Signal Processing and Speech Communication Laboratory
hometheses & projects › Generative Deep Learning for the Cration of 3D Objects

Generative Deep Learning for the Cration of 3D Objects

Master Thesis
Announcement date
05 May 2022
Research Areas

This project is a collaboration between the Signal Processing and Speech Communication Laboratory and the Institute of Architecture and Media at TU Graz. The overall topic and aim is the development of generative Deep Learning methods for the purpose of 3D object creation. This project combines state of the art Neural Network algorithms with the field of digital art and 3D design.

For generative deep learning applications usable datasets are rare and hard to come by. However, the Institute of architecture and media has developed multiple parametric models that produce a diverse dataset of 3d geometries/objects. The aim is to train various neural networks with these datasets to create new objects. The current focus lies on Variational Autoencoders, however other generative methods (GANS, Style-transfer, etc) are open for exploration. Further the topic of the 3d model representation (Voxel, Mesh, Point Clouds, etc) in neural networks will be part of this exploration.

This project is hands-on, meaning: coding, experimenting with different deep learning methods, development of new generative workflows, and creation of digital 3d objects are part of the main work. This project combines the disciplines of computer science/ML with the field of 3D design, digital art, and architecture. The student will with his/her master thesis be part and contribute to this ongoing research project.

Your Profile

  • Interest in machine learning and in particular, generative deep learning methods
  • Good experience in python, and knowledge about Deep Learning frameworks (for example TensorFlow 2.0, Pytoch, etc…)
  • Interest in the topics of 3D design, 3D modeling, digital art

Your tasks

  • literature research, esp. on suitable algorithms
  • implementation of model for 3D object creation
  • evaluation of the model

Reference Papers:



Franz Pernkopf (