MotionCrafter: Dense Geometry and Motion Reconstruction with a 4D VAE

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zhu, Ruijie, Lu, Jiahao, Hu, Wenbo, Han, Xiaoguang, Cai, Jianfei, Shan, Ying, Zheng, Chuanxia
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914428698492928
author Zhu, Ruijie
Lu, Jiahao
Hu, Wenbo
Han, Xiaoguang
Cai, Jianfei
Shan, Ying
Zheng, Chuanxia
author_facet Zhu, Ruijie
Lu, Jiahao
Hu, Wenbo
Han, Xiaoguang
Cai, Jianfei
Shan, Ying
Zheng, Chuanxia
contents We present MotionCrafter, a framework that leverages video generators to jointly reconstruct 4D geometry and estimate dense motion from a monocular video. The key idea is a joint representation of dense 3D point maps and 3D scene flows in a shared coordinate system, together with a 4D VAE tailored to learn this representation effectively. Unlike prior work that strictly aligns 3D values and latents with RGB VAE latents-despite their fundamentally different distributions-we show that such alignment is unnecessary and can hurt performance. Instead, we propose a new data normalization and VAE training strategy that better transfers diffusion priors and greatly improves reconstruction quality. Extensive experiments on multiple datasets show that MotionCrafter achieves state-of-the-art performance in both geometry reconstruction and dense scene flow estimation, delivering 38.64% and 25.0% improvements in geometry and motion reconstruction, respectively, all without any post-optimization. Project page: https://ruijiezhu94.github.io/MotionCrafter_Page
format Preprint
id arxiv_https___arxiv_org_abs_2602_08961
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle MotionCrafter: Dense Geometry and Motion Reconstruction with a 4D VAE
Zhu, Ruijie
Lu, Jiahao
Hu, Wenbo
Han, Xiaoguang
Cai, Jianfei
Shan, Ying
Zheng, Chuanxia
Computer Vision and Pattern Recognition
Artificial Intelligence
Computational Geometry
Machine Learning
We present MotionCrafter, a framework that leverages video generators to jointly reconstruct 4D geometry and estimate dense motion from a monocular video. The key idea is a joint representation of dense 3D point maps and 3D scene flows in a shared coordinate system, together with a 4D VAE tailored to learn this representation effectively. Unlike prior work that strictly aligns 3D values and latents with RGB VAE latents-despite their fundamentally different distributions-we show that such alignment is unnecessary and can hurt performance. Instead, we propose a new data normalization and VAE training strategy that better transfers diffusion priors and greatly improves reconstruction quality. Extensive experiments on multiple datasets show that MotionCrafter achieves state-of-the-art performance in both geometry reconstruction and dense scene flow estimation, delivering 38.64% and 25.0% improvements in geometry and motion reconstruction, respectively, all without any post-optimization. Project page: https://ruijiezhu94.github.io/MotionCrafter_Page
title MotionCrafter: Dense Geometry and Motion Reconstruction with a 4D VAE
topic Computer Vision and Pattern Recognition
Artificial Intelligence
Computational Geometry
Machine Learning
url https://arxiv.org/abs/2602.08961