PeopleofSuperintelligence

Researcher

Kihyuk Sohn

Kihyuk Sohn is a research scientist whose work spans semi-supervised learning, computer vision and generative models. He earned a B.S. at KAIST and a Ph.D. at the University of Michigan, then developed learning methods at NEC and Google, including multi-class N-pair loss and FixMatch. StyleDrop adapted text-to-image models to visual styles, while VideoPoet used a language model for zero-shot video generation and won a Best Paper award at ICML 2024; he coauthored the ICLR 2024 papers MAGVIT-v2 and DreamFlow. He is a research scientist at Meta Reality Labs in Burlingame, where his recent work includes WorldGen, 3D-ARD+ and AssetGen.

Now at
Meta
Badges

Timeline

  1. 2003–2015

    KAIST and a Michigan PhD

    At KAIST in Daejeon, he studies electrical engineering and computer science and mathematical science from March 2003 to February 2008, earning a B.S. He completes a Ph.D. in electrical engineering: systems at the University of Michigan in June 2015 under thesis advisor Honglak Lee.

  2. 2015

    Deep conditional generative models

    In 2015, he coauthors “Learning Structured Output Representation using Deep Conditional Generative Models” with Xinchen Yan and Honglak Lee at NIPS. The model uses Gaussian latent variables and stochastic gradient variational Bayes to predict structured outputs, with experiments on pixel-level object segmentation and semantic labeling.

  3. 2015–2019

    Research at NEC Laboratories

    From July 2015 to July 2019, he works as a researcher in NEC Laboratories America’s Media Analytics group. In 2016, he introduces multi-class N-pair loss as sole author; the objective compares each example with N−1 negative examples in one update.

  4. 2019–2020

    FixMatch at Google Research

    He joins Google Research in Mountain View as a research scientist in July 2019. In January 2020, he is first author of FixMatch, a semi-supervised method that combines consistency regularization with pseudo-labeling, reaches 94.93% accuracy on CIFAR-10 with 250 labels and appears at NeurIPS 2020.

  5. 2023

    StyleDrop for text-to-image

    In June 2023, he is first author of StyleDrop, a method that tunes a text-to-image model to follow a style from one reference image while fine-tuning fewer than 1% of model parameters. In December 2023, Google Research publishes a post by him and Dilip Krishnan presenting StyleDrop as built on Muse.

  6. 2023–2024

    VideoPoet wins at ICML

    In December 2023, he coauthors VideoPoet with Dan Kondratyuk, Lu Jiang and colleagues; the decoder-only language model processes images, video, text and audio for zero-shot video generation. At ICML 2024, VideoPoet receives a Best Paper award.

  7. 2024

    MAGVIT-v2 and DreamFlow at ICLR

    In 2024, his coauthored work MAGVIT-v2 reaches ICLR; its video tokenizer represents images and videos with a shared vocabulary for language models. DreamFlow, another ICLR 2024 paper he coauthors, uses a three-stage, coarse-to-fine framework for text-to-3D generation.

  8. Since May 2024

    Research scientist at Meta

    He joins Meta Reality Labs in Burlingame, California, as a research scientist in May 2024. In September 2025, he seeks summer 2026 interns for image, video and 3D generation; his November 2025 paper WorldGen describes interactive 3D worlds generated from text prompts. In April and May 2026, he coauthors 3D-ARD+, a 7B autoregressive 3D diffusion model for text-directed object placement, and AssetGen, which creates a textured mesh from one reference image in 30 seconds, while he remains a Meta Reality Labs research scientist in October 2026.

Career

  1. Since 2024MetaResearch Scientist
  2. From 2019Google ResearchResearch Scientist
  3. 2015–2019NECResearcher

Education

  1. 2015University of MichiganPhD
  2. 2008KAISTBS

From the biography.

  1. Of 943 people shaping AI who worked at Google, 446 work there nowIncludes Kihyuk Sohn. Our estimates of where the 943 people shaping AI who worked at Google and Google DeepMind work now: 446 at Google and 452 across 278 other organizations.
  2. Google alumni have founded or led 133 other organizations, more than any other AI labIncludes Kihyuk Sohn. We estimate that alumni of 17 major AI labs went on to found or lead 350 other organizations, including 226 they founded, and we trace 472 career moves directly between 23 labs since 2010.
  3. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other universityIncludes Kihyuk Sohn. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other university. We compare doctorate and 5+ year employment links, career timing and current workplaces for 187 major-prize and 1,111 best-paper winners.
  4. The company share of AI researchers’ top papers was 31% in 2020–2024 and 8% in 1970–1974Includes Meta. We estimate where the people shaping AI worked when their most-cited papers appeared: the company share was 31% in 2020–2024 and 8% in 1970–1974, and Google led all organizations in 2020–2024.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Kihyuk Sohn yet. Send us a free-licensed photo.

  1. Kihyuk Sohn sites.google.com. Checked 21 October 2026.
  2. Learning Structured Output Representation using Deep Conditional Generative Models proceedings.neurips.cc. Checked 21 October 2026.
  3. Improved Deep Metric Learning with Multi-class N-pair Loss Objective proceedings.neurips.cc. Checked 21 October 2026.
  4. [2001.07685] FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence arxiv.org. Checked 21 October 2026.
  5. [2306.00983] StyleDrop: Text-to-Image Generation in Any Style arxiv.org. Checked 21 October 2026.
  6. StyleDrop: Text-to-image generation in any style research.google. Checked 21 October 2026.
  7. ICML 2024 Awards icml.cc. Checked 21 October 2026.
  8. [2312.14125] VideoPoet: A Large Language Model for Zero-Shot Video Generation arxiv.org. Checked 21 October 2026.
  9. Language Model Beats Diffusion -- Tokenizer is Key to Visual Generation research.google. Checked 21 October 2026.
  10. DreamFlow: High-quality text-to-3D generation by Approximating Probability Flow proceedings.iclr.cc. Checked 21 October 2026.
  11. [2511.16825] WorldGen: From Text to Traversable and Interactive 3D Worlds arxiv.org. Checked 21 October 2026.
  12. [2604.16552] Co-generation of Layout and Shape from Text via Autoregressive 3D Diffusion arxiv.org. Checked 21 October 2026.
  13. [2605.26137] AssetGen: Deployable 3D Asset Generation at Interactive Speed arxiv.org. Checked 21 October 2026.