Efficient Deep CNN for Transport-Based Neural Style Transfer

As part of my master thesis, I have implemented a feed-forward version of 3D smoke stylization while considering temporal coherence. I have obtained promising results with a novel patch stitching technique that removes the need for overlaps with a customized multi-scale 3D U-Net architecture.

Dates:Sept 2019 - Jun 2020

Type:Master thesis at the Computer Graphics Laboratory at ETH. Paper submision planned for SIGGRAPH 2021.

Achievements:Awarded the highest grade (6.0/6.0)

Related skills:Machine Learning, PyTorch, Tensorflow, High-Performance Computing, Graphics, Fluid Simulation, Python, Tensorboard, Batch Processing, Research

Above there is a summary video of my thesis work. Note that the project is still on research at the Computer Graphics Laboratory of ETH where Vinicius and Byungsoo are taking over my work to further evaluate the implementation for preparation for SIGGRAPH 2021.

Supervised by: Byungsoo Kim, Vinicius Azevedo, Markus Gross, and Barbara Solenthaler.

contact me

Have a project in mind? I would love to hear from you.

Do you have something interesting? What can I do for you?