Xuhan SHENG (盛 栩涵)

Education

• Ranked 18th globally and 1st in Japan in Computer Science (QS Subject Rankings 2026).
• Incoming Ph.D. Student | Computer Science | Advisor: Rio Yokota

• Master's Degree | Computer Application Technology | GPA: 3.87/4.00 | Advisor: Jian Zhang (VILLA)

• Bachelor's Degree | Artificial Intelligence | GPA: 4.17/5.00 | Advisor: Xu Jia (IIAU-LAB)

Internships

• Image Super-Resolution based on Qwen-Image for game content and production-level visual assets.
• Built data pipeline operators for video generation models, including streaming shot segmentation for long videos.
• These techniques have been deployed and integrated into the production pipeline.

• AI Talent Program: Class of 2026 Dream-Seeking Internship
• Vision-Language Models for Assisting AI-based Image Enhancement.
• Advisor: Lei Zhang (IEEE Fellow, Chair Professor of Hong Kong Polytechnic University)

Research Background

• Diffusion-based Super-Resolution for Panoramic Images: Leveraging diffusion priors to restore ultra-high-resolution (4K×8K) omnidirectional images under real-world degradations for VR/AR applications.
• Vision-Language Models for AI-based Image Enhancement: Training VLMs to localize and annotate perceptual artifacts in AI-generated images, guiding more accurate super-resolution and enhancement.

Future Research Interest

• World Models for Interactive Long Video Generation: focusing on post-training and distillation acceleration for interactive and controllable video synthesis.
• 3D / Spatial Memory for World Models: investigating explicit 3D representation-based spatial memory mechanisms that provide world models with updatable, retrievable scene-level memory, so as to maintain spatial consistency of object appearance, scene layout, camera motion, and occlusion relationships during long-horizon generation and interaction.
• World Models for Visual Reasoning: exploring the use of world models for modeling physical laws, spatial relations, and causal dynamics, enabling visual systems not only to generate plausible scene evolutions but also to support reasoning over future states, latent relations, and complex interaction processes.

Publications

(* denotes co-first author)

• First Author Papers (ECCV'24 Oral; Under Review)
  1. RealOSR: Latent Unfolding Boosts Diffusion-based Real-world Omnidirectional Image Super-Resolution
     Xuhan Sheng^*, Runyi Li^*, Bin Chen, Weiqi Li, Xu Jiang, Jian Zhang.
     Under Review.
     • Overview: One-step diffusion-based omnidirectional image super-resolution guided by latent space unfolding, with lightweight domain alignment and degradation-aware modules for real-world restoration.
     • Contributions: Major contributor to codebase, experiments, and manuscript writing.
  2. OmniSSR: Zero-shot Omnidirectional Image Super-Resolution using Stable Diffusion Model
     Runyi Li^*, Xuhan Sheng^*, Weiqi Li, Jian Zhang.
     ECCV 2024 (Oral, Acceptance Rate: 2.3%) (Top-tier international conference in computer vision).
     • Overview: First zero-shot omnidirectional image super-resolution method leveraging Stable Diffusion priors, integrating Octadecaplex Tangent Information Interaction and Gradient Decomposition without training.
     • Contributions: Code implementation, method design, and experimental evaluation.
• Other Contributions (IJCV'25; ICME'25 Oral; PRCV'25; CVPRW'23 NTIRE Champion)
  1. OmniDrag: Enabling Motion Control for Omnidirectional Image-to-Video Generation
     Weiqi Li, Shijie Zhao, Chong Mou, Xuhan Sheng, et al.
     IJCV 2025 (Top-tier journal in computer vision).
     • Overview: First diffusion-based framework enabling precise drag-style motion control for omnidirectional image-to-video generation.
     • Contributions: A large-motion dataset construction (Move360).
  2. Bridging the Point to Boundary Gap for Point-supervised Temporal Action Localization with Single-stage Inference
     Junshi Yang, Shenglan Liu, Xuhan Sheng, et al.
     PRCV 2025 (The largest and most comprehensive technical conference in China, focusing on pattern recognition).
     • Overview: A hierarchical framework converting sparse point annotations into boundary-accurate pseudo labels via Gaussian-prior boundary enhancement.
     • Contributions: Assisted in paper writing and submission.
  3. Label-guided Facial Retouching Reversion
     Guanhua Zhao^*, Yu Gu^*, Xuhan Sheng, Yujie Hu, Jian Zhang.
     ICME 2025 (Oral) (Top-tier international conference in multimedia).
     • Overview: A diffusion-based method for reversing facial retouching effects such as skin smoothing, eye enlargement, and face slimming.
     • Contributions: Developed hierarchical adaptive instance normalization to mitigate color drift, achieving >20 FID improvement and +1.4 dB PSNR.
  4. OPDN: Omnidirectional Position-Aware Deformable Network for Omnidirectional Image Super-Resolution
     Xiaopeng Sun^*, Weiqi Li^*, Zhenyu Zhang, Qiufang Ma, Xuhan Sheng, et al.
     CVPRW 2023 (NTIRE 360° Panoramic SR Challenge Champion).
     • Overview: Champion solution for the NTIRE 2023 Challenge on 360° panoramic image super-resolution.
     • Contributions: Proposed a spatial-frequency fusion module and collected fisheye camera data to support model training.

Skills & Languages

• Programming: Python (PyTorch)
• IELTS Academic: 7.5 / 9.0

Honors and Awards

• Peking University Academic Excellence Award 2024
• Dalian University of Technology University-level Outstanding Graduate 2023
• Dalian University of Technology First-class Academic Excellence Scholarship 2022
• Dalian University of Technology Second-class Academic Excellence Scholarship 2021
• Dalian University of Technology First-class Academic Excellence Scholarship 2020