MatSynth

MatSynth is a Physically Based Rendering (PBR) materials dataset designed for modern AI applications. This dataset consists of over 4,000 ultra-high resolution, offering unparalleled scale, diversity, and detail.

Meticulously collected and curated, MatSynth is poised to drive innovation in material acquisition and generation applications, providing a rich resource for researchers, developers, and enthusiasts in computer graphics and related fields.

For further information, refer to our paper: "MatSynth: A Modern PBR Materials Dataset" available on arXiv.

🔍 Dataset Details

Dataset Description

MatSynth is a new large-scale dataset comprising over 4,000 ultra-high resolution Physically Based Rendering (PBR) materials, all released under permissive licensing.

All materials in the dataset are represented by a common set of maps (Basecolor, Diffuse, Normal, Height, Roughness, Metallic, Specular and, when useful, Opacity), modelling both the reflectance and mesostructure of the material.

Each material in the dataset comes with rich metadata, including information on its origin, licensing details, category, tags, creation method, and, when available, descriptions and physical size. This comprehensive metadata facilitates precise material selection and usage, catering to the specific needs of users.

📂 Dataset Structure

The MatSynth dataset is divided into two splits: the test split, containing 89 materials, and the train split, consisting of 3,980 materials.

🔨 Dataset Creation

The MatSynth dataset is designed to support modern, learning-based techniques for a variety of material-related tasks including, but not limited to, material acquisition, material generation and synthetic data generation e.g. for retrieval or segmentation.

🗃️ Source Data

The MatSynth dataset is the result of an extensively collection of data from multiple online sources operating under the CC0 and CC-BY licensing framework. This collection strategy allows to capture a broad spectrum of materials, from commonly used ones to more niche or specialized variants while guaranteeing that the data can be used for a variety of usecases.

Materials under CC0 license were collected from AmbientCG, CGBookCase, PolyHeaven, ShateTexture, and TextureCan. The dataset also includes limited set of materials from the artist Julio Sillet, distributed under CC-BY license.

We collected over 6000 materials which we meticulously filter to keep only tileable, 4K materials. This high resolution allows us to extract many different crops from each sample at different scale for augmentation. Additionally, we discard blurry or low-quality materials (by visual inspection). The resulting dataset consists of 3736 unique materials which we augment by blending semantically compatible materials (e.g.: snow over ground). In total, our dataset contains 4069 unique 4K materials.

✒️ Annotations

The dataset is composed of material maps (Basecolor, Diffuse, Normal, Height, Roughness, Metallic, Specular and, when useful, opacity) and associated renderings under varying environmental illuminations, and multi-scale crops. We adopt the OpenGL standard for the Normal map (Y-axis pointing upward). The Height map is given in a 16-bit single channel format for higher precision.

In addition to these maps, the dataset includes other annotations providing context to each material: the capture method (photogrammetry, procedural generation, or approximation); list of descriptive tags; source name (website); source link; licensing and a timestamps for eventual future versioning. For a subset of materials, when the information is available, we also provide the author name (387), text description (572) and a physical size, presented as the length of the edge in centimeters (358).

🧑‍💻 Usage

MatSynth is accessible through the datasets python library. Following a usage example:

import torchvision.transforms.functional as TF
from datasets import load_dataset
from torch.utils.data import DataLoader

# image processing function
def process_img(x):
    x = TF.resize(x, (1024, 1024))
    x = TF.to_tensor(x)
    return x

# item processing function
def process_batch(examples):
    examples["basecolor"] = [process_img(x) for x in examples["basecolor"]]
    return examples

# load the dataset in streaming mode
ds = load_dataset(
    "gvecchio/MatSynth", 
    streaming = True,
)

# remove unwanted columns
ds = ds.remove_columns(["diffuse", "specular", "displacement", "opacity", "blend_mask"])
# or keep only specified columns
ds = ds.select_columns(["metadata", "basecolor"])

# shuffle data
ds = ds.shuffle(buffer_size=100)

# filter data matching a specific criteria, e.g.: only CC0 materials
ds = ds.filter(lambda x: x["metadata"]["license"] == "CC0")
# filter out data from Deschaintre et al. 2018
ds = ds.filter(lambda x: x["metadata"]["source"] != "deschaintre_2020")

# Set up processing
ds = ds.map(process_batch, batched=True, batch_size=8)

# set format for usage in torch
ds = ds.with_format("torch")

# iterate over the dataset
for x in ds:
    print(x)

⚠️ Note: Streaming can be slow. We strongly suggest to cache data locally.

📜 Citation

@inproceedings{vecchio2023matsynth,
  title={MatSynth: A Modern PBR Materials Dataset},
  author={Vecchio, Giuseppe and Deschaintre, Valentin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2024}
}

If you use the data from Deschaintre et al. contained in this dataset, please also cite:

@article{deschaintre2018single,
  title={Single-image svbrdf capture with a rendering-aware deep network},
  author={Deschaintre, Valentin and Aittala, Miika and Durand, Fredo and Drettakis, George and Bousseau, Adrien},
  journal={ACM Transactions on Graphics (ToG)},
  volume={37},
  number={4},
  pages={1--15},
  year={2018},
  publisher={ACM New York, NY, USA}
}