Tero Karras

Timeline

1. By 2013

   Joining NVIDIA Research

   He joins NVIDIA Research in 2009 after three years with NVIDIA's Tegra business. In 2013 he is a senior research scientist at NVIDIA. Earlier work covers mobile graphics hardware and software at Hybrid Graphics.

2. 2018

   Progressive growing of GANs

   With Timo Aila, Samuli Laine, and Jaakko Lehtinen, he publishes Progressive Growing of GANs for Improved Quality, Stability, and Variation at ICLR 2018. The method grows both the generator and the discriminator from a low resolution and adds new layers that model increasingly fine details as training progresses.

3. 2019

   A style-based generator

   A Style-Based Generator Architecture for Generative Adversarial Networks appears at CVPR 2019, with Samuli Laine and Timo Aila. The architecture learns an unsupervised separation of high-level attributes, such as pose and identity when trained on human faces, and stochastic variation such as freckles and hair.

4. 2020

   Improving StyleGAN image quality

   Analyzing and Improving the Image Quality of StyleGAN appears at CVPR 2020 with Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. The paper redesigns generator normalization, revisits progressive growing, and regularizes the generator to encourage good conditioning in the mapping from latent codes to images. A path length regularizer makes the generator easier to invert.

5. 2021

   Alias-free generative networks

   Alias-Free Generative Adversarial Networks appears at NeurIPS 2021 with Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. The resulting networks are fully equivariant to translation and rotation even at subpixel scales.

6. 2022

   A NeurIPS outstanding paper

   Elucidating the Design Space of Diffusion-Based Generative Models, with Miika Aittala, Timo Aila, and Samuli Laine, receives a NeurIPS outstanding paper award in 2022. It is also a NeurIPS 2022 oral. The work reports a class-conditional CIFAR-10 FID of 1.79.

7. 2024

   A best paper runner-up

   Guiding a Diffusion Model with a Bad Version of Itself is a NeurIPS 2024 best paper runner-up. Coauthors include Miika Aittala, Tuomas Kynkäänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. ImageNet generation reaches FID scores of 1.01 for 64x64 and 1.25 for 512x512 on publicly available networks.

8. As of October 2026

   Senior scientist at NVIDIA

   He works as a senior distinguished research scientist at NVIDIA Research. His research covers deep learning, generative models, and digital content creation. He is the primary author of the StyleGAN family of generative models and has had a role in the development of NVIDIA's RTX technology, including both hardware and software design.