Research Experience

To address deformation, blur, and structural-detail loss, proposed SymTrack, the first detection-free scene text tracking framework, and built a dedicated benchmark from three video text spotting datasets. Independently completed the implementation, ablations, reproduction of 10+ recent trackers, evaluation, and paper writing. The work achieved state-of-the-art results on all three benchmarks and was accepted by ICML 2026 as a sole first-author paper. Paper · Code

To improve glyph rendering and reduce background distortion, developed the first training-free video text editing framework based on Wan 2.2, using frequency-guided glyph injection and mask-trajectory preservation. It outperformed VACE, FlowEdit, and FlowAlign in overall quality. Led the first VTE benchmark spanning real and synthetic data, reproduced leading methods, and independently completed the implementation, ablations, and manuscript. The work is under review at AAAI 2027.

To mitigate visual hallucinations caused by overreliance on semantic priors, contributed to a hallucination-resistant dual-branch diffusion framework with a three-source evidence gating module for adaptive structural and semantic fusion. The method achieved state-of-the-art results on CTR and Real-CE under ×2 and ×4 settings. Worked on the implementation, ablation studies, and manuscript; the second-author paper was accepted by ACM Multimedia 2026.

Studied handwritten formula recognition and table parsing, reproduced recent state-of-the-art models, and independently ported TAMER (AAAI 2025) from PyTorch to Jittor with aligned accuracy. Completed the program and passed the final presentation.

Investigating a unified-representation framework that addresses conflicts between discriminative visual understanding and generative modeling. The framework unifies diverse conditional inputs and outputs so that a single model can perform conditional generation alongside visual tasks such as object detection, semantic segmentation, and depth estimation.