Biography

I am a researcher at Alibaba Tongyi Lab, where I lead Z-Image's reinforcement learning alignment. My work mainly focuses on foundation model post-training, spanning SFT, RL, and training infrastructure, while also contributing to pre-training for advanced model capabilities. My research centers on multimodal generation and understanding. Earlier, I worked on multimodal large language models at Tencent Hunyuan, with a focus on improving their advanced reasoning capabilities.

Featured projects

Z-Image Image generation Z-Reward Reward modeling

News

  • 2026 New paper on pixel-space text-to-image diffusion models released on arXiv.
  • 2026 Released Z-Reward, a reasoning-internalized teacher-student reward modeling framework for text-to-image post-training.
  • 2026 Released Z-Image (Base), the high-quality foundation model specializing in rich aesthetics and controllability. The series is now the #2 most popular Chinese model on Hugging Face and ranks 2nd in global Image Generation API usage on fal.ai.
  • 2025 Released Z-Image-Turbo, an efficient 6B foundation model specializing in photorealistic image generation. It ranked #1 among open-source text-to-image models on the Artificial Analysis Image Arena, while achieving sub-second inference latency.
  • 2024 Released MM-IQ, a new benchmark for assessing the core reasoning capabilities of large multimodal models.
  • 2024 New paper on Self-Correction in LLMs released on arXiv.
  • 2024 Paper System-2 Mathematical Reasoning accepted to TMLR.
  • 2023 The extended paper Tri-token Equipped Transformer Model for Image Matting released on arXiv.
  • 2022 TransMatting accepted to ECCV.
  • 2021 Won the 2nd Place Award in NTIRE 2021 Challenge on Multi-modal Aerial View Object Classification at CVPR 2021.

Publications

* Equal contribution. † Project Lead.

Selected work

Tech Report, 2025

Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer GitHub Stars

Huanqia Cai, Sihan Cao, Ruoyi Du, Peng Gao, Steven Hoi, et al.

Z-Image Teaser

Z-Image is a state-of-the-art foundation model designed for high efficiency. Community Impact: It is the #2 most popular Chinese model on Hugging Face and ranks 2nd in global API usage on fal.ai (next to Gemini Nano). It excels at generating photorealistic images with superior visual quality while maintaining low computational cost.

Paper Code HuggingFace
Tech Report, 2026 Project Lead

Z-Reward: Beyond Scalar Rewards by Internalizing Reasoning into Score Distributions

Xin Jin*, Huanqia Cai*†, Zhen Li, Zechao Zhan, Dengyang Jiang, Aiming Hao, Yuming Jiang, Chunle Guo, Peng Gao, Ming-Ming Cheng, Steven C.H. Hoi

Z-Reward Teaser

Z-Reward is a reasoning-internalized reward modeling framework for visual generation. It decouples reasoning-heavy judgment from efficient reward deployment: a 27B GDSO teacher reasons over rubric-aligned score distributions, while a compact 9B RISD student internalizes the teacher's distributional judgment without generating explicit reasoning chains at inference time.

Project Paper PDF HF Daily Paper

More publications

5 papers
Showing 2 of 5Showing all 5 Show 3 more publicationsShow fewer publications