Holo4D: Holistic 4D Reconstruction as Geometric Control for Video Diffusion

Yushi Lan, Zeren Jiang, Koichi Namekata, Xingang Pan, Chuanxia Zheng, Andrea Vedaldi

University of Oxford · Nanyang Technological University

Published October 18, 2026 · on the feed October 30, 2026 · ▲ 6 · arXiv 2610.25276

Video3D, worlds & gamesAgents Full code arXiv ↗Hugging Face ↗

The abstract

Synthesizing a monocular video from a new camera trajectory requires preserving the observed scene geometry and motion while completing regions that become newly visible. Feed-forward 4D reconstructors recover the structure of the observed scene, whereas video diffusion models provide a generative prior for completing unseen content. We introduce Holo4D, a video reshooting framework that uses reconstructed 3D trajectories as a holistic source-view 4D scaffold. We project these trajectories into the requested views and encode their geometry and source appearance to guide a pretrained video diffusion model. Together with the source video and target cameras, the scaffold grounds generation in reconstructed observations while leaving unsupported regions to the generative prior. Experiments demonstrate the value of trajectory-based conditioning for camera control and novel-view synthesis. The framework accommodates different 4D reconstructors without retraining the video generator, allowing it to benefit from future advances in source reconstruction. More details can be found in our project page.

The Brief

1:46

A short podcast summary of the paper. Written and voiced by AI.

Transcript

Holo4D. Holistic 4D Reconstruction as Geometric Control for Video Diffusion. By Yushi Lan and five others, from the University of Oxford and Nanyang Technological University. Published on the eighteenth of October.

Reshooting a monocular video from a new camera trajectory asks for two things at once. The scene geometry and motion that were observed have to be preserved, faithfully, from the original footage. And the regions that only become visible from the new viewpoint have to be filled in, because the original camera never saw them. Two kinds of model each cover one half of that. Feed-forward 4D reconstructors recover the structure of what was observed. Video diffusion models supply a generative prior for completing what was not.

Holo4D joins the two halves. It is a video reshooting framework that takes reconstructed 3D trajectories and uses them as a holistic 4D scaffold of the source view, then lets that scaffold steer a pretrained video diffusion model. The reconstruction supplies the structure, and the diffusion model works inside it.

The trajectories are projected into the requested views, and their geometry and source appearance are encoded to guide the diffusion model. The scaffold does not act alone. Together with the source video and the target cameras, it grounds generation in what was actually reconstructed, while regions the reconstruction does not support are left to the generative prior to complete.

Experiments demonstrate the value of conditioning on trajectories, both for controlling the camera and for synthesizing novel views. The framework also accommodates different 4D reconstructors without retraining the video generator, so one trained generator can be paired with more than one way of recovering the source scene.

The division of labour is the design. Reconstruction says what is known, diffusion fills in what is not, and since the reconstructor can be swapped without retraining, future advances in source reconstruction should carry straight through to the reshot video.

More with an account

Listen to every day’s papers

Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.

Join free
KipQuest: the day’s AI research, as podcasts and quests. All papers Android app Privacy Contact Episodes are written and voiced by AI from each paper. Summaries can contain mistakes: the paper is the source.