3DCodeVerse

// code → 3D

Read the program, see the world it builds.

One case per language across all four tracks. The left pane is the actual source file that produced the render on the right; the GLSL shader is compiled and running in your browser, and the URDF assets move between their real joint states.

Blender Python3D objectCadQuery3D objectOpenSCAD3D objectThree.js3D objectURDFArticulatedURDFArticulatedThree.jsSceneBlender PythonSceneGLSLGraphicsOpenGLGraphics

nacre.py· 174 linesBlender Python

output· rendered by the runtime

NACRE · sectioned nautilus

57 chambers, nacreous layers and fine growth lines, on a stand that actually touches the shell.

// 3DCodeVerse-2M

A large-scale corpus of executable 3D code.

Programs come from public datasets, from converting specialized representations (DeepCAD sequences and Articraft geometry become standalone CadQuery), and from running the harness itself. Each program ships with the sources and assets it needs to execute, renders, three kinds of captions (instruction, visual description, construction procedure) and its verification record.

Packaged catalogue

269,597 programs across ten collections, expanded into 2,013,629 render and caption entries: the 2.03M code–condition pairs. 3DCodeBench factories, instances and test splits are held out for evaluation.

ilabai/3dcodeverse on Hugging Face

DeepCAD · DeepCAD → CadQuery

128,201

Shadertoy · GLSL

119,961

Articraft · CadQuery

6,882

Bioinspired3D · Blender Python

6,343

Articraft · URDF

6,146

Thingiverse · OpenSCAD

1,193

3DCodeBench factories · Blender Python

486

Agent-written · Three.js

280

Agent-written · Blender Python

80

Three.js repositories · Three.js / R3F

25

// 3DCodeVerse Harness

A data engine that checks its own work.

Visually plausible is not the same as geometrically valid. The harness pairs a VLM judge with deterministic measurements, turns every finding into a targeted edit, and loops without human feedback, scaling at inference time to projects with thousands of lines of code.

1. 01
   Plan

   An LLM planner turns the prompt into components, dimensions, materials, spatial relations and acceptance criteria.

2. 02
   Build

   The plan becomes a DAG of coding and assembly tasks. Sub-agents own separate files and write them in parallel.

3. 03
   Check

   Execution logs, then spatial tools: contacts, collisions, joint sweeps and placement, each reported per part in millimetres.

4. 04
   Judge

   A VLM judge reads labelled multi-view renders against the plan. Measurements cap the score and veto claims they rule out.

5. 05
   Fix

   Findings become localized repair tasks for a fix agent. Every round is kept, and the highest-scoring program wins.

auto-generated feedback → targeted edit

“Belly plates sink 3.2 mm into the clouds (≤ 2 mm).”
“Scales read as domed buttons; flatten the dome to 0.05w.”

adversarial loop engineering

During development the judge and the generator are refined in turns. The judge is tuned on a fixed bank of stored artifacts to reduce score variance and contradictions with measurements; then, with the judge frozen, harness designs are compared on matched prompts, trading judged quality against workflow failures.

// results

Same model, better harness. Same model, better data.

Across four backbones, outputs from each model's native harness are rarely preferred, while the same model inside the 3DCodeVerse harness wins most comparisons. Fine-tuning open Qwen models on 3DCodeVerse-2M turns mostly failing programs into runnable ones.

Gemini 3.8 Flash

vs. Gemini CLI

13%

87%

Claude Opus 5.5

vs. Claude Code

21%

79%

GPT-6 Astra

vs. Codex

31%

69%

GPT-6 Luna

vs. Codex

11%

89%

Claude Opus 5.5

vs. Claude Design

10%

90%

Gemini 3.8 Flash

vs. Gemini CLI

86.6

Claude Opus 5.5

vs. Claude Code

88.5

GPT-6 Astra

vs. Codex

87.7

GPT-6 Luna

vs. Codex

85.9

Claude Opus 5.5

vs. Claude Design

87.7

0255075100

Qwen3.8-27B (pretrained)Qwen3.8-27B + SFT + DPO

3DCodeBench

212 tasks

92.9%

Blender

103 tasks

93.2%

CadQuery

200 tasks

98.5%

OpenSCAD

50 tasks

90%

GLSL

200 tasks

71.5%

Three.js

40 tasks

92.5%

0255075100

Show as a table

Model	3DCodeBench(212)	Blender(103)	CadQuery(200)	OpenSCAD(50)	GLSL(200)	Three.js(40)
Qwen3.5-9B	0.0	1.9	10.5	50.0	33.0	72.5
+ SFT	84.9 +84.9	90.3 +88.4	98.5 +88.0	86.0 +36.0	56.0 +23.0	74.5 +2.0
Qwen3.8-27B	51.9	74.8	65.5	80.0	57.0	72.5
+ SFT + DPO	92.9 +41.0	93.2 +18.4	98.5 +33.0	90.0 +10.0	71.5 +14.5	92.5 +20.0

The preference and judge charts cover 100 text-to-3D object prompts. Human preference also reaches 78–88% on articulated objects, 50 scene prompts and 50 graphics prompts.

// gallery

24 worlds, each one a program.

Code-authored showcase scenes. Geometry, procedural materials, lighting, camera and animation are all kept as executable source. Hover to play a clip, and click to read the code next to the render.

Glasswing172 lines · py

The Gilded Kingdom461 lines · py

Meridian Outpost173 lines · py

Hangar 07257 lines · py

The Drowned Nave207 lines · py

River of a Thousand Lanterns166 lines · py

The Cloud Atlas193 lines · py

After Midnight203 lines · py

The Last Orchard120 lines · py

Boreal Observatory104 lines · py

The Jade Guardian160 lines · py

The Amber Passage120 lines · py

Sporelight126 lines · py

The Glasswing Atelier172 lines · py

The Hours of Totality126 lines · py

Pelagic Station167 lines · py

The Canopy Commons147 lines · py

Meridian Terminal204 lines · py

Between the Falls201 lines · py

Selene Departure117 lines · py

The Night Crossing95 lines · py

The Ember Works125 lines · py

The Little Bird Factory164 lines · py

The Pelagic Archive314 lines · py

These are showcase studies written with the harness and polished over several rounds. They are not the random samples or model comparisons behind the reported results.

// citation

BibTeX

@article{gao2026_3dcodeverse,
  title   = {3DCodeVerse: Exploring 3D Worlds as Code},
  author  = {Gao, Yipeng and Zeng, Ziyao and Huang, Jinfa and Li, Yijiang and Xiong, Haolin and Qin, Wang and Yu, Xinlei and Yao, Youheng and Zhang, Jie and Lin, Kun-Yu and Liu, Rong and Zhao, Yinyuan and Zhao, Yajie and Guo, Meiqi and Luo, Jiebo and Shu, Lei and Ge, Yunhao and Hu, Xinyu and Itti, Laurent},
  journal = {arXiv preprint},
  year    = {2026}
}