Chaoqi Chen

Novel classes are rarely entirely novel—treating them as unseen compositions of known visual primitives enables robust open-world category discovery without modifying downstream heads or objective functions.

Temporal degradation in video diffusion isn't just caused by weak cross-frame attention—it is driven by an underlying spectral imbalance between local diffuseness and global diversity that can be selectively repaired at inference time without retraining.

Discovering new categories in open-world settings hinges on reshaping feature spaces, and CPF-GCD reveals how low-rank compositional structures can unlock this potential.

Instead of just pruning redundant tokens, ST-SimDiff dramatically cuts MLLM video processing costs by intelligently preserving tokens representing *changes* in the video.