Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


Create by Keunhong Park, Konstantinos Rematas, Ali Farhadi, and Steven M. Seitz.

[Project Page] [arXiv]

Citing PhotoShape

If you find PhotoShape useful, please consider citing:

 author = {Park, Keunhong and Rematas, Konstantinos and Farhadi, Ali and Seitz, Steven M.},
 title = {PhotoShape: Photorealistic Materials for Large-Scale Shape Collections},
 journal = {ACM Trans. Graph.},
 issue_date = {November 2018},
 volume = {37},
 number = {6},
 month = nov,
 year = {2018},
 articleno = {192},

Getting Started

We try to provide everything required to get up and running, but due to licensing and copyright restrictions you will have to download/purchase some data directly from the source.

Python Environment

I recommend pyenv for managing python environments, but anything should work.

# Install pyenv
curl -L | bash
pyenv update

# Install dependencies and python
sudo apt-get install -y make build-essential libssl-dev zlib1g-dev libbz2-dev \
libreadline-dev libsqlite3-dev wget curl llvm libncurses5-dev libncursesw5-dev \
xz-utils tk-dev libffi-dev liblzma-dev
pyenv install 3.7.1
pyenv virtualenv 3.7.1 photoshape

# Enter the environment and set paths

Install Dependencies

Non-Python Dependencies

  • Blender 2.79 or higher
  • Docker and Docker-Compose
  • NVIDIA GPU compatible with PyTorch and OpenGL rendering.

Python Dependencies

If you just need to use the network, all you'll need is the basic dependencies.

pip install torch torchvision
pip install -r requirements.txt

If you want to run code that requires the dense CRF implementation, you'll need to install that package. This includes the fine alignment step and the substance segmentation for exemplars.

# For dense CRF (needed for substance segmentation).
pip install git+

Unfortunately our code for rendering segments and alignment renderings involves some custom shaders with vispy. This can be finnicky to set up. You'll need to install the following fork of PyOpenGL which supports NVIDIA extensions.

# For custom shader support (needed for rendering certain stuff in pipeline)
pip install git+

Please create an issue if something doesn't work as expected and I'll try to address it.


First we have to fetch this codebase.

# Clone this repository
git clone
git submodule update --init --recursive

Database Setup

To facilitate efficient querying and data storage we use PostgreSQL for most metadata.

First install docker and docker-compose following the instructions linked. Once you have those installed, run the database using the following command:

# Remove -d to show logs
docker-compose up -d

# You can stop the service by running 
# docker-compose down

Once the database is initialized and running, you can import data from the provided SQL dumps:

docker exec -i $(docker-compose ps -q postgres) psql -U photoshape_user photoshape_db < data/postgres/envmaps.sql
docker exec -i $(docker-compose ps -q postgres) psql -U photoshape_user photoshape_db < data/postgres/exemplars.sql
docker exec -i $(docker-compose ps -q postgres) psql -U photoshape_user photoshape_db < data/postgres/shapes.sql
docker exec -i $(docker-compose ps -q postgres) psql -U photoshape_user photoshape_db < data/postgres/materials.sql
docker exec -i $(docker-compose ps -q postgres) psql -U photoshape_user photoshape_db < data/postgres/exemplar_shape_pairs.sql


Data directories and paths are defined in src/terial/ Modify the directories so that they are pointing to the correct location on your disk(s). We've created a basic skeleton of the directory structure in data.

Blender Module

We use Blender for shape processing and rendering. Unfortunately installing Blender as a Python module is a pretty involved process. Follow directions here.

There also seems to be some effort to have this prepackaged at but I cannot guarantee that it will work.

Data Sets

Blob data associated with rows in the database are stores in a blobs directory. This directory can be configured in by modifying BLOB_ROOT. We recommend using a large disk for this since the data generated can be quite large.

Exemplar Images

We've preprocessed exemplar images. Download them to data/blobs

cd data/blobs
curl -L -O
tar xvzf exemplars.tar.gz


Herman Miller

You will have to download and process Herman Miller 3D shapes. Please review their terms of use before proceeding.

cd data/hermanmiller

# This will download zip files to the zip directory and then 
# extract *.3ds models to the extracted directory.

Next we will import the downloaded shapes and link them to rows in our database. Make sure that all of the *.3ds files are in the root of the directory. Some may be in random directories.

python -m terial.shapes.link_3ds data/hermanmiller/extracted/

The associated shape paths in data/blobs/shapes should now have uvmapped_v2.obj files.


Download and extract the file from the ShapeNet website and extract it to data/shapenet (or somewhere else, but remember to update the config if you do).

You will need to preprocess the ShapeNet shapes so that they have UV maps. Ensure that SHAPENET_CORE_DIR, SHAPENET_TAXONOMY_PATH and SHAPENET_META_DIR are properly set and run

# Preprocess a certain category.
python -m terial.shapes.preprocess_shapenet --category chair

Then we need to import the processed shape files so that the meshes are associated with the database rows.

python -m terial.shapes.link_shapenet


You can download the preview renders

cd data/blobs
curl -L -O
tar xvzf materials.tar.gz

Adobe Stock

Materials from Adobe Stock are subject to copyright and must be downloaded from their source. Please refer to the source_url field in materials.json.

Place the extracted materials sorted by substance type in the data/materials/adobe-stock directory.

Adobe stock materials should look like this:

├── denim_track_rough
│   ├── denim_track_rough_basecolor.png
│   ├── denim_track_rough_height.png
│   ├── denim_track_rough_metallic.png
│   ├── denim_track_rough_normal.png
│   └── denim_track_rough_roughness.png
└── denim_track_rough.mdl


Materials from Poliigon are subject to copyright and must be downloaded from their source. Please refer to the source_url field in materials.json.

Place the extracted materials sorted by substance type in the data/materials/poliigon directory.

Download the 1K version of the materials. Poliigon materials should look like this:

├── FabricCanvas001_COL_VAR1_1K.jpg
├── FabricCanvas001_COL_VAR2_1K.jpg
├── FabricCanvas001_DISP_1K.jpg
├── FabricCanvas001_GLOSS_1K.jpg
├── FabricCanvas001_NRM_1K.jpg
├── FabricCanvas001_OBJECTID_1K.png
└── FabricCanvas001_REFL_1K.jpg

Aittala Beckmann

We provide materials adapted from Aittala et al. and scanned ourselves. These SVBRDFs have been adapted from the original Aittala BRDF to the Beckmann model in order to support rendering using Blender.

Download the files and extract to data/materials.

curl -L -O
tar xvzf aittala-beckmann.tar.gz

V-Ray Materials

Materials from

Download the files and extract to data/materials.

curl -L -O
tar xvzf vray-materials-de.tar.gz

Parametric Materials

Materials with type BLINN_PHONG and PRINCIPLED are defined in the params field of each material in materials.json.


We've process our data using PostgreSQL. For your convenience we've exported the data in an easier to use collection of JSON files found in data/*.json.


This file contains meta-data regarding the 3D shapes used in this project.


  • id: the unique identifier for the shape
  • source: the source of the shape (either shapenet or hermanmiller)
  • source_id: the identifier from the source
  • category: the object category of the shape (e.g., chair, cabinet, etc.)
  • split_set: our train/validation split used for training our material classifier
  • azimuth_correction: a correction to the azimuth of the model orientation used to create preview renderings


Contains meta-data regarding the exemplar images.


Contains meta-data for Exemplar-Shape pairs.


  • id: a unique identifier for the pair
  • shape_id: the id for the shape associated with this pair
  • exemplar_id: the id for the exemplar associated wiith this pair
  • fov: the field of view of the aligned camera
  • azimuth: the azimuthal angle in radians of the aligned camera
  • elevation: the elevation angle in radians of the aligned camera
  • rank: the rank of this pair in terms of HoG distance between the shape and exemplar.


Contains meta-data for the materials.


  • id: the unique identifier for the material
  • name: a name for the material
  • author: the author of the material
  • source: the source of the material (e.g., poliigon, adobe_stock, aittala, merl, etc.)
  • source_id: the unique identifier from the source
  • source_url: the URL the material can be downloaded from
  • default_scale: the UV mapping should be scaled by 2^s where s is this value


  • MDL is the NVIDIA material definition language.
  • PRINCIPLED materials are the Disney principled BRDF as implemented in Blender.
  • AITTALA_BECKMANN materials are the Aittala TwoShot SVBRDF materials fitted to the Beckmann model as implemented in Blender.
  • For BLINN_PHONG and PRINCIPLED materials, the params field will contain the material properties.
  • For MDL materials, the params field will contain material properties and/or locations of textures.


This contains the mapping from material_id to output class labels for the material classifier.

Material Classifier

The weights for our Download the our network parameters

mkdir -p data/classifier
cd data/classifier
curl -L -O
tar xvzf model.tar.gz

Example Inference

Let's try to infer the materials on our sample input in data/classifier/input. For this example you'll need the material blobs downloaded as above and the set up properly. You'll also need to have set up the database since the script will query the database for the materials.

We'll show the inference results using visdom. Please install visdom, and have it running in the background somewhere:

pip install visdom
python -m visdom.server

Now you can run the example script:

python -m terial.classifier.inference.infer_one \
  --checkpoint-path data/classifier/model/model_best.pth.tar \
  data/classifier/inputs/chair1.png \

You should see the following in visdom if everything works correctly:

Inference Results