Feiran Wang

We present SpatialMind, a vision-language model that combines metric geometry and progressive state-chain reasoning to understand observed spatial dynamics and predict future spatial states.

Kepler4D uses an explicit 4D proxy state and editable object trajectories to guide future video generation, supporting controllable motion and novel viewpoints.

ForeAct3D grounds future semantic 3D scene prediction in policy-generated actions and physical consistency to improve VLA policies without requiring future prediction at inference.

We revisit and observe that RayMap-based predictions exhibit inherent static scene bias and propose RayMap3R, a training-free streaming framework for dynamic scene reconstruction.

We present LIST3R, an instance-aware framework for long-sequence 3D reconstruction inspired by the way humans organize spatial memory around stable and recognizable objects.

CIF formulates a continuous probabilistic field over object existence and identity in space-time, enabling consistent instance representations across views for dynamic scene understanding.

We present CogniMap3D, a bioinspired framework that maintains a persistent memory bank of static scenes, enabling efficient spatial knowledge storage and rapid retrieval.

Rooted in the X-ray imaging process, X-Field presents a representation specifically for high-quality X-ray Novel View Synthesis and CT Reconstruction.

GPF reformulates photon mapping as a continuous radiance field of 3D Gaussian primitives, achieving photon-level accuracy for global illumination with greatly reduced computation.

To mitigate medical data scarcity, ZECO synthesizes high-quality 3D MRI images across various modalities, conditioned on segmentation masks.

Point clouds captured by LiDAR are often colorless; PCCN-RE enables high-quality colorization with a relevance embedding module built on a Conditional GAN.

Understanding morphology and distribution of neurons remains a significant challenge in modern neuroscience. We aim to develop a unified framework for sparse-to-dense neural generation and unsupervised segmentation, providing deeper insights into neural activity and connectivity.