Summary
Vista4D is a robust and flexible video reshooting framework that anchors both the input video and a target camera within a 4D point cloud. Given an input video, the method re-synthesizes the same dynamic scene from new camera trajectories and viewpoints. Existing video reshooting approaches struggle with depth estimation artifacts in real-world dynamic videos, fail to preserve content appearance, and lack precise camera control along challenging trajectories. Vista4D addresses this by constructing a 4D anchored point cloud representation with static-pixel segmentation and 4D reconstruction, explicitly preserving seen content while providing rich camera signals. The model is trained on reconstructed multi-view dynamic data, making it robust to point cloud artifacts during real-world inference. Results show improved 4D consistency, camera control, and visual quality over state-of-the-art baselines across diverse videos and camera paths. The framework also generalizes to real-world applications such as dynamic scene extension and 4D scene re-composition. Authors include Kuan Heng Lin, Zhizheng Liu, Pablo Salamanca, Yash Kant, Ryan Burgert, Yuancheng Xu, Koichi Namekata, Yiwei Zhao, Bolei Zhou, Micah Goldblum, Paul Debevec, and Ning Yu. arXiv: 2604.21915.
Paper Overview
Field: Computer Vision (CV)
Authors: Kuan Heng Lin, Zhizheng Liu, Pablo Salamanca, Yash Kant, Ryan Burgert, Yuancheng Xu, Koichi Namekata, Yiwei Zhao, Bolei Zhou, Micah Goldblum, Paul Debevec, Ning Yu
Release date: 2026-04-23
arXiv: 2604.21915
Abstract (translated from Chinese)
We present Vista4D, a robust and flexible video reshooting framework that anchors both the input video and target cameras in 4D point clouds. Specifically, given an input video, our method re-synthesizes a scene with the same dynamics from diverse camera trajectories and viewpoints.
Existing video reshooting methods typically struggle with depth estimation artifacts in real-world dynamic videos, fail to preserve content appearance, and cannot maintain precise camera control along challenging new trajectories. We construct a 4D anchored point cloud representation with static-pixel segmentation and 4D reconstruction to explicitly preserve seen content and provide rich camera signals. We train on reconstructed multi-view dynamic data to be robust to point cloud artifacts during real-world inference.
Our results show improvements in 4D consistency, camera control, and visual quality compared to state-of-the-art baselines across a variety of videos and camera paths. Moreover, our method generalizes to real-world applications such as dynamic scene extension and 4D scene re-composition.
Links
- arXiv: https://arxiv.org/abs/2604.21915
---
*Auto-collected on 2026-04-27*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/177618799