Keyu Tian
Keyu Tian is a computer-vision researcher whose work spans image augmentation, self-supervised learning, and generative models. His papers in 2020 and 2021 sped up image-augmentation search and explored continuous optimization, and he began interning at ByteDance in 2021. After switching from materials science to software engineering, he entered Peking University’s data-science master’s program in 2022. In 2023, he coauthored SparK for masked pretraining on convolutional networks; his 2024 VAR paper introduced next-scale image generation and won a NeurIPS Best Paper award. ByteDance dismissed and sued him over alleged interference with internal AI training in 2024; as of October 2026, he is building a startup for world models.
- Badges
Timeline
-
2020
Weight sharing for image augmentation
In 2020, he coauthors “Improving Auto-Augment via Augmentation-Wise Weight Sharing,” which evaluates image-augmentation policies with a fast proxy task based on shared weights.
-
2021
Continuous image augmentation
At NeurIPS in 2021, he coauthors “A Continuous Mapping for Augmentation Design,” which uses continuous optimization to explore image transformations. He begins interning at ByteDance in 2021.
-
2022
Materials to software engineering
He switches from materials science to software engineering during college. In 2022, he enrolls at Peking University for a master’s degree in data science after graduating from Beihang University’s School of Software.
-
2023
Sparse masked pretraining
In 2023, he coauthors “Designing BERT for Convolutional Networks: Sparse and Hierarchical Masked Modeling,” which presents SparK for masked pretraining on convolutional networks with sparse convolution and a hierarchical decoder.
-
2024
Next-scale image generation
In 2024, he coauthors “Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction,” which generates images through coarse-to-fine next-scale prediction.
-
August 2024–October 2026
A ByteDance lawsuit
ByteDance dismisses him in August 2024 and files a suit in November 2024, alleging code changes disrupted an internal AI training project. In 2026, he says he uses code to stop another intern’s program to free computing resources; a first-instance court orders him to pay 500,000 yuan, and ByteDance appeals.
-
2024
A NeurIPS Best Paper
NeurIPS selects his coauthored VAR paper for a 2024 main-track Best Paper award.
-
As of October 2026
Building world models
He is building a startup that develops world models for robotics and interactive video. His team has made an AI dictionary with 200,000 symbols to process video.
Career
- From 2021ByteDanceIntern
Education
- 2022Beihang University
- From 2022Peking UniversityMaster’s degree in data science
Sources and credits
Photo: We have no published photo of Keyu Tian yet. Send us a free-licensed photo.
- [2009.14737] Improving Auto-Augment via Augmentation-Wise Weight Sharing
- Publication
- A Continuous Mapping For Augmentation Design
- A Conversation with Tian Keyu: The Year After Being Fired by ByteDance, Sued, and Besieged by Public Opinion | PANews English
- Fired for Poisoning ByteDance's AI, His Startup Now Hits $200 Million Valuation
- [2301.03580] Designing BERT for Convolutional Networks: Sparse and Hierarchical Masked Modeling
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction
- ByteDance seeks $1.1 million damages from intern in AI breach case, report says
- NeurIPS 2024 Awards
- Ex-ByteDance intern raises $30M for stealth China world-models startup | Dealroom News
