Research code availableProcedural 3D · Blender · Agentic AI

TreeStruct3D

Enabling structural editability in agentic procedural 3D modeling.

Generate Blender programs whose semantic parts stay attached when a parent or child is independently resized—not just assets that look correct once.

Explore the repositorySee how it works
Structure-aware generationOpen-source implementation

Controlled editing at a glance

Structure that survives the edit.

TreeStruct3D represents each object as a directed tree of semantic parts, realizes every attachment with shared anchors, then verifies those relationships after controlled parameter changes.

Controlled editing comparison between 3DCodeBench and TreeStruct3D across four vision-language models
Figure 01 · Controlled editability comparisonDefault · parent scale · child scale

System pipeline

From a prompt to a verified program.

The system makes structure explicit before code generation and keeps structural verification inside the agentic loop.

  1. 01 / 04

    Extract

    Structure blueprint

    Turn the object description into semantic parts, a root, directed attachments, and shared-anchor requirements.

  2. 02 / 04

    Generate

    Blender program

    Condition code generation on the text prompt and the separately extracted structure blueprint.

  3. 03 / 04

    Validate

    Structural behavior

    Execute the program and check hierarchy, contact, anchors, and independent part perturbations.

  4. 04 / 04

    Repair

    Localized feedback

    Return precise structural failures for repair before the final visual-refinement pass.

Editable by construction

Appearance is a snapshot. Structure is behavior.

A plausible render can still break when one part changes. TreeStruct3D evaluates the generated program as an editable system: each semantic parameter must update geometry without invalidating its attachments.

Static quality

Does it look right?

Structural quality

Does it keep working?

Generated program protocolPython · Blender
PART_PARAMS = {
  "trunk":  { "scale": 1.0 },
  "branch": { "scale": 1.0 },
}

SHARED_ANCHORS = {
  "trunk_to_branch": {
    "parent": "trunk",
    "child":  "branch",
    "world_position": anchor,
  }
}

# Parent and child edits are checked independently.
build_scene(PART_PARAMS, SHARED_ANCHORS)

Reproducible input boundary

Built on a pinned benchmark snapshot.

212

categories

3DCodeBench evaluation split

636

source files

Verified against one revision

01

model input

Description or instruction text

Upstream provenance

The included 3DCodeBench split comes from the official YipengGao/3DCode dataset and is pinned to an immutable revision.

Dataset on Hugging Face3DCodeBench source

Start with one object

One config. Two model-facing phases.

API format, endpoint, credentials, model identity, token policy, and retries all come from the selected YAML configuration.

Open the quick start
Bird_seed0config.local.yaml
cp configs/config.example.yaml config.local.yaml
export TREESTRUCT3D_API_KEY=your-api-key

./extract_structure.sh \
  --config config.local.yaml \
  --instances Bird_seed0

./generate_3d.sh \
  --config config.local.yaml \
  --instances Bird_seed0

Qualitative editing gallery

Across four vision-language models.

Each panel shows the default asset beside exaggerated 0.4× and 1.6× parent- and child-side edits. Quantitative evaluation uses 0.8× and 1.2×.

(a) GPT-5.5
Figure 06
TreeStruct3D controlled editing examples generated with GPT-5.5
(b) GPT-5.6 Sol
Figure 06
TreeStruct3D controlled editing examples generated with GPT-5.6 Sol
(c) Gemini 3.1 Pro
Figure 06
TreeStruct3D controlled editing examples generated with Gemini 3.1 Pro
(d) Gemini 3.5 Flash
Figure 06
TreeStruct3D controlled editing examples generated with Gemini 3.5 Flash