Hongsong Wang

Replacing rigid autoregressive ordering with masked discrete diffusion and active self-correction resolves early error propagation in bidirectional motion-language modeling.

ScenePilot achieves superior scene generation by combining incremental growth with intelligent rectification, resulting in more physically plausible and functionally coherent 3D environments.

Bridging the semantic-kinematic gap, GenPrior achieves unprecedented accuracy in zero-shot action recognition by integrating generative motion priors with textual embeddings.

Uncertain samples can now be safely categorized into virtual classes, leading to enhanced learning from unlabeled data without the risk of introducing noise.

Noise amplification reveals hidden artifacts in AI-generated videos, outperforming traditional detection methods by focusing on the nuances of noise rather than just visual content.

Achieving over 4x speedup in text-to-image generation without sacrificing quality could revolutionize real-time creative applications.