Wangguandong Zheng

Zhiyuan Zhang*, Wangguandong Zheng*, Yansong Qu, Quanjian Song, Zesong Yang, Yuhao Liu, Jing Liao

MotionMem couples future trajectory prediction with memory retrieval in a plan–retrieve–generate framework for long-horizon world generation. Single-step motion forecasting guides the retrieval and alignment of historical object appearance, preserving scene and dynamic-object consistency during causal, few-step streaming generation.

Yansong Qu*, Wangguandong Zheng*, Quanjian Song, Zhiyuan Zhang, Xiao Cao

ScaleWorld uses one shared backbone for four-step camera-controlled world prediction and one-step high-resolution refinement. Separate causal histories reduce context-storage costs, while task-asymmetric on-policy distillation consolidates both capabilities into a single streaming model that refines its own predictions.

Zesong Yang, Weikai Chen, Liyuan Cui, Lutao Jiang, Runze Zhang, Yingda Yin, Xiaoyang Huang, Kai Yan, Keyang Luo, Wangguandong Zheng, Xin Wang, Hujun Bao, Zhaopeng Cui

GEAR uses geometry to route each target token to relevant historical visual memory for long-horizon camera-controlled video generation. Geometric Correspondence Attention retrieves matched features while an Invisible Octree filters occluded matches, enabling minute-long generation with consistent scene revisits without relying on globally fused 3D geometry.

HomepagePaperCode

Haodong Li, Wangguandong Zheng, Jing He, Yuhao Liu, Xin Lin, Xin Yang, Ying-Cong Chen, Chunchao Guo

DA² estimates dense depth directly from a single 360° panorama, with strong zero-shot generalization across scenes. It combines large-scale panoramic training data with SphereViT, a transformer that uses spherical coordinates to account for panoramic distortion and preserve geometric consistency without splitting the image into perspective views.

HomepagePaperCode

Tianyu Huang*, Wangguandong Zheng*, Tengfei Wang, Yuhao Liu, Zhenwei Wang, Junta Wu, Jie Jiang, Hui Li, Rynson W.H. Lau, Wangmeng Zuo, Chunchao Guo

Voyager turns a single image and a user-defined camera path into an explorable 3D scene by jointly generating RGB and depth videos. A persistent world cache and autoregressive scene extension preserve geometric consistency during long-range exploration, without requiring a separate 3D reconstruction pipeline at inference time.

HomepagePaperCode

Wenyue Chen, Peng Li, Wangguandong Zheng, Chengfeng Zhao, Mengfei Li, Yaolong Zhu, Zhiyang Dou, Ronggang Wang, Yuan Liu

SyncHuman combines 2D multiview and native 3D generative models to reconstruct clothed human meshes from a single image, including challenging poses. Pixel-aligned synchronization attention keeps the generated views and 3D shapes geometrically aligned, while feature injection transfers fine image details into the final textured mesh.

HomepagePaperCode

Peng Li, Wangguandong Zheng, Yuan Liu, Tao Yu, Yangguang Li, Xingqun Qi, Mengfei Li, Xiaowei Chi, Siyu Xia, Wei Xue, Wenhan Luo, Qifeng Liu, Yike Guo

PSHuman reconstructs a detailed, textured human mesh from a single image. Its SMPL-X-conditioned cross-scale multiview diffusion jointly models the full body and local facial details to preserve identity, while explicit mesh carving and remeshing recover geometry and appearance from the generated color and normal maps.

HomepagePaperCode

Wangguandong Zheng, Haifeng Xia, Rui Chen, Ming Shao, Siyu Xia, Zhengming Ding

Sketch3D creates 3D assets whose shapes follow an input sketch and whose colors match a text description. It combines a shape-preserving reference image with style-consistent multiview guidance to optimize 3D Gaussians, jointly refining structure, color, and similarity to the sketch.