VideoHandles: Editing 3D Object Compositions in Videos Using Video Generative Priors
TL;DR: VideoHandles is the first method for 3D object composition
editing in videos without any training. It achieves temporally consistent and context-aware
edits using a video generative prior.
Context-Aware 3D Edits in Videos
Input Video
Edited Video
Solid axes represnt the original 3D position, dotted axes the user-provided target position. VideoHandles
produces plausible edits, such as generating a new reflection for the wine glass and handling the disocclusion
of the lamp revealed behind the moved book pile.
Overview
Through 3D reconstruction from the source video, we introduce a 3D-aware warping function to ensure
frame-consistent transformations. This function warps the intermediate features of a video generative
model, guiding the generative process to position the object at the target position, while also
maintaining the plausibility of effects like shadows and reflections.
More Results
VideoHandles with CogVideoX
Input
Output
VideoHandles is independent of the choice of video generative models. Applying VideoHandles to a more
recent advanced video generative model, CogVideoX, produces much sharper and more detailed outputs.
Long and Stylized Video Editing
Input
Output
Real-World Video Editing
Input
Output
Baseline Comparison
Input
Direct Warping
Direct Warp. + SDEdit
DiffusionHandles
Ours
Click here for more comparison results.
Ablation Study
Input
w/ Temporal Feat.
w/o Null-Text
w/o Self-Attn
Ours
Click here for more ablation study results.
BibTeX
@inproceedings{Koo:2025VideoHandles,
title = {VideoHandles: Editing 3D Object Compositions in Videos Using Video Generative Priors},
author = {Koo, Juil and Guerrero, Paul and Huang, Chun-Hao Paul and Ceylan, Duygu and Sung, Minhyuk},
booktitle = {CVPR},
year = {2025}
}