Zilong Huang / 黄子龙

I’m now a Principal Researcher at Tencent Singapore. Before that, I was a Research Scientist at ByteDance. I received my Ph.D. degree and B.E. degree from Huazhong University of Science and Technology (HUST) in 2020 and 2015 respectively, advised by Prof. Wenyu Liu and Prof. Xinggang Wang. I was working as a visiting student (2018-2019) in the IFP group at the University of Illinois at Urbana-Champaign (UIUC), advised by Prof. Thomas S. Huang, Prof. Yunchao Wei and Prof. Humphrey Shi.

I am currently working on Multimodal Interaction, specializing in multimodal agents, multimodal pretraining, and efficient model architecture.

CV  /  Email  /  GitHub  /  Google Scholar  /  LinkedIn

[Hiring!] Our team is actively recruiting scientists, engineers and interns at all levels in China and Singapore, with a particular focus on GUI Agent/MLLM. Feel free to contact me if you are interested.

• [Sep 2025] Consistently ranked among the World's Top 2% Scientists by Stanford University/Elsevier (2022-2025)
• [Sep 2025] SuperCLIP is accepted by NeurIPS 2025.
• [Aug 2025] Invited to be an Area Chair for ICLR 2026.
• [Jun 2025] Three papers (SAIL(Highlight), GiGaTok and QK-Edit) are accepted by ICCV 2025.
• [May 2025] Seed1.5-VL Technical Report is released.
• [Feb 2025] DiG and Video Depth-Anything(Spotlight) are accepted by CVPR 2025.
• [Jan 2025] SeaFormer++ is accepted by IJCV.
• [Oct 2024] SuperClass and Depth-Anything V2 are accepted by NeurIPS 2024.
• [May 2024] Co-organizing the ECCV 2024 Workshop Synthetic Data for Computer Vision
• [Feb 2024] Depth-Anything is accepted by CVPR 2024.
• [Dec 2021] Delighted to receive CSIG Excellent Doctoral Dissertation Award

My selected publications are listed here. The complete list of publications can be seen from my Google Scholar page.

^ Intern. * equal contribution. + corresponding author

We propose SuperCLIP, a simple yet effective framework that augments contrastive learning with classification-based supervision. With just a 0.077% increase in total FLOPs and no need for additional annotated data, it can significantly improve the performance of visual models.

My primary responsibility was to conduct research on vision-language pre-training and to build the training pipeline of Seed-ViT.

We systematically compare SAIL’s properties—including scalability, cross-modal information flow patterns, and visual representation capabilities—with those of modular MLLMs.

We introduce GigaTok, the first method for scaling visual tokenizers to 3 billion parameters.

It's the first study to explore the simplest architecture for pixel-wise MLLM tasks, including referring segmentation and visual prompt understanding.

We propose Video Depth Anything for high-quality, consistent depth estimation in super-long videos (over several minutes) without sacrificing efficiency.

This work presents Diffusion GLA, the first exploration for diffusion backbone with linear attention transformer.

we introduce a new method squeeze-enhanced Axial Transformer (SeaFormer) for mobile visual recognition.

We introduce SuperClass, a super simple classification method for vision-language pre-training on image-text data.

Depth Anything V2 is trained from 595K synthetic labeled images and 62M+ real unlabeled images.

This work presents Meta Prompts, a simple yet effective scheme to harness a diffusion model for visual perception tasks.

This work presents Depth Anything, a highly practical solution for robust monocular depth estimation by training on a combination of 1.5M labeled images and 62M+ unlabeled images.

we propose the first self-supervised large-scale pretraining approach for image matting.

we propose a Motion Latent-based Diffusion model (MLD) could produce vivid motion sequences conforming to the given conditional inputs.

we design a generic attention block characterized by the formulation of squeeze Axial and detail enhancement for mobile vision.

CoCo-INR is a novel framework for implicit neural 3D representations, which builds a connection between each coordinate and the prior information.

Topformer is the first work that makes transformer real-time on mobile devices for segmentation tasks.

we revisit the spatial shuffle as an efficient way to build connections among windows in window-based self-attention.

we focus on the feature misalignment issue in previous popular feature aggregation architectures for semantic segmentation.

we tackle the problem of human de-occlusion which reasons about occluded segmentation masks and invisible appearance content of humans.

we propose HDMatt, a first deep learning based image matting approach for high-resolution inputs.

we present Agriculture-Vision: a large-scale aerial farmland image dataset for semantic segmentation of agricultural patterns.

More than 3000 citations, PaperDigest Most Influential ICCV 2019 papers (5th). Applications of CCNet also include AlphaFold2.

we propose a Criss-Cross Network (CCNet) for obtaining full-image contextual information in a very effective and efficient way.

we propose a Semantic Prediction Guidance (SPG) module which learns to re-weight the local features through the guidance from pixel-wise semantic prediction.

we propose a Scale-Adaptive Network (SAN) which consists of multiple branches with each one taking charge of the segmentation of the objects of a certain range of scales.

we identify several useful properties, including feature resolution, global context information and edge details, and perform rigorous analyses to reveal how to leverage them to benefit the human parsing task.

we propose to train a semantic segmentation network starting from the discriminative regions and progressively increase the pixel-level supervision using by seeded region growing.

Last updated on March 09, 2022. Thanks to Jon Barron for this minimalist website template.