A unified image model for text-to-image, reference-based generation, and multi-turn conversational editing.
Yutao Cui 崔玉涛
Senior Researcher, Tencent Hunyuan
About Me
I am a Senior Researcher at Tencent Hunyuan. My work focuses on the pre-training and post-training of multi-modal foundation models, spanning image, video, and audio generation as well as native unified multi-modal architectures.
As a core contributor, I have been involved in the development and release of HY Image 3.5 Preview, HunyuanImage 3.0, HunyuanVideo, and related multi-modal systems. I also lead the projects of MixGRPO and the HYDRA / HYDRA-X series.
I received my Ph.D. in Computer Science and Technology from Nanjing University in 2024, advised by Prof. Limin Wang.
News
- 🥳 HYDRA and HYDRA-X were accepted to NeurIPS 2026.
- 🚀 HY Image 3.5 Preview was released.
- 🎉 Received WAICYOP Young Outstanding Paper — Honorable Mention.
- 🥳 MixGRPO was accepted to ECCV 2026.
- 🚀 HunyuanImage 3.0 was released. [Code] [Report]
Released Models
Selected industrial multi-modal models I contributed to at Tencent Hunyuan (with or without a public technical report).
HunyuanImage 3.0
An industrial-scale native multi-modal model for high-quality image generation.
HunyuanVideo
A systematic framework for large-scale video generation from Tencent Hunyuan.
HunyuanVideo-Foley
High-fidelity video-to-audio generation with multimodal diffusion and representation alignment.
Selected Publications
* indicates equal contribution. Full list on Google Scholar.
MixGRPO: Unlocking Flow-based GRPO Efficiency with Mixed ODE-SDE
An efficient RL method that mixes ODE/SDE sampling to accelerate flow-based preference alignment.
HYDRA: Unifying Multi-modal Generation and Understanding via Representation-Harmonized Tokenization
A unified multi-modal architecture that harmonizes representations for joint generation and understanding.
HYDRA-X: Native Unified Multimodal Models with Holistic Visual Tokenizers
Native unified multimodal models with holistic visual tokenizers for both images and videos.
MixFormer: End-to-End Tracking with Iterative Mixed Attention
An end-to-end transformer tracker with iterative mixed attention for joint feature extraction and target integration.
StableDrag: Stable Dragging for Point-based Image Editing
A stable point-based image editing method that improves dragging control and visual consistency.
MixFormerV2: Efficient Fully Transformer Tracking
A fully transformer tracker designed for high efficiency and real-time deployment.
SportsMOT: A Large Multi-Object Tracking Dataset in Multiple Sports Scenes
A large multi-object tracking dataset and benchmark focused on challenging sports scenes.
MixFormer: End-to-End Tracking with Iterative Mixed Attention
Fully Convolutional Online Tracking
Target Transformed Regression for Accurate Tracking
Awards & Honors
- WAICYOP Young Outstanding Paper — Honorable Mention, 2026
- Tencent Tech Elite Talent Program, 2024
- IEEE TPAMI Featured Article (MixFormer), 2024
- Runner-up in Object Tracking, The First Perception Test Challenge, Google DeepMind, 2023
- CVPR Oral Presentation (MixFormer), 2022