CVPR2025
AC3D: Analyzing and Improving 3D Camera Control in Video Diffusion Transformers
Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B. Lindell, Sergey Tulyakov
摘要
2 Vector Institute 3 Snap Inc. 4 SFU *equal contribution https://snap-research.github.io/ac3d Figure 1. Camera-controlled video generation. Our method enables precise camera controllability in pre-trained video diffusion transformers, allowing joint conditioning of text and camera sequences. We synthesize the same scene with two different camera trajectories as input. The inset images visualize the cameras for the videos in the corresponding columns. The left camera sequence consists of a rotation to the right, while the right camera visualizes a zoom-out and up trajectory.