Lucas Relic

I'm currently a PhD student at ETH Zürich, advised by Markus Gross. My PhD is part of a joint program between the Computer Graphics Laboratory and Disney Research|Studios. I work on learning-based methods for data compression.

I'm interested in applied research of deep learning and generative AI to domains such as data compression and computer graphics. My current research is about using generative models as a strong prior for data compression, particularly at extremely low bitrates. I previously worked in computational neuroscience, focusing on improving the quality of prosthetic vision.

Earlier, I earned my B.S. and M.S. at UCSB in computer engineering and computer science, respectively.

Email  /  Google Scholar

Research

We developed a quanitzation-based diffusion forward process and a uniform noise diffusion model, which improves compression performance at extremely low bitrates.

Foundation diffusion models can be built into an image codec by using them to "undo" quantization error.

We developed a novel transformer block to better capture spatio-temporal dependencies between frames in neural video compression.

By using a differentiable model of the human retina within an autoencoder, we optimized input stimuli for retinal implants (or any neuroprosthesis with a differentiable computational model).

We optimized the input to visual prosthetics using a neural approximation of a human retinal model.

Website design from Jon Barron.