Junhui Hou

Achieving a staggering $3,000\times$ reduction in computational complexity, REFINE redefines efficiency in 3D Gaussian splatting without sacrificing quality.

PRISM achieves superior performance in 3D representation learning by leveraging intrinsic geometric properties, outperforming traditional methods that rely on extrinsic features.

Achieve high-fidelity 3D reconstruction of thin structures from point clouds by decoupling metric distance and topological phase, sidestepping the limitations of both SDFs and UDFs.

Concept erasure in text-to-image models is mostly smoke and mirrors: a text-free attack can still regenerate "forgotten" concepts by exploiting the model's latent visual knowledge.

Stream 3D Gaussian Splatting scenes with higher visual quality and lower bandwidth by predicting user viewpoints and dynamically adapting bitrate using deep reinforcement learning.

Achieve 45x compression of 3D Gaussian Splatting data while *improving* visual fidelity by over 10% with a streaming-friendly octree-based codec.

By distilling visual foundation models, this work achieves a significant leap in event stream representation learning, surpassing prior methods in generalization, data efficiency, and transferability.