Tero Karras

Career

Karras worked for three years in NVIDIA's Tegra business before joining NVIDIA Research in 2009, where Karras is now a Senior Distinguished Research Scientist. NVIDIA credits Karras with a central role in developing its RTX technology, covering both hardware and software design, and lists Karras's research areas as computer graphics, generative AI and real-time rendering.²¹

Karras's earlier interests, per the NVIDIA Technical Blog, included real-time ray tracing, representations for detailed 3D content, parallel algorithms and GPU computing. In late 2012 Karras wrote the blog's three-part "Thinking Parallel" series on collision detection, tree traversal and tree construction on the GPU.²

Research

Karras's generative modeling work began with "Progressive Growing of GANs" (2017), which trains generator and discriminator progressively from low to high resolution. The StyleGAN line followed: the style-based generator (2018), StyleGAN2 (2019), adaptive discriminator augmentation for training GANs with limited data (2020) and the alias-free StyleGAN3 (2021), each with Karras as first author.^7389101

Karras then turned to diffusion models. "Elucidating the Design Space of Diffusion-Based Generative Models" (EDM, 2022) separated the design choices in diffusion samplers and training, EDM2 (2023) analyzed and improved their training dynamics, and the 2024 autoguidance paper showed that guiding a diffusion model with a weaker version of itself can improve image quality. NVIDIA publishes the official implementations in its NVlabs GitHub organization.^5111264

Work and contributions

• StyleGAN family (StyleGAN to StyleGAN3) Published research

  First author of "A Style-Based Generator Architecture for Generative Adversarial Networks" (2018) and of "Alias-Free Generative Adversarial Networks" (StyleGAN3, NeurIPS 2021). The official StyleGAN3 implementation in NVlabs lists Karras first and is released under the NVIDIA Source Code License.³¹⁰⁴¹

• EDM and EDM2 diffusion models Published research

  First author of "Elucidating the Design Space of Diffusion-Based Generative Models" (EDM, NeurIPS 2022), "Analyzing and Improving the Training Dynamics of Diffusion Models" (EDM2, CVPR 2024), and "Guiding a Diffusion Model with a Bad Version of Itself" (NeurIPS 2024). The EDM2 repository releases source code and pretrained models under CC BY-NC-SA 4.0.⁵¹¹¹²⁶

Timeline

1. 2009Joined NVIDIA Research after three years with NVIDIA's Tegra business.²¹
2. Nov 2012Began the "Thinking Parallel" series on GPU algorithms on the NVIDIA developer blog.²
3. Oct 2017Posted "Progressive Growing of GANs for Improved Quality, Stability, and Variation" as first author.⁷
4. Dec 2018Posted the StyleGAN paper, "A Style-Based Generator Architecture for Generative Adversarial Networks".³
5. Dec 2019Posted StyleGAN2, "Analyzing and Improving the Image Quality of StyleGAN", as first author.⁸
6. Jun 2020Posted "Training Generative Adversarial Networks with Limited Data" as first author.⁹
7. Jun 2021Posted StyleGAN3, "Alias-Free Generative Adversarial Networks", as first author.¹⁰
8. Jun 2022Posted EDM, on the design space of diffusion-based generative models, as first author.⁵
9. Dec 2023Posted EDM2, on the training dynamics of diffusion models, as first author.¹¹
10. Jun 2024Posted "Guiding a Diffusion Model with a Bad Version of Itself" as first author.¹²

Where they work

• NVIDIA 43 people in the wiki

Related people

• David LuebkeAlso at NVIDIA
• Miika AittalaAlso at NVIDIA
• Sanja FidlerAlso at NVIDIA
• Ming-Yu LiuAlso at NVIDIA
• Tsung-Yi LinAlso at NVIDIA
• Justin JohnsonComputer vision · Image, video, and 3D generation
• Bilawal SidhuImage, video, and 3D generation · Computer vision
• Richard ZhangComputer vision · Image, video, and 3D generation
• Ron MokadyImage, video, and 3D generation · Computer vision
• Bichen WuComputer vision · Image, video, and 3D generation