A sample from the A23D corpus of 100,000+ human-authored PBR materials, spanning a wide range of real-world material categories. This sample is provided for AI teams to inspect the quality, consistency, and per-map ground truth of A23D materials before licensing the full corpus.

Every material is authored in-house to a single production standard — giving consistency across resolution, channel packing, map conventions, naming, and metadata that open-source and marketplace texture sets do not provide.

Key highlights

Human-authored, single production standard. Every material is created in-house to one specification. No web-scraped, user-generated, or AI-generated content in the corpus.

Complete, explicit PBR ground truth. Each material ships a full set of physically based maps with explicit color space, bit depth, channel count, and normal-map convention — the per-map metadata AI training pipelines need, not just flat image files.

Rich, structured metadata. Every asset carries a natural-language caption, category hierarchy, material attributes, resolution, tiling and workflow declarations, and a machine-readable list of exactly which maps it ships.

Consistent at scale. The same schema, naming and conventions hold across all 100,000+ materials. What you validate on these 25 sets is what you get on the rest.

What's in the sample


Material sets	25
Images	268 — 219 maps, 24 previews, 25 thumbnails
Texture resolution	4096 × 4096, seamless, power-of-two
PBR workflow	metallic-roughness, DirectX normal convention
Captions	25 natural-language descriptions, 76–121 words each
Taxonomy	10 level-2 categories, 17 level-3
Total size	2.05 GB

Categories represented: Wood, Fabric, Metal, Plaster, Brick, Stone, Rock, Tiles, Glass, Wicker.

Repository structure

metadata.jsonl          one row per material, 26 columns — loads directly
metadata/<SKU>.json     full record per material, including per-bitmap detail
maps/<SKU>/             the PBR map stack — Albedo.png, Normal.png, Height.exr, …
previews/<SKU>.png      1920×1080 render, material tiled on a flat plane
thumbnails/<SKU>.png    750×750 render, material on a sphere

Paths in metadata.jsonl are relative to the repository root, and every path is constructible from the SKU alone. No filename contains a space.

Map stack

Eight base maps ship with every material, so ingestion never has to branch on which channels exist. Additional maps appear where the material calls for them — the transmissive glass set carries IOR and translucency channels.

Map	Sets	Format	Colour space	Bit depth	Channels
`Albedo`	25	PNG	sRGB	8	3
`Normal`	25	PNG	Linear	8	3
`Roughness`	25	PNG	Linear	8	1
`Metallic`	25	PNG	Linear	8	1
`Ambient Occlusion`	25	PNG	Linear	8	1
`ORM`	25	PNG	Linear	8	3
`Opacity`	25	PNG	Linear	8	1
`Height`	25	EXR	Linear	16 / 32 float	1
`Specular Level`	13	PNG	Linear	8	1
`Specular`	3	PNG	Linear	8	1
`IOR`	1	PNG	Linear	8	1
`Translucency`	1	PNG	Linear	8	1
`Translucency Roughness`	1	PNG	Linear	8	1

Conventions

Declared explicitly so results are reproducible and no team has to guess:

ORM packs Occlusion, Roughness and Metallic into one RGB image, in that channel order.
Colour management — only Albedo is display-referred. Every other map is data and must be sampled linearly; treating them as sRGB produces physically wrong shading.
Normal maps follow the DirectX convention, green channel +Y down. Flip green for OpenGL or Blender.
Height is single-channel float OpenEXR, stored as displacement — higher value means higher surface. Bit depth is declared per file in each material's record, read from the file itself.
Tiling — every material is seamless and power-of-two, declared per record as tileable and power_of_two.

Metadata

metadata.jsonl loads directly with datasets or pandas; metadata/<SKU>.json carries the same record plus full per-bitmap technical detail.

Field	Type	Description
`sku`	string	Stable asset ID. Primary key across the full corpus
`name`	string	Short human-readable title
`human_authored`	bool	Provenance attestation — `true` for every asset in the corpus
`caption`	string	Natural-language description of the material
`category_l1` … `category_l4`	string	Three-level category hierarchy
`material_type`, `pattern`, `finish`, `surface_texture`, `effects`, `color`, `condition`	string	Structured material attributes
`resolution`	string	Texture resolution
`tileable`, `power_of_two`	bool	Tiling guarantees
`pbr_workflow`	string	`metallic-roughness`
`normal_convention`	string	`DirectX`
`maps`, `maps_count`	list, int	Machine-readable declaration of which maps this material ships
`map_paths`	list	Paths parallel to `maps`
`total_size_mb`	float	Measured size of the map stack
`image`, `preview`	image	Thumbnail and preview render, decoded by the viewer
`metadata`	string	Path to the full per-material record

Per-bitmap detail in metadata/<SKU>.json gives colorspace, bit_depth, channels, format and resolution for every individual file.

Usage

from datasets import load_dataset

ds = load_dataset("A23D/a23d-pbr-materials-sample", split="train")
ds[0]["caption"]    # natural-language description
ds[0]["preview"]    # 1920x1080 preview render
ds[0]["image"]      # 750x750 thumbnail

Text–image pairs for contrastive or generative training:

pairs = [(r["caption"], r["preview"]) for r in ds if r["preview"] is not None]

The map stack is referenced by path rather than decoded, since each material carries 8–12 maps:

import os
from PIL import Image
from huggingface_hub import snapshot_download

root = snapshot_download("A23D/a23d-pbr-materials-sample", repo_type="dataset")
row = ds[0]
stack = {m: os.path.join(root, p) for m, p in zip(row["maps"], row["map_paths"])}
albedo = Image.open(stack["Albedo"])

Height.exr requires an OpenEXR-capable reader — opencv-python with OPENCV_IO_ENABLE_OPENEXR=1, or OpenImageIO. It will not open with Pillow.

The full corpus

This sample is a small slice of the complete A23D library:

100,000+ human-authored PBR materials
Broad coverage across architectural surfaces, natural materials, fabrics, metals, wood, stone, ground and terrain, and many more
Uniform schema and production standard across every material
Complete provenance — authored by our in-house team
Bulk delivery, custom curation, and category-targeted subsets

Available for licensing for AI and ML training, including generative model development, and for robotics, embodied AI, world models, digital twins, synthetic data and spatial AI.

Get in touch:
Website: www.a23d.co/ai
Email: enterprise@a23d.co

License

This sample is released under the A23D Sample Dataset License for evaluation.

You may use it internally to assess the corpus — inspecting, rendering, validating against your pipelines, computing embeddings and benchmarks, and limited-scale fine-tuning to judge suitability. You may not redistribute it, use it commercially, build derivative datasets from it, or deploy or distribute any model trained on it.

Production training rights are granted under a commercial license for the full corpus.

Citation

@misc{a23d_materials_sample_dataset_2026,
  title        = {A23D PBR Materials Sample Dataset for AI/ML Training},
  author       = {A23D},
  year         = {2026},
  publisher    = {A2VR Technologies LLP},
  url          = {https://huggingface.co/datasets/A23D/a23d-pbr-materials-sample}
}

About A23D

A23D develops enterprise-scale 3D assets and PBR material datasets for artificial intelligence, robotics, simulation and digital content creation — trusted, high-quality, human-authored 3D data for the next generation of AI systems.

One concept, one vision, one structure.

Downloads last month: 63

Total file size:

2.05 GB