TT-VidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining
TT-VidT’s reported edge is motion sensitivity at lower encoder cost.
The paper says its matched 24-run study isolates architecture and objective choices across similarly scaled video pretraining setups. TT-VidT pairs a DINOv3-initialized per-frame ViT path with a small temporal transfer layer, then trains with Diff Compression to separate appearance from frame-specific motion. In the authors’ final comparison, it leads on Jester, Something-Something V2, ARID, and Diving48 fine-tuning, while using fewer encoder FLOPs than DisMo, VideoMAE, and V-JEPA2. The claim is narrower on HMDB51, IARD, and EPIC-Kitchens. HF Daily Papers' note
The paper says its matched 24-run study isolates architecture and objective choices across similarly scaled video pretraining setups. TT-VidT pairs a DINOv3-initialized per-frame ViT path with a small temporal transfer layer, then trains with Diff Compression to separate appearance from frame-specific motion. In the authors’ final comparison, it leads on Jester, Something-Something V2, ARID, and Diving48 fine-tuning, while using fewer encoder FLOPs than DisMo, VideoMAE, and V-JEPA2. The claim is narrower on HMDB51, IARD, and EPIC-Kitchens. HF Daily Papers' note
score 4