VidEdit3D: Flexible 3D Gaussian Editing
with Video Diffusion

1 Westlake University2 University of Leeds3 Nanyang Technological University4 University of California, Merced

† Corresponding author

Qualitative results of VidEdit3D.

VidEdit3D propagates an edit from a single anchor view to a 3D Gaussian scene, supporting object additions, removals, appearance changes, and user-provided references.

Method

VidEdit3D reformulates 3D scene editing as trajectory-based video editing. Starting with a reconstructed scene and a user-edited anchor, it organizes smooth camera paths, propagates the edit across their rendered views, and refines the 3D representation.

VidEdit3D pipeline: Trajectory-Guided Rendering organizes source views into video clips, Flow-Aligned Video Editing propagates an anchor edit, and Region-Selective 3DGS Refinement updates target Gaussians.
Overview of VidEdit3D. Vector PDF ↗

Trajectory-Guided Rendering (TGR) selects target-supported views and constructs smooth video clips. Flow-Aligned Video Editing (FAVE) propagates the anchor edit with a pretrained video diffusion model and Editing Consistency Guidance. Region-Selective 3DGS Refinement (RS-GS) transfers the edited frames back into 3D while restricting updates to the target Gaussians.

Demo

Editing Results

Select a case to view the source and edited scenes along the same camera trajectory.

3D scene editing

Loading results…

Source
Ours
Explore views

Comparisons

Each example shows the source scene, four baselines, and our result.

Source · Baseline · Ours

Citation

If you find this work useful, please consider citing:

BibTeX
@misc{yang2026videdit3dflexible3dgaussian,
  title={VidEdit3D: Flexible 3D Gaussian Editing with Video Diffusion},
  author={Yanming Yang and Chenxi Song and Tong Zhao and Juntao Xu and Wenhao Rao and Yiwen Chen and Yiwei Wang and Chi Zhang},
  year={2026},
  eprint={2610.35756},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.35756},
}

3D editing result