- Alternating-Attention Transformer: Inputs patch tokens and leverages global attention and frame attention, combining per-view detail with multi-view consistency.
- Training Strategy: Conducts multi-task supervision, including pose estimation, reconstruction and point tracking.
- Causal Streaming Architecture: Replaces global self-attention with temporal causal attention to enable real-time, sequential 4D updates for up to 100 frames.
- Knowledge Distillation: Leverages a pre-trained VGGT teacher to maintain high-fidelity reconstruction quality in the streaming student model.
- 3D point tracking Decomposition: Decomposes world-space 3D motion into scene geometry, camera ego-motion, and pixel-wise object motion
- A unified and differentiable pipeline: A fully differentiable pipeline that supports scalable joint training across heterogeneous data.