Skip to content
Bryce Newell
All work

Machine learning

3D GAN

Using deep learning to start creating game-ready 3D assets from a dataset of cars.

  • TensorFlow
  • TensorLayer
  • Deep learning
  • GAN
  • Voxels
  • Unity
A voxel-based 3D car model generated by a neural network, shown in the Unity editor.

A voxel-based 3D car model generated by a neural network, shown in the Unity editor.

Framework
TensorFlow and TensorLayer
Training
24+ hours, ~3,000 epochs
Dataset
3D car models
Output
Voxel assets, viewed in Unity

Highlights

Deep learning

Using some of the latest deep learning techniques to bring voxel assets to life.

A chart of discriminator and generator loss against training epochs.

A chart of discriminator and generator loss against training epochs.

02

TensorFlow

TensorFlow was used to train the network. Discriminator and generator losses were tracked across the run.

A generated voxel car shown inside the Unity editor.

A generated voxel car shown inside the Unity editor.

03

Game-ready assets

Although the assets are a little unrefined, these images are taken directly from the Unity editor.

The story

The why

Deep learning has a vast range of applications, from recognising numbers to generating entire games. In this project I used TensorFlow and TensorLayer to generate 3D models from a dataset, focusing on cars.

Training the Generative Adversarial Network took over 24 hours and around 3,000 epochs to produce the results shown here. The generated cars are slightly undefined, but they are fully usable as game assets. It highlights the potential of deep learning for creating complex 3D models across a range of applications.