Gen4P: Video Generators are 4D Perceivers
Synthetic voice
Gen4P, by Yitong Xia, Zeren Jiang, and Chuanxia Zheng, repurposes a pretrained video generator into a single-step feed-forward 4D perceiver that recovers geometry, depth, camera poses, and 3D point trajectories from monocular video. It processes clean video latents in one encoder pass, uses continuous-time cross-attention queries for temporal flexibility, and optionally plugs in a frozen video-depth LoRA as a geometry prior without adding trainable parameters. The authors report state-of-the-art or competitive performance on 3D point tracking, video depth, and camera pose estimation.
- Gen4P converts a multi-step video generator into a single-step feed-forward 4D perceiver without iterative denoising.
- A frozen video-depth LoRA serves as a plug-in geometry prior, improving performance without additional trainable parameters.
- The authors report state-of-the-art or competitive results on 3D point tracking, video depth, and camera pose estimation.