VitronEdit

VitronEditUniversal Visual Editing
across Images, Videos, and 3D Scenes

Jing Wang1,†, Yanlin Li1,†, Mong-Li Lee1, Wynne Hsu1,
Shengqiong Wu2,*, Hao Fei2,*

1 National University of Singapore 2 University of Oxford

† Equal contribution    * Corresponding authors

Read the paperExplore the code

“Make the camera a wooden toy.”

“Change the car to a jeep.”

“Make the bonsai snowy.”

Cross-modal knowledge transfer · Geometry-aware post-training

Shared editing semantics enable transfer from abundant image and video data to scarce 3D data; shared geometric priors preserve consistency within images, across frames, and across viewpoints.

Two questions.
One universal editor.

  1. Can abundant image and video editing data help data-scarce 3D editing?
  2. Can shared geometric priors improve consistency across images, videos, and 3D scenes?

A universal editing framework.

VitronEdit architecture: visual and text conditioning, shared generation backbone, and a frozen geometry decoder for depth and camera-pose supervision

Unified representation

Static frames, temporal frames, and multiview observations share a sequence-processing pipeline.

Cross-modal transfer

Image and video data transfer editing semantics; 3D supervision improves multiview consistency.

Geometry-aware learning

A frozen teacher provides depth and camera-pose signals to preserve structural consistency.

Abstract

Recent advances in unified visual editing have begun to consolidate image and video editing within a single model, while 3D scene editing remains largely separate and constrained by scarce paired training data. Given the rich image and video editing supervision now available and the common editing semantics shared across modalities, a natural question is whether this knowledge can be transferred to data-scarce 3D editing. We present VitronEdit, a universal framework for image, video, and 3D scene editing. VitronEdit adopts a shared sequence representation and progressive multimodal training to transfer editing knowledge across modalities. We further introduce geometry-aware post-training, which leverages a pretrained geometry reconstruction model to provide depth and camera-pose supervision. Surprisingly, applying geometric supervision only to 3D samples yields the best overall performance, while image and video editing also improve despite receiving no direct geometric supervision, suggesting that geometric regularization transfers through the shared backbone. Our analysis further shows that image and video training mainly transfers editing semantics to 3D, whereas 3D supervision primarily improves multiview consistency.

BibTeX

Download BibTeX
@misc{wang2026vitronedit,
  title  = {VitronEdit: Universal Visual Editing across Images, Videos, and 3D Scenes},
  author = {Wang, Jing and Li, Yanlin and Lee, Mong-Li and Hsu, Wynne and Wu, Shengqiong and Fei, Hao},
  year   = {2026}
}

Editing demo