Wang Zhao
Publications
SCoPE: Sightline-Coordinate Positional Encoding for Video Diffusion Transformers
Video diffusion transformers address their tokens by position on the pixel-time grid: an address in the tensor, not in the world. The address we would want, the world point a token depicts, lies on a surface not yet generated, while its camera ray is fixed once the user specifies a trajectory. SCoPE therefore treats the ray as a second positional coordinate, and camera control becomes a property of the coordinate system, not an added module. The ray is added to the pretrained attention's queries and keys, and the score gains a term that reads the two rays alone. Its canonical form, the reciprocal product of line geometry, measures how nearly two lines of sight meet. Normalize-Gate-Inject makes a single encoding trainable across metric and up-to-scale pose sources. The retrofit keeps RoPE bit-exact, starts from the unchanged pretrained DiT, and adds under 0.1/% new parameters. On Wan2.2 at 5B and 14B under matched data and budget, SCoPE improves every camera-controllability and fidelity metric, leads all closed-loop revisit metrics, and shows widening margins with model size. At 14B, rotation error falls 29/% and FVD 43/% below the strongest baseline.
LLMTrack: Semantic Multi-Object Tracking with Multi-modal Large Language Models
Traditional Multi-Object Tracking (MOT) systems have achieved remarkable precision in localization and association, effectively answering \textit{where} and \textit{who}. However, they often function as autistic observers, capable of tracing geometric paths but blind to the semantic \textit{what} and \textit{why} behind object behaviors. To bridge the gap between geometric perception and cognitive reasoning, we propose \textbf{LLMTrack}, a novel end-to-end framework for Semantic Multi-Object Tracking (SMOT). We adopt a bionic design philosophy that decouples strong localization from deep understanding, utilizing Grounding DINO as the eyes and the LLaVA-OneVision multimodal large model as the brain. We introduce a Spatio-Temporal Fusion Module that aggregates instance-level interaction features and video-level contexts, enabling the Large Language Model (LLM) to comprehend complex trajectories. Furthermore, we design a progressive three-stage training strategy, Visual Alignment, Temporal Fine-tuning, and Semantic Injection via LoRA to efficiently adapt the massive model to the tracking domain. Extensive experiments on the BenSMOT benchmark demonstrate that LLMTrack achieves state-of-the-art performance, significantly outperforming existing methods in instance description, interaction recognition, and video summarization while maintaining robust tracking stability.