I'm Tianheng Cheng, and have finished my Ph.D. career at the HUST Vision Lab of Huazhong University of Science and Technology. I’m now a researcher at ByteDance Seed Team and working on cutting-edge large multimodal models and world models.

My lifelong research goal is to enable machines/robots to see, understand, and live like human beings.

Previous works/publications are listed at Google Scholar 📚.

Currently, I'm devoted to research about large multimodal models, foundational visual-language modeling, and image generation. Before that, I mainly focused on fundamental tasks such as object detection and instance segmentation, as well as visual perception for autonomous driving.

Highlighted Works of those pinned works:

• ControlAR (ICLR 2025) explores controllable image generation with autoregressive models and empowers autoregressive models with arbitrary-resolution generation.
• MaskAdapter (CVPR 2025) integrates seamlessly into open-vocabulary segmentation methods based on mask pooling in a plug-and-play manner, delivering more accurate classification results.
• EVF-SAM (arXiv) empowers segment-anything (SAM, SAM-2) with the strong text-prompting ability. Try our demo on HuggingFace.
• OSP (ECCV 2024) explores sparse set of points to predict 3D semantic occupancy for autonomous vehicles, which is a brand new formulation!
• YOLO-World (CVPR 2024) for real-time open-vocabulary object detection; Symphonies (CVPR 2024) for camera-based 3D scene completion.
• SparseInst (CVPR 2022) aims for real-time instance segmentation with a simple fully convolutional framework! MobileInst (AAAI 2024) further explores temporal consistency and kernel reuse for efficient mobile video instance segmentation.
• BoxTeacher (CVPR 2023) bridges the gap between fully supervised and box-supervised instance segmentation. With ~1/10 annotation cost, BoxTeacher can achieve 93% performance versus fully supervised methods.