Chunyu Wang

About

I am building a startup, working on personalized multi-modal computing. Previously, I was a Team Lead at Tencent Hunyuan, where I worked on image/video generation and was responsible for developing Hunyuan Image Editing models (now powering Yuanbao, Tencent Games, Tencent Videos, etc), HunyuanImage 3.0 post-training, and a few application models. Before that, I was a Principal Researcher at Microsoft Research Asia working on computer vision. I serve as area chair and reviewer for a few top computer vision and machine learning conferences, and received the AAAI Best Paper Award in 2026. I obtained my PhD degree from Peking University, working with Prof. Yizhou Wang and Alan L. Yuille (UCLA).

Google Scholar / 知乎

🚀 News

• Apr 2026 Released HY-SOAR, a reward-free post-training method for diffusion models that goes beyond SFT and RL — no reward models, preference labels, or negative samples needed.
• Apr 2026 Released HiVG, a 3B-parameter model that beats GPT-5 and Gemini 2.5 on image-to-SVG via hierarchical SVG tokenization.
• Feb 2026 Five papers on image editing are accepted by CVPR 2026!
• Jan 2026 Our paper "LLM2CLIP" received the AAAI 2026 Best Paper Award
• Oct 2025 Released HY-Edit 1.0, our image editing model, now powering Yuanbao.
• Oct 2025 Released HunyuanImage 3.0, the first open-source unified understanding and generation image model, ranking #1 on LMArena.
• Sep 2025 Released PromptEnhancer, powering Hunyuan Image 2.1 to rank #1 among open-source models on Artificial Analysis Arena.
• Aug 2025 Open-sourced SRPO, an RLHF algorithm that boosts realism of generated images, powering Flux-dev to rank #2 among open-source models on Artificial Analysis Arena.

⭐ Selected Publications

• HY-SOAR: Self-Correction for Optimal Alignment and Refinement in Diffusion Models
  2026 [arXiv] [Project] [Code]
• Hierarchical SVG Tokenization: Learning Compact Visual Programs for Scalable Vector Graphics Modeling
  2026 [arXiv] [Project] [Code]
• LLM2CLIP: Powerful Language Model Unlocks Richer Cross-Modality Representation
  2026 [arXiv] [Project]
• HunyuanImage 3.0 Technical Report
  2025 [arXiv] [Project]
• HunyuanVideo: A Systematic Framework For Large Video Generative Models
  2024 [arXiv] [Project]
• Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone
  2024 [arXiv] [Project]
• PromptEnhancer: A Simple Approach to Enhance Text-to-Image Models via Chain-of-Thought Prompt Rewriting
  2025 [arXiv] [Project]
• Directly Aligning the Full Diffusion Trajectory with Fine-Grained Human Preference
  2025 [arXiv] [Project]
• JarvisEvo: Towards a Self-Evolving Photo Editing Agent with Synergistic Editor-Evaluator Optimization
  2025 [arXiv] [Project]
• FairMOT: On the Fairness of Detection and Re-Identification in Multiple Object Tracking
  International Journal of Computer Vision (IJCV), 2021 [arXiv] [Project]
• VoxelPose: Towards Multi-Camera 3D Human Pose Estimation in Wild Environment
  ECCV 2020 (Oral) [arXiv] [Project]

Talks and Blogs

• On-Policy Distillation去蒸馏多个专家为什么没成为多模态生成的标配？ (知乎分享)
  Apr 2026
• 视觉理解与生成统一应该发生在哪一层？任务还是表达？ (知乎分享)
  Apr 2026
• 《NextFlow论文阅读》 (知乎分享)
  Feb 2026
• 图像生成模型的后训练实践 (中国图象图形学学会珠峰论坛)
  Jan 2026
• Graphic Design Generation: An Early Attempt (稀土掘金大会)
  Jun 2024
• Structured Representation for Human Pose (中文) (知乎分享)
  Apr 2023
• Avoiding Model Collapse in Semi-supervised Learning (中文) (知乎分享)
  Nov 2020
• Handling Occlusion in Human Pose Estimation (Max Planck Institute)
  Apr 2020
• 人体姿态估计：一眼万里，洞见千姿百态 (Valse Online)
  Aug 2020

Contact

Email: chunyu.wangdlut@gmail.com