Go with Your Gut: Scaling Confidence for Autoregressive Image Generation
Confidence scaling for better autoregressive image generation.
Professional Summary
Wenjie Shu is a Research Assistant at The Hong Kong University of Science and Technology (HKUST), supervised by Prof. Harry Yang and Prof. Qifeng Chen. He obtained his B.E. degree from the University of Electronic Science and Technology of China (UESTC) in 2025, where he worked closely with Prof. Liangjian Deng. He is also fortunate to collaborate with Dr. Xiaogang Xu and Prof. Ser-Nam Lim.
He is always open to research collaborations and is currently applying to PhD programs, actively seeking opportunities for the next intake. Feel free to get in touch if you are interested in working with him! His research interests include Video Generation & Understanding, Reinforcement Learning and Computer Vision.
Education
B.E. in Information Engineering
University of Electronic Science and Technology of China (UESTC)
Visiting Student, Generative AI
The Hong Kong University of Science and Technology (HKUST)
Visiting Student, Video Generation
The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen)
Interests
Generative model
Computer Vision
Reinforcement Learning
📚 My Research
My research lies at the intersection of generative modeling and reinforcement learning, aiming for controllable, reliable, and efficient image/video generation.
- Video Generation & Temporal Coherence: text/image-to-video generation, multi-shot composition, and temporal regularization for stronger consistency.
- Alignment with Human Preferences: preference modeling and RL for diffusion models.
- Efficient Generation: knowledge distillation and training-efficient pipelines for lightweight diffusion models.
- Evaluation & Benchmarking: visual reasoning and robustness evaluation for video generators.
- Low-level vision: image fusion, super-resolution and low-light image enhancement.
I actively collaborate across academia and industry. If you’re interested in collaboration, feel free to reach out.
Featured Publications
CoRe-GRPO: Consensus-driven and Region-focused RL for Human-Centric Generation in Lightweight Diffusion Models
Region-focused RL for human-centric diffusion generation.
ThinkVid: Benchmarking Visual Reasoning in Video Generative Models
Benchmarking visual reasoning in video generation.
CMT: Cross Modulation Transformer with Hybrid Loss for Pansharpening
Cross Modulation Transformer with frequency-domain hybrid loss for pansharpening.
Exploring the Low-Pass Filtering Behavior in Image Super-Resolution
We explore frequency behavior in super-resolution models and its impact on reconstruction performance.
Recent Publications
(2025).
Enhancing Diffusion-based Restoration Models via Difficulty-Adaptive Reinforcement Learning with IQA Reward.
Under review.
(2025).
Go with Your Gut: Scaling Confidence for Autoregressive Image Generation.
Preprint.
(2025).
AlignVid: Training-Free Attention Scaling for Semantic Fidelity in Text-Guided Image-to-Video Generation.
Under review.
(2025).
CoRe-GRPO: Consensus-driven and Region-focused RL for Human-Centric Generation in Lightweight Diffusion Models.
Under review.
(2025).
ThinkVid: Benchmarking Visual Reasoning in Video Generative Models.
Under review.
Blog
