generating static meshes. And it reportedly beats Claude Design in 90% of human comparisons!

Researchers have introduced 3DCodeVerse, a new ecosystem for generating 3D objects, articulated characters, complex environments, animations, and even physics simulations through executable code.

The key difference from traditional AI 3D generators?

The output isn't a baked mesh. It's an editable, executable, and verifiable program.

That means AI agents can modify geometry, adjust materials, change animations, and rebuild entire scenes by editing code.

The project has three main components:

▪️ 3DCodeVerse-2M: A dataset containing approximately 269,600 executable 3D programs, expanded into 2.03 million code-condition pairs. Each program includes source files, required assets, descriptions, and verification data.

▪️ 3DCodeVerse Harness: A multi-agent framework that organizes 3D creation through a dependency graph. Agents write code, execute it, inspect geometry, analyze renders, identify problems, and automatically refine the results.

▪️ 3D coding research: Training recipes, evaluation tools, and practical recommendations for improving how AI models generate 3D content.

How does the agent workflow operate?

▪️ Plan: Break a request into objects, components, dimensions, materials, and spatial relationships.

▪️ Build: Multiple coding agents generate different parts of the scene in parallel.

▪️ Verify: Execute the code and check collisions, object placement, connectivity, and joint movements.

▪️ Evaluate: A vision-language model examines rendered views and identifies visual defects.

▪️ Refine: Agents automatically fix problems and repeat the process until they achieve better results.

All without requiring human intervention during generation.

The benchmark results are particularly interesting.

In the researchers' human evaluations of 100 text-to-3D object prompts:

▪️ 90% preference over Claude Design when both systems used Claude Opus 5.5.

▪️ 79% preference over standard Claude Code using the same model.

▪️ 69% preference over Codex using GPT-6 Astra.

▪️ 87% preference over Gemini CLI using Gemini 3.8 Flash.

These are results reported by the researchers, not independent evaluations.

Fine-tuning also produced major improvements: Qwen3.8-27B's execution rate on 3DCodeBench increased from 51.9% to 92.9%.

Supported environments include Blender Python, Three.js, CadQuery, OpenSCAD, URDF, GLSL, and OpenGL.

And the applications extend beyond static modeling:

▪️ Procedural game environments and reusable assets

▪️ Rigged characters and animated objects

▪️ Robotics and physics simulations

▪️ Programmable worlds with precise spatial control

▪️ AI video generation using 3D scenes as motion and camera references

This is a fascinating direction for AI-assisted game development.

Instead of asking AI to generate a finished 3D asset, you give it the ability to program, inspect, debug, and continuously improve an entire 3D world.

Project and interactive demos:
https://3dcodeverse.com/
GitHub:
https://github.com/gaoypeng/3dcodeverse

Research paper:
https://arxiv.org/abs/2610.22341
#AIAgents #3DGeneration #GameDev