CGAxis/cgaxis-3d-models-sample

CGAxis 3D Models - Free Sample (Furniture / Chairs)

A free, licensed sample of human-authored 3D models from CGAxis, a 3D content studio operating since 2008. This sample is a taster of the full CGAxis AI Data corpus (4,200+ 3D models + 7,913 PBR material sets) available for commercial AI-training licenses.

Every model ships as GLB and USDZ (the USDZ with UsdPhysics authored: rigid body, collision, mass, physics material), with geometry statistics, real-world scale in centimetres, semantic tags, a natural-language caption, per-file SHA-256 and a machine-readable compliance block, plus a dataset-level MLCommons Croissant file for direct loading into ML pipelines.

Why this dataset is different

Most 3D training data is scraped from the open web (objects pulled under contested or unknown licenses, with no one to indemnify you when the model ships). This one is not:

• Single owner, clean provenance. Human-authored in-house by CGAxis since 2008. No scraping, no unknown rights, no user-generated content.
• Real-world scale. Each model carries its bounding box in centimetres, derived from the source geometry - not normalised-to-unit-cube like most 3D datasets. Essential for physically consistent generation, simulation and robotics.
• Production-ready formats. GLB (the ML ingestion standard) and USDZ (USD / Apple / simulation pipelines) for every model. Clean UVs.
• Simulation-ready USDZ. Every USDZ carries standard UsdPhysics schema (RigidBodyAPI, CollisionAPI with convex decomposition, MassAPI with a per-material density, physics material), so it drops into NVIDIA Omniverse / Isaac Sim and other PhysX-based simulators as a dynamic body.
• AI-native metadata. Per-model metadata.json (geometry stats, bbox, tags, caption, per-file SHA-256, compliance) plus a loadable croissant.json and a deterministic train/test split.

This sample

Property	Value
Models	40 (furniture - chairs, dining chairs, armchairs, lounge)
Formats	GLB + USDZ (both, every model)
USDZ with UsdPhysics (rigid body + collision + mass)	40 / 40
With real-world scale (bbox in cm)	40 / 40
UV-mapped	40 / 40
Triangles (min / median / max)	2,732 / 64,220 / 1,256,392
Preview render	40 / 40 (WebP / JPG)
Per-file SHA-256	40 / 40
Caption + semantic tags	40 / 40
Split	38 train / 2 test (deterministic, seed cgaxis-3d-v1)

Intended uses

Training and evaluating geometry- and material-aware 3D models: text-to-3D, image-to-3D, single- and multi-view 3D reconstruction, mesh/shape generative models, 3D asset retrieval, and (via the baked PBR textures inside each GLB) material-aware 3D generation. The real-world scale also makes the set useful as licensed assets for simulation / embodied-AI scenes.

Load the metadata (Croissant)

import mlcroissant as mlc
ds = mlc.Dataset("croissant.json")
records = list(ds.records(record_set="models"))

Or read models_index.json directly (full per-model records, including geometry, bbox, tags, caption and the split field).

Structure

/                                  # one folder per model
  <model_id>/
    <Model>.glb                    # glTF binary (geometry + baked PBR textures)
    <Model>_USDZ.zip               # USDZ with UsdPhysics, packaged in a zip
    <model_id>.render.jpg|.webp    # preview render
    <model_id>.metadata.json       # geometry, bbox_cm, tags, caption, per-file sha256, compliance
croissant.json                     # MLCommons Croissant 1.0 (loadable)
models_index.json / models_index.csv   # all models (JSON has the split field; CSV backs Croissant)
splits.json + train.txt / test.txt     # deterministic train/test split
DATASET_CARD.md                    # plain-text summary
LICENSE.txt                        # this sample's license (evaluation only)

Compliance (EU AI Act)

Every model carries a machine-readable compliance block (vendor assertions): human_authored=true, ai_generated=false, contains_pii=false, contains_trademarks_or_logos=false, web_scraped=false, plus a training_data_summary - so training-data documentation is trivial. CGAxis is an EU company.

License

Free for internal evaluation and prototyping under the CGAxis AI Data Free Sample License (see LICENSE.txt). This sample is for assessing quality, format and metadata, and building internal proofs of concept - not for training models you ship or sell, and it comes without warranty or indemnification.

License the full corpus (4,200+ 3D models + 7,913 PBR sets)

Production training - with an IP warranty and full indemnification - is licensed per Category Pack, Full Corpus or Exclusive tier.

• 3D models: GLB + USDZ, real-world scale in cm, clean UVs, Croissant + per-file SHA-256.
• PBR materials: metallic-roughness and specular-glossiness, GL + DX normals, 16-bit displacement, real-world tile size.
• EU AI Act training-data summary, deterministic train/test splits.

Request a quote or a larger sample: cgaxis.com/ai-training-data

Changelog

• v1.1.0 (2026-10-06) - 22 models replaced with other chairs from the catalogue; every USDZ now carries UsdPhysics; wooden_chair_3d_model_9904 GLB and USDZ re-exported clean (stray vertices removed, grounded, 44.9 x 46.4 x 82 cm); index, Croissant and per-model metadata regenerated.
• v1.0.0 (2026-07-13) - first release.

Citation

@misc{cgaxis_ai_data_3d_2026,
  title  = {CGAxis AI Data: Licensed 3D Models and PBR Materials for AI Training},
  author = {CGAxis (DTS S.C.)},
  year   = {2026},
  url    = {https://cgaxis.com/ai-training-data/}
}