EVENT DOSSIER
On September 7, 2026, arXiv cs.CV published the paper ASTRA. This study proposes an asynchronous spatio-temporal reconstruction framework called ASTRA, aimed at addressing the temporal inconsistencies in existing video generation models. The core of this method lies in introducing a trajectory alignment mechanism, which synchronizes spatial features with temporal dynamics to achieve more coherent video reconstruction. Experiments show that ASTRA significantly improves the temporal consistency and spatial detail quality of video generation in various benchmark tests.
ASTRA 提出一种基于轨迹对齐的异步时空重建框架，旨在解决多相机系统中因时间不同步导致的运动模糊和几何伪影问题。该方法通过引入 2D 运动轨迹作为显式、抗纹理的监督信号，联合优化时间偏移与动态 3D 表示，利用动态掩码抑制不可靠约束。在多种动态高斯泼溅骨干上的实验表明，ASTRA 在高达 25 帧的时间不同步下仍能保持鲁棒性，相比现有方法 PSNR 提升约 1.4 dB，时间偏移均方误差降低 54.0%，同步成功率几乎提高四倍。