Export
Yuanrong Xu
1 other author named Yuanrong Xu: Yuanrong Xu (2 papers)
Research Overview
Yuanrong Xu focuses on computer vision, multi-modal representation learning, and image restoration. His trajectory spans from multi-view learning and scene text spotting to medical image analysis and robust vision-language adaptation. Most recently, his work addresses text editing in natural scenes and geometric consistency for few-shot generalization in foundation models.
Research Trajectory
Xu's early work explored foundational methods in representation learning. He developed techniques for multi-view learning, specifically introducing partial tubal nuclear norm regularization to better capture complex data structures in Partial Tubal Nuclear Norm Regularized Multi-view Learning. Expanding into applied computer vision the following year, he participated in research on scene text spotting, contributing to a single-shot architecture that decoupled text recognition from detection in Decoupling Recognition from Detection: Single Shot Self-Reliant Scene Text Spotter.
In 2024, Xu collaborated on medical image analysis, addressing the scarcity of high-quality data for automated dental caries detection. This effort produced CariesXrays: Enhancing Caries Detection in Hospital-Scale Panoramic Dental X-rays via Feature Pyramid Contrastive Learning, which introduced a large-scale hospital-sourced dataset alongside a feature pyramid contrastive learning framework. This work intersects with the broader research area of Dynamic Balancing of Competing Training Objectives by integrating dual-directional feature pyramids and proposals-prototype contrastive learning to improve the detection of small and varied lesions. Returning to core vision tasks that same year, he guided research on image enhancement, proposing a unified approach to blind image restoration using frequency adapters and spatial prompt networks in Frequency Adapter and Spatial Prompt Network for All-in-One Blind Image Restoration.
Building on his prior experience with scene text, Xu recently directed the development of a unified transformer-based architecture for modifying text in natural images. The resulting framework, detailed in Recognition-Synergistic Scene Text Editing, treats text recognition and text editing as complementary tasks. By leveraging the inherent capacity of recognition models to separate text content from visual style, the method replaces complex multi-stage pipelines. This contribution advances several distinct research directions, including the area of Disentangled Representation Learning for Content and Style Control, the field of Generative Cross-Modal Synthesis for Low-Data Vision, and the broader domain of Cycle Consistency for Unsupervised Image Transformation.
Continuing his focus on robust representation learning, Xu contributed to adapting vision-language models for downstream tasks. In Prompt Tuning for CLIP on the Pretrained Manifold, he and his collaborators addressed "manifold drift"—a phenomenon where prompt tuning causes models like CLIP to overfit to dataset-specific shortcuts. By enforcing geometric consistency and introducing an additive structural bias, the method ensures adapted features remain close to the robust pretrained manifold. This work directly supports the research direction of Addressing Overfitting in Few-Shot Learning.
Current Directions
The most recent papers in the evidence indicate a dual focus on generative image manipulation and the robust adaptation of foundation models. One primary direction involves synergistic approaches to scene text editing, where the disentanglement of visual style and textual content enables high-fidelity modifications in natural images without relying on multi-stage generation pipelines. Concurrently, recent work investigates the geometric properties of vision-language models during fine-tuning. By mitigating manifold drift and enforcing structural consistency, this research aims to improve cross-domain performance and few-shot generalization in large-scale pretrained models.
Profile updated
| Title | Venue | Year |
|---|---|---|
| Prompt Tuning for CLIP on the Pretrained Manifold | ICML | 2026 |
| Coordinate direction normalization using point cloud projection density for 3D ear | Openalex | 2009 |
| Paper | Venue | Year |
|---|---|---|
| Recognition-Synergistic Scene Text EditingGitHubZhengyaoFang/ | CVPR | 2025 |
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake