Project3DCodeVerse

3DCodeVerse: Towards Building the Universe via Code

3DCodeVerse aims to gather any code that creates, renders, or governs 3D— every dialect, one corpus.3D Code Data

Datasets & Progress

Data format

Every entry is a self-contained, runnable project— the code, its rendered outputs, a metadata card, and text captions. One source folder at the top, coarse-category subfolders beneath (the same ones in the table above), and one folder per sample.

3dcodebench/                          one source folder per origin
  factories_geo/                      subfolder — a coarse category
    <sample>/                         one self-contained, runnable sample
      code.py                         the code — single file, or many; runs as-is
      renders/                        rendered outputs for this sample
        view_00.png  view_01.png  …   multi-view rendered images
        video.mp4                     video (optional)
        object.glb                    3D mesh — .glb / .obj / .stl (optional)
      meta.json                       identity card (see below)
      captions.json                   text descriptions, keyed by type (see below)
  factories_tex/                      … same structure

{
  "id": "factories_geo/0007",
  "type": "3D Objects",            // Attribute — what it is
  "language": "Blender Python",    // code dialect
  "entry": "code.py",              // file to run
  "multi_file": false,             // single file, or a multi-file project
  "environment": "Blender 5.0",    // exact runtime needed to execute
  "renders": [                     // artifacts under renders/
    "renders/view_00.png",
    "renders/object.glb"
  ],
  "source": "3DCodeBench",         // origin dataset
  "license": "commercial-ok",      // reuse terms
  "curator": "yipeng",             // who reviewed / curated it
  "status": "curated"              // curated | pending | revise
}

{
  "brief":    "A wooden four-drawer dresser.",
  "detailed": "A rectangular wooden dresser with four stacked drawers, each with a small round knob, a flat top and short tapered legs.",
  "geometry": "Box body ~0.8 x 0.45 x 1.1 m, split into four equal drawer compartments; cylindrical handles centered on each drawer front.",
  "function": "Bedroom storage; the four drawers slide out along the front face."
}

On our radar — candidate sources

3DCodeVerse aims to unify multiple dialectsof 3D code under one quality-gated roof. Today: Infinigen (organic / natural, Blender Python). On our radar — all permissively licensed, so combinable with attribution:

Open challenges

Where the harness work lives.

• •Translators / executors — command-sequence → code, .blend → Blender Python, AI2-THOR → Blender scene scripts, FeatureScript ↔ CadQuery.
• •Cross-dataset dedup — DeepCAD ↔ CADFS ↔ Omni-CAD ↔ CAD-Coder share lineage; geometry / ID dedup before merge.
• •License & attribution tracking — keep provenance per entry so the merged corpus stays redistributable.

Best practices

Lessons from building the pipeline so far.

1. 1
   Coding agent does the conversion

   A coding agent transcribes each asset into standalone, runnable code — the heavy lifting of the asset → code transform.

2. 2
   Human feedback closes the loop

   Humans keep / drop / revise; their feedback drives the auto-refine loop instead of hand-rewriting code.

3. 3
   Give the agent eyes

   Build visualization tools the coding agent can call — render previews, diffs — so it verifies and self-corrects its output instead of working blind.

4. 4
   Anomaly detection as a final gate

   After curation, sweep for outliers — over-long code files, excessive character counts, degenerate geometry — to catch what slips past human review.

5. 5
   Scale with Claude Code

   The 20× (Max) plan plus headless mode (claude -p) lets you batch-convert and clean data in bulk, not one file at a time.

Contribute data — how it works

Bring your own (code → 3D) projects into the corpus with the 3dcode toolkit. Data uploads straight from your own machineto a private staging bucket, gets auto-validated + dedup-checked, then a maintainer reviews and ingests it into the corpus. Your bandwidth, not ours — nothing routes through a central server.

<source>/                  one source per origin
  <sample>/                one folder per data sample
    model.py               the code (runnable)
    renders/               rendered images + .glb mesh
    prompt.txt             text prompt
    meta.json              auto-generated (id, dialect, hashes, status)

• Validates

  layout + anomaly checks before anything uploads

• Fingerprints

  code / geometry / visual hashes — flags duplicates vs the corpus

• Uploads direct

  from your machine to the private bucket, no central bottleneck

Toolkit & docs: github.com/gaoypeng/3dcode_toolkit