Ziqi Jiang

I'm a 2-year PhD student at Long Group, HKUST, advised by Long Chen. Previouly I got my B.S. degree from ZheJiang University.

My research interests lie in computer vision, machine learning, with a special focus on grenative models, both their theory and applications.

Email  /  Scholar   

Publications

DP-FM is a Riemannian framework that decouples radial and angular feature dynamics on a cylindrical manifold to eliminate angular speed distortion and preserve modality confidence for effective few-shot vision-language adaptation.

FlowCIR reframes zero-shot composed image retrieval as a lightweight conditional flow matching problem and use an inference-only steering strategy to handle negation and removal edits.

We proposed path-decoupled Hyperbolic Flow Matching (HFM), which exploits the exponential expansion of the Lorentz manifold to decouple transport trajectories.

We are the first to explore the potential to adapt the multi-step recitification ability of flow matching for better few-shot learning performance.

We propose CLIPDrag, a solution that incorporates text signals into drag-based methods by using text signals as global information

We propose a causal multi-teacher distillation(CausalD) framework, which realizes FDA to estimate the causal effect and preserves the inference efficiency.

We devise a novel WED framework that disentangles fingerspelling letter representations through VAE and masked reconstruction, followed by the DMM that leverages the disentangled knowledge for detection and recognition as humans do.