Xinyao Zhang 张欣瑶

We address the fundamental challenge in instruction-guided video editing: simultaneously achieving precise semantic modification and faithful motion preservation. SAMA factorizes video editing into two sub-tasks — Semantic Anchoring (jointly predicting semantic tokens and video latents on sparse anchor frames to establish reliable visual anchpoints) and Motion Alignment (pre-training on motion-centric video inpainting tasks including cube inpainting, speed perturbation, and tube shuffle). A two-stage training strategy enables learning from unpaired data in Stage 1, achieving strong zero-shot editing, before supervised fine-tuning in Stage 2. SAMA achieves state-of-the-art among open-source methods on VIE-Bench, OpenVE-Bench, and ReCo-Bench, competitive with commercial systems like Kling-Omni.

Visual place recognition (VPR) under domain shifts remains challenging due to high-dimensional descriptors and distribution bias in traditional NetVLAD. SuperVLAD proposes a cluster-center-free framework that eliminates inter-domain distribution shift, achieving superior cross-domain generalization. With as few as 4 clusters, SuperVLAD matches or exceeds standard NetVLAD with many more clusters, significantly reducing descriptor dimensionality while achieving state-of-the-art performance on Pitts30k, MSLS-val, and multiple VPR benchmarks.