About Me

Hi there, I am Changliang Xia (Chinese name: 夏常亮), a second-year Master’s student in Computer Science and Technology at Xi’an Jiaotong University, supervised by Prof. Minnan Luo. My research focuses on Computer Vision, with a specialization in image and video generation.

Currently, my work involves:

• Advanced methods for image and video generation and editing
• Diffusion-based visual perceptual tasks (e.g., depth estimation, normal estimation)
• Agent for Multimodal Creative Intelligence

I am passionate about building generative systems that are both controllable and creative, and I enjoy working at the intersection of vision, generative models, and multimodal intelligence.

I am actively seeking research internships for 2025/2026. If you are interested in collaboration or opportunities, please feel free to reach out to me at 202066@stu.xjtu.edu.cn.

🔥 News

• 2025.12: 🎉 Our papers [D³-Predictor, PaCo-RL] are recently released.
• 2025.06: 🎉 Our papers [T2IS, DenseDiT] are recently released.
• 2025.02: 🎉 ChatGen is accepeted by CVPR 2025.
• 2024.11: 🎉 Our papers [ChatGen] are recently released.

📝 Publications

Preprint

D³-Predictor: Noise-Free Deterministic Diffusion for Dense Prediction
Changliang Xia*, Chengyou Jia*, Minnan Luo, Zhuohang Dang, Xin Shen, Bowen Ping (* means equal contributions)

• We identify a key limitation in current diffusion-based dense prediction methods: the intrinsic stochastic noise of diffusion models disrupts the geometric structure and fine-grained spatial reasoning crucial for dense prediction tasks. To overcome this, we propose D³-Predictor, a deterministic framework that leverages noise-free diffusion priors for fast and accurate dense prediction.

CVPR 2025

ChatGen: Automatic Text-to-Image Generation From FreeStyle Chatting
Chengyou Jia*, Changliang Xia*, Zhuohang Dang, Weijia Wu, Hangwei Qian, Minnan Luo (* means equal contributions)

• This work tackles the trial-and-error challenge in text-to-image generation by automating the entire pipeline, enabling image creation through simple chat. We introduce the ChatGenBench benchmark and propose ChatGen-Evo, a novel training strategy that equips models with multi-step automation reasoning. Our approach significantly advances performance in generating high-quality images from freestyle user requests.

• Preprint From Ideal to Real: Unified and Data-Efficient Dense Prediction for Real-World Scenarios, Changliang Xia*, Chengyou Jia*, Zhuohang Dang, Minnan Luo (* means equal contributions)
• Preprint PaCo-RL: Advancing Reinforcement Learning for Consistent Image Generation with Pairwise Reward Modeling, Bowen Ping, Chengyou Jia, Minnan Luo, Changliang Xia, Xin Shen, Zhuohang Dang, Hangwei Qian Yu, Qian Liu, Chao Yang
• Preprint Why Settle for One? Text-to-ImageSet Generation and Evaluation, Chengyou Jia, Xin Shen, Zhuohang Dang, Changliang Xia, Weijia Wu, Xinyu Zhang, Hangwei Qian, Ivor W.Tsang, Minnan Luo

🎖 Honors and Awards

• 2023.04: 🏆 Finalist, Mathematical Contest in Modeling/Interdisciplinary Contest in Modeling (MCM/ICM)
• 2023.09: 🏆 Outstanding Student in Xi’an Jiaotong University
• 2022.09: 🏆 Outstanding Student in Xi’an Jiaotong University
• 2021.09: 🏆 Outstanding Student in Xi’an Jiaotong University
• 2025.11: 💰 Xi’an Jiaotong University Special Grade Scholarship
• 2024.11: 💰 Xi’an Jiaotong University First-Class Scholarship

📖 Educations

• 2024.09 - Present, M.S. Student, Computer Science, Xi’an Jiaotong University.
• 2020.09 - 2024.06, B.S. in Computer Science, Xi’an Jiaotong University.