Research

Propose PreCoG, an efficient video correction framework combining low-cost preview assessment with training-free editing that accounts for source-condition misalignment during generation.

Introduce TempCloze, a video cloze benchmark for evaluating visual temporal reasoning beyond language shortcuts by identifying the missing middle from beginning and ending clips.

Propose SlotMem, a character-addressable internal memory framework for preserving long-range character consistency in narrative long-video generation.

Propose MIDiff, a multivariate-imaging diffusion framework for generating realistic and diverse mobile usage traces under sparsity and usage imbalance.

Propose U-MASK, a uniform model for multi-tasks generazabiltity through spatio-temporal masks for contexts-retrival under sparse observations.

Address the challenge of users’ daily mobile usage modeling under usage sparsity by correlation imaging for data augmentation.