README.md

---
dataset_info:
  features:
  - name: item_id
    dtype: string
  - name: prompt
    dtype: string
  - name: domain
    dtype: string
  - name: level
    dtype: string
  - name: conditioning
    dtype: string
  - name: reference_image
    dtype: image
  - name: nova3d_preview
    dtype: image
  - name: nova3d_code
    dtype: string
  - name: meshy_preview
    dtype: image
  - name: trellis2_preview
    dtype: image
  - name: triposg_preview
    dtype: image
  - name: partcrafter_preview
    dtype: image
  - name: cubepart_preview
    dtype: image
  - name: meshanything_preview
    dtype: image
  - name: cadcoder_preview
    dtype: image
  - name: cadcoder_code
    dtype: string
  - name: blenderllm_preview
    dtype: image
  - name: blenderllm_code
    dtype: string
  - name: text2cad_preview
    dtype: image
  - name: text2cadquery_preview
    dtype: image
  - name: text2cadquery_code
    dtype: string
  - name: llamamesh_preview
    dtype: image
  - name: naive_blender_llm_preview
    dtype: image
  - name: naive_blender_llm_code
    dtype: string
  splits:
  - name: train
    num_bytes: 96231418
    num_examples: 54
  download_size: 95198118
  dataset_size: 96231418
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: cc-by-4.0
task_categories:
- text-to-3d
- image-to-3d
pretty_name: Nova3D-Bench (preview)
size_categories:
- n<1K
---

# Nova3D-Bench (preview)

A frozen, spec-grounded benchmark from *Nova3D: Code-Native Generation of
Programmable 3D Assets*: 54 prompts across six domains and three difficulty
levels, each with a machine-checkable ground-truth spec. It compares
**code-native** 3D generation — a program that produces the model, not just
the model — against mesh-native, CAD, and part-structured baselines.

This is a **preview release**: one row per benchmark item, one column per
pipeline. See [Data completeness](#data-completeness) below — full `.glb`
models are currently only available for 2 of the 13 pipelines; the rest are
render previews only, pending baseline re-runs.

## Dataset summary

| | |
|---|---|
| Benchmark items | 54 |
| Domains | 6 (architecture, characters, furniture, mechanical, tools, vehicles) |
| Difficulty levels | 3 (L1 ≤8 parts / L2 9–25 parts / L3 >25 parts) |
| Conditioning | 24 text-only, 30 image+text |
| Pipelines run | 13 (Nova3D + 12 baselines) |

## Schema

One row per benchmark item. Metadata columns, then one `<pipeline>_preview`
image column per pipeline, plus a `<pipeline>_code` text column for the 5
pipelines that emit source.

| Column | Type | Description |
|---|---|---|
| `item_id` | string | e.g. `arch-L1-01` |
| `prompt` | string | Full instruction given to every system (image items still get a text prompt — see note below) |
| `domain` | string | One of the 6 domains |
| `level` | string | `L1` / `L2` / `L3` |
| `conditioning` | string | `text` or `image` |
| `reference_image` | image | Synthetic reference photo/sketch; a shared "n/a" placeholder for text-only items |
| `<pipeline>_preview` | image | Rendered textured view of that pipeline's output; "n/a" placeholder if the pipeline doesn't apply to this item's conditioning type, or the run failed |
| `<pipeline>_code` | string | Generated source (Nova3D `code.py`, CAD-Coder `model.py`, BlenderLLM `model.bpy.py`, Text2CADQuery `code.py`, naive-Blender-LLM `code.py`); `null` where not applicable |

**Note on `reference_image`/`<pipeline>_preview`:** every Image column always
holds a real image, never `null` — the Hugging Face Viewer renders `null` in
an Image column as a broken "Not supported with pagination yet" cell instead
of blank, so a shared dark "n/a" placeholder is used wherever nothing
applies.

**Note on prompts:** image-conditioned items are still fully text-prompted —
e.g. `"Recreate the street bollard shown in the reference photo."` — and any
numeric constraints on an image item are spelled out in the prompt text too,
since a constraint the system can't see in text isn't a fair test.

## Pipeline registry — 13 total

| Category | Pipelines | What it ships |
|---|---|---|
| Code-native | Nova3D, BlenderLLM | A program that produces the model |
| CAD | CAD-Coder, Text2CAD, Text2CADQuery | Parametric solids via a CAD kernel (CadQuery / STEP) |
| Part-structured | PartCrafter, CubePart | Mesh output with explicit part segmentation |
| Mesh-native | Meshy, TRELLIS.2, TripoSG, MeshAnything, LlamaMesh, naive-Blender-LLM* | Direct mesh surface, no source, no part structure |

\* naive-Blender-LLM is an **ablation of Nova3D itself** (same underlying
LLM route, no master prompt / repair loop / validation pass) — included to
isolate what Nova3D's engineering contributes, not a third-party system.

Coverage varies by pipeline because most baselines only support one
conditioning type: Nova3D and Meshy ran on all 54 items; the 5 image-only
baselines (TRELLIS.2, TripoSG, PartCrafter, CubePart, MeshAnything) only
apply to the 30 image-conditioned items; the 5 text-only baselines
(CAD-Coder, BlenderLLM, Text2CAD, Text2CADQuery, naive-Blender-LLM) only
apply to the 24 text-conditioned items, and further drop below 24 from real
generation failures on that baseline.

## Data completeness

This preview currently ships **render previews for every pipeline** but
**full `.glb` models for only 2 of the 13 pipelines**:

| Status | Pipelines |
|---|---|
| `.glb` available | Nova3D (54/54), naive-Blender-LLM (39/54 — rest were failed exports) |
| `.glb` not currently recoverable | Meshy, TRELLIS.2, TripoSG, PartCrafter, CubePart, MeshAnything, CAD-Coder, Text2CAD, Text2CADQuery, LlamaMesh, BlenderLLM |

Nova3D and naive-Blender-LLM route through persistent Azure Blob Storage, so
their models were recoverable after the fact. The other 11 baselines were
run against ephemeral Cloudflare Quick Tunnels or local/private-network
servers at generation time; those endpoints no longer exist, so their
`.glb`s are not recoverable from this repo and would require re-running the
baseline or locating a separate archive of the original outputs.

## License

CC BY 4.0. No named or trademarked objects are included in the benchmark
(see IP policy in the paper's dataset section).